skill-iter-tune — Skillopedia

Skill Iter Tune Iterative skill refinement through execute-evaluate-improve feedback loops. Each iteration runs the skill via Claude, evaluates output via Gemini, and applies improvements via Agent. Architecture Overview Chain Mode Extension Key Design Principles 1. Iteration Loop : Phases 2-3-4 repeat until quality threshold, max iterations, or convergence 2. Two-Tool Pipeline : Claude (write/execute) + Gemini (analyze/evaluate) = complementary perspectives 3. Pure Orchestrator : SKILL.md coordinates only — execution detail lives in phase files 4. Progressive Phase Loading : Phase docs read…

).replace(/`/g, '\\\\`');\n}\n\nconst cliCommand = `ccw cli -p \"${escapeForShell(executePrompt)}\" --tool claude --mode write --cd \"${iterDir}/artifacts\"`;\n\n// Execute in background, wait for hook callback\nBash({\n command: cliCommand,\n run_in_background: true,\n timeout: 600000 // 10 minutes max\n});\n\n// STOP HERE -- wait for hook callback to resume\n// After callback, verify artifacts were produced\n```\n\n### Step 2.2b: Chain Execution Path\n\n> Skip this step if `state.execution_mode === 'single'`.\n\nIn chain mode, execute each skill sequentially. Each skill receives the previous skill's artifacts as input context.\n\n```javascript\n// Chain execution: iterate through chain_order\nlet previousArtifacts = ''; // Accumulates upstream output\n\nfor (let i = 0; i \u003c state.chain_order.length; i++) {\n const skillName = state.chain_order[i];\n const skill = state.target_skills.find(s => s.name === skillName);\n const skillArtifactDir = `${iterDir}/artifacts/${skillName}`;\n\n // Build this skill's content\n const skillMd = Read(`${skill.path}/SKILL.md`);\n const phaseFiles = Glob(`${skill.path}/phases/*.md`).sort().map(f => ({\n relativePath: f.replace(skill.path + '/', ''),\n content: Read(f)\n }));\n const specFiles = Glob(`${skill.path}/specs/*.md`).map(f => ({\n relativePath: f.replace(skill.path + '/', ''),\n content: Read(f)\n }));\n\n const singleSkillContent = `### File: SKILL.md\\n${skillMd}\\n\\n` +\n phaseFiles.map(f => `### File: ${f.relativePath}\\n${f.content}`).join('\\n\\n') +\n (specFiles.length > 0 ? '\\n\\n' + specFiles.map(f => `### File: ${f.relativePath}\\n${f.content}`).join('\\n\\n') : '');\n\n // Build chain context from previous skill's artifacts\n const chainInputContext = previousArtifacts\n ? `\\nPREVIOUS CHAIN OUTPUT (from upstream skill \"${state.chain_order[i - 1]}\"):\\n${previousArtifacts}\\n\\nIMPORTANT: Use the above output as input context for this skill's execution.\\n`\n : '';\n\n // Construct per-skill execution prompt\n // Ref: templates/execute-prompt.md\n const chainPrompt = `PURPOSE: Simulate executing the following workflow skill against a test scenario. Produce all expected output artifacts.\n\nSKILL CONTENT (${skillName} — chain position ${i + 1}/${state.chain_order.length}):\n${singleSkillContent}\n${chainInputContext}\nTEST SCENARIO:\nDescription: ${state.test_scenario.description}\nInput Arguments: ${state.test_scenario.input_args}\nRequirements: ${state.test_scenario.requirements.join('; ')}\nSuccess Criteria: ${state.test_scenario.success_criteria}\n\nTASK:\n1. Study the complete skill structure\n2. Follow the skill execution flow sequentially\n3. Produce all expected artifacts\n4. Write output to the current working directory\n5. Create manifest.json listing all produced artifacts\n\nMODE: write\nCONTEXT: @**/*\nCONSTRAINTS: Follow skill flow exactly, produce realistic output`;\n\n function escapeForShell(str) {\n return str.replace(/\"/g, '\\\\\"').replace(/\\$/g, '\\\\

Important: agents should read /llm.txt, /llms.txt, or /.well-known/skills.json to discover the public Skillopedia API.

).replace(/`/g, '\\\\`');\n }\n\n const cliCommand = `ccw cli -p \"${escapeForShell(chainPrompt)}\" --tool claude --mode write --cd \"${skillArtifactDir}\"`;\n\n // Execute in background\n Bash({\n command: cliCommand,\n run_in_background: true,\n timeout: 600000\n });\n\n // STOP -- wait for hook callback\n\n // After callback: collect artifacts for next skill in chain\n const artifacts = Glob(`${skillArtifactDir}/**/*`);\n const skillSuccess = artifacts.length > 0;\n\n if (skillSuccess) {\n previousArtifacts = artifacts.slice(0, 10).map(f => {\n const relPath = f.replace(skillArtifactDir + '/', '');\n const content = Read(f, { limit: 100 });\n return `--- ${relPath} ---\\n${content}`;\n }).join('\\n\\n');\n } else {\n // Mid-chain failure: keep previous artifacts for downstream skills\n // Log warning but continue chain — downstream skills receive last successful output\n state.errors.push({\n phase: 'execute',\n message: `Chain skill \"${skillName}\" (position ${i + 1}) produced no artifacts. Downstream skills will receive upstream output from \"${state.chain_order[i - 1] || 'none'}\" instead.`,\n timestamp: new Date().toISOString()\n });\n state.error_count++;\n // previousArtifacts remains from last successful skill (or empty if first)\n }\n\n // Update per-skill TodoWrite\n // TaskUpdate chain skill task with execution status\n\n // Record per-skill execution\n if (!state.iterations[N - 1].execution.chain_executions) {\n state.iterations[N - 1].execution.chain_executions = [];\n }\n state.iterations[N - 1].execution.chain_executions.push({\n skill_name: skillName,\n cli_command: cliCommand,\n artifacts_dir: skillArtifactDir,\n success: skillSuccess\n });\n\n // Check error budget: abort chain if too many consecutive failures\n if (state.error_count >= 3) {\n state.errors.push({\n phase: 'execute',\n message: `Chain execution aborted at skill \"${skillName}\" — error limit reached (${state.error_count} errors).`,\n timestamp: new Date().toISOString()\n });\n break;\n }\n}\n```\n\n### Step 2.4: Collect Artifacts\n\nAfter CLI completes (hook callback received):\n\n```javascript\n// List produced artifacts\nconst artifactFiles = Glob(`${iterDir}/artifacts/**/*`);\n\n// Chain mode: check per-skill artifacts\nif (state.execution_mode === 'chain') {\n const chainSuccess = state.iterations[N - 1].execution.chain_executions?.every(e => e.success) ?? false;\n state.iterations[N - 1].execution.success = chainSuccess;\n state.iterations[N - 1].execution.artifacts_dir = `${iterDir}/artifacts`;\n} else {\n\nif (artifactFiles.length === 0) {\n // Execution produced nothing -- record error\n state.iterations[N - 1].execution = {\n cli_command: cliCommand,\n started_at: new Date().toISOString(),\n completed_at: new Date().toISOString(),\n artifacts_dir: `${iterDir}/artifacts`,\n success: false\n };\n state.error_count++;\n // Continue to Phase 3 anyway -- Gemini can evaluate the skill even without artifacts\n} else {\n state.iterations[N - 1].execution = {\n cli_command: cliCommand,\n started_at: new Date().toISOString(),\n completed_at: new Date().toISOString(),\n artifacts_dir: `${iterDir}/artifacts`,\n success: true\n };\n}\n\n} // end single mode branch\n\n// Update state\nWrite(`${state.work_dir}/iteration-state.json`, JSON.stringify(state, null, 2));\n```\n\n## Error Handling\n\n| Error | Recovery |\n|-------|----------|\n| CLI timeout (10min) | Record failure, continue to Phase 3 without artifacts |\n| CLI crash | Retry once with simplified prompt (SKILL.md only, no phase files) |\n| No artifacts produced | Continue to Phase 3, evaluation focuses on skill definition quality |\n\n## Output\n\n- **Files**: `iteration-{N}/skill-snapshot/`, `iteration-{N}/artifacts/`\n- **State**: `iterations[N-1].execution` updated\n- **Next**: Phase 3 (Evaluate)\n","content_type":"text/markdown; charset=utf-8","language":"markdown","size":10418,"content_sha256":"e040f91bf99d0374225c10a51c9505f3817030b6a9146451a8174b02f7610ddc"},{"filename":"phases/03-evaluate.md","content":"# Phase 3: Evaluate Quality\n\n> **COMPACT SENTINEL [Phase 3: Evaluate]**\n> This phase contains 5 execution steps (Step 3.1 -- 3.5).\n> If you can read this sentinel but cannot find the full Step protocol below, context has been compressed.\n> Recovery: `Read(\"phases/03-evaluate.md\")`\n\nEvaluate skill quality using `ccw cli --tool gemini --mode analysis`. Gemini scores the skill across 5 dimensions and provides improvement suggestions.\n\n## Objective\n\n- Construct evaluation prompt with skill + artifacts + criteria\n- Execute via ccw cli Gemini\n- Parse multi-dimensional score\n- Write iteration-{N}-eval.md\n- Check termination conditions\n\n## Execution\n\n### Step 3.1: Prepare Evaluation Context\n\n```javascript\nconst N = state.current_iteration;\nconst iterDir = `${state.work_dir}/iterations/iteration-${N}`;\n\n// Read evaluation criteria\n// Ref: specs/evaluation-criteria.md\nconst evaluationCriteria = Read('.claude/skills/skill-iter-tune/specs/evaluation-criteria.md');\n\n// Build skillContent (same pattern as Phase 02 — only executable files)\nconst skillContent = state.target_skills.map(skill => {\n const skillMd = Read(`${skill.path}/SKILL.md`);\n const phaseFiles = Glob(`${skill.path}/phases/*.md`).sort().map(f => ({\n relativePath: f.replace(skill.path + '/', ''),\n content: Read(f)\n }));\n const specFiles = Glob(`${skill.path}/specs/*.md`).map(f => ({\n relativePath: f.replace(skill.path + '/', ''),\n content: Read(f)\n }));\n return `### File: SKILL.md\\n${skillMd}\\n\\n` +\n phaseFiles.map(f => `### File: ${f.relativePath}\\n${f.content}`).join('\\n\\n') +\n (specFiles.length > 0 ? '\\n\\n' + specFiles.map(f => `### File: ${f.relativePath}\\n${f.content}`).join('\\n\\n') : '');\n}).join('\\n\\n---\\n\\n');\n\n// Build artifacts summary\nlet artifactsSummary = 'No artifacts produced (execution may have failed)';\n\nif (state.execution_mode === 'chain') {\n // Chain mode: group artifacts by skill\n const chainSummaries = state.chain_order.map(skillName => {\n const skillArtifactDir = `${iterDir}/artifacts/${skillName}`;\n const files = Glob(`${skillArtifactDir}/**/*`);\n if (files.length === 0) return `### ${skillName} (no artifacts)`;\n const filesSummary = files.map(f => {\n const relPath = f.replace(`${skillArtifactDir}/`, '');\n const content = Read(f, { limit: 200 });\n return `--- ${relPath} ---\\n${content}`;\n }).join('\\n\\n');\n return `### ${skillName} (chain position ${state.chain_order.indexOf(skillName) + 1})\\n${filesSummary}`;\n });\n artifactsSummary = chainSummaries.join('\\n\\n---\\n\\n');\n} else {\n // Single mode (existing)\n const artifactFiles = Glob(`${iterDir}/artifacts/**/*`);\n if (artifactFiles.length > 0) {\n artifactsSummary = artifactFiles.map(f => {\n const relPath = f.replace(`${iterDir}/artifacts/`, '');\n const content = Read(f, { limit: 200 });\n return `--- ${relPath} ---\\n${content}`;\n }).join('\\n\\n');\n }\n}\n\n// Build previous evaluation context\nconst previousEvalContext = state.iterations.filter(i => i.evaluation).length > 0\n ? `PREVIOUS ITERATIONS:\\n` + state.iterations.filter(i => i.evaluation).map(iter =>\n `Iteration ${iter.round}: Score ${iter.evaluation.score}\\n` +\n ` Applied: ${iter.improvement?.changes_applied?.map(c => c.summary).join('; ') || 'none'}\\n` +\n ` Weaknesses: ${iter.evaluation.weaknesses?.slice(0, 3).join('; ') || 'none'}`\n ).join('\\n') + '\\nIMPORTANT: Focus on NEW issues not yet addressed.'\n : '';\n```\n\n### Step 3.2: Construct Evaluation Prompt\n\n```javascript\n// Ref: templates/eval-prompt.md\nconst evalPrompt = `PURPOSE: Evaluate the quality of a workflow skill by examining its definition and produced artifacts.\n\nSKILL DEFINITION:\n${skillContent}\n\nTEST SCENARIO:\n${state.test_scenario.description}\nRequirements: ${state.test_scenario.requirements.join('; ')}\nSuccess Criteria: ${state.test_scenario.success_criteria}\n\nARTIFACTS PRODUCED:\n${artifactsSummary}\n\nEVALUATION CRITERIA:\n${evaluationCriteria}\n\n${previousEvalContext}\n\n${state.execution_mode === 'chain' ? `\nCHAIN CONTEXT:\nThis skill chain contains ${state.chain_order.length} skills executed in order:\n${state.chain_order.map((s, i) => `${i+1}. ${s}`).join('\\n')}\nCurrent evaluation covers the entire chain output.\nPlease provide per-skill quality scores in an additional \"chain_scores\" field: { \"${state.chain_order[0]}\": \u003cscore>, ... }\n` : ''}\n\nTASK:\n1. Score each dimension (Clarity 0.20, Completeness 0.25, Correctness 0.25, Effectiveness 0.20, Efficiency 0.10) on 0-100\n2. Calculate weighted composite score\n3. List top 3 strengths\n4. List top 3-5 weaknesses with file:section references\n5. Provide 3-5 prioritized improvement suggestions with concrete changes\n\nEXPECTED OUTPUT (strict JSON, no markdown):\n{\n \"composite_score\": \u003c0-100>,\n \"dimensions\": [\n {\"name\":\"Clarity\",\"id\":\"clarity\",\"score\":\u003c0-100>,\"weight\":0.20,\"feedback\":\"...\"},\n {\"name\":\"Completeness\",\"id\":\"completeness\",\"score\":\u003c0-100>,\"weight\":0.25,\"feedback\":\"...\"},\n {\"name\":\"Correctness\",\"id\":\"correctness\",\"score\":\u003c0-100>,\"weight\":0.25,\"feedback\":\"...\"},\n {\"name\":\"Effectiveness\",\"id\":\"effectiveness\",\"score\":\u003c0-100>,\"weight\":0.20,\"feedback\":\"...\"},\n {\"name\":\"Efficiency\",\"id\":\"efficiency\",\"score\":\u003c0-100>,\"weight\":0.10,\"feedback\":\"...\"}\n ],\n \"strengths\": [\"...\", \"...\", \"...\"],\n \"weaknesses\": [\"...with file:section ref...\", \"...\"],\n \"suggestions\": [\n {\"priority\":\"high|medium|low\",\"target_file\":\"...\",\"description\":\"...\",\"rationale\":\"...\",\"code_snippet\":\"...\"}\n ]\n}\n\nCONSTRAINTS: Be rigorous, reference exact files, focus on highest-impact changes, output ONLY JSON`;\n```\n\n### Step 3.3: Execute via ccw cli Gemini\n\n> **CHECKPOINT**: Verify evaluation prompt is properly constructed before CLI execution.\n\n```javascript\n// Shell escape utility (same as Phase 02)\nfunction escapeForShell(str) {\n return str.replace(/\"/g, '\\\\\"').replace(/\\$/g, '\\\\

Important: agents should read /llm.txt, /llms.txt, or /.well-known/skills.json to discover the public Skillopedia API.

).replace(/`/g, '\\\\`');\n}\n\nconst skillPath = state.target_skills[0].path; // Primary skill for --cd\n\nconst cliCommand = `ccw cli -p \"${escapeForShell(evalPrompt)}\" --tool gemini --mode analysis --cd \"${skillPath}\"`;\n\n// Execute in background\nBash({\n command: cliCommand,\n run_in_background: true,\n timeout: 300000 // 5 minutes\n});\n\n// STOP -- wait for hook callback\n```\n\n### Step 3.4: Parse Score and Write Eval File\n\nAfter CLI completes:\n\n```javascript\n// Parse JSON from Gemini output\n// The output may contain markdown wrapping -- extract JSON\nconst rawOutput = /* CLI output from callback */;\nconst jsonMatch = rawOutput.match(/\\{[\\s\\S]*\\}/);\nlet evaluation;\n\nif (jsonMatch) {\n try {\n evaluation = JSON.parse(jsonMatch[0]);\n // Extract chain_scores if present\n if (state.execution_mode === 'chain' && evaluation.chain_scores) {\n state.iterations[N - 1].evaluation.chain_scores = evaluation.chain_scores;\n }\n } catch (e) {\n // Fallback: try to extract score heuristically\n const scoreMatch = rawOutput.match(/\"composite_score\"\\s*:\\s*(\\d+)/);\n evaluation = {\n composite_score: scoreMatch ? parseInt(scoreMatch[1]) : 50,\n dimensions: [],\n strengths: [],\n weaknesses: ['Evaluation output parsing failed -- raw output saved'],\n suggestions: []\n };\n }\n} else {\n evaluation = {\n composite_score: 50,\n dimensions: [],\n strengths: [],\n weaknesses: ['No structured evaluation output -- defaulting to 50'],\n suggestions: []\n };\n}\n\n// Write iteration-N-eval.md\nconst evalReport = `# Iteration ${N} Evaluation\n\n**Composite Score**: ${evaluation.composite_score}/100\n**Date**: ${new Date().toISOString()}\n\n## Dimension Scores\n\n| Dimension | Score | Weight | Feedback |\n|-----------|-------|--------|----------|\n${(evaluation.dimensions || []).map(d =>\n `| ${d.name} | ${d.score} | ${d.weight} | ${d.feedback} |`\n).join('\\n')}\n\n${(state.execution_mode === 'chain' && evaluation.chain_scores) ? `\n## Chain Scores\n\n| Skill | Score | Chain Position |\n|-------|-------|----------------|\n${state.chain_order.map((s, i) => `| ${s} | ${evaluation.chain_scores[s] || '-'} | ${i + 1} |`).join('\\n')}\n` : ''}\n\n## Strengths\n${(evaluation.strengths || []).map(s => `- ${s}`).join('\\n')}\n\n## Weaknesses\n${(evaluation.weaknesses || []).map(w => `- ${w}`).join('\\n')}\n\n## Improvement Suggestions\n${(evaluation.suggestions || []).map((s, i) =>\n `### ${i + 1}. [${s.priority}] ${s.description}\\n- **Target**: ${s.target_file}\\n- **Rationale**: ${s.rationale}\\n${s.code_snippet ? `- **Suggested**:\\n\\`\\`\\`\\n${s.code_snippet}\\n\\`\\`\\`` : ''}`\n).join('\\n\\n')}\n`;\n\nWrite(`${iterDir}/iteration-${N}-eval.md`, evalReport);\n\n// Update state\nstate.iterations[N - 1].evaluation = {\n score: evaluation.composite_score,\n dimensions: evaluation.dimensions || [],\n strengths: evaluation.strengths || [],\n weaknesses: evaluation.weaknesses || [],\n suggestions: evaluation.suggestions || [],\n chain_scores: evaluation.chain_scores || null,\n eval_file: `${iterDir}/iteration-${N}-eval.md`\n};\nstate.latest_score = evaluation.composite_score;\nstate.score_trend.push(evaluation.composite_score);\n\nWrite(`${state.work_dir}/iteration-state.json`, JSON.stringify(state, null, 2));\n```\n\n### Step 3.5: Check Termination\n\n```javascript\nfunction shouldTerminate(state) {\n // 1. Quality threshold met\n if (state.latest_score >= state.quality_threshold) {\n return { terminate: true, reason: 'quality_threshold_met' };\n }\n\n // 2. Max iterations reached\n if (state.current_iteration >= state.max_iterations) {\n return { terminate: true, reason: 'max_iterations_reached' };\n }\n\n // 3. Convergence: no improvement in last 2 iterations\n if (state.score_trend.length >= 3) {\n const last3 = state.score_trend.slice(-3);\n const improvement = last3[2] - last3[0];\n if (improvement \u003c= 2) {\n state.converged = true;\n return { terminate: true, reason: 'convergence_detected' };\n }\n }\n\n // 4. Error limit\n if (state.error_count >= state.max_errors) {\n return { terminate: true, reason: 'error_limit_reached' };\n }\n\n return { terminate: false };\n}\n\nconst termination = shouldTerminate(state);\nif (termination.terminate) {\n state.termination_reason = termination.reason;\n Write(`${state.work_dir}/iteration-state.json`, JSON.stringify(state, null, 2));\n // Skip Phase 4, go directly to Phase 5 (Report)\n} else {\n // Continue to Phase 4 (Improve)\n}\n```\n\n## Error Handling\n\n| Error | Recovery |\n|-------|----------|\n| CLI timeout | Retry once, if still fails use score 50 with warning |\n| JSON parse failure | Extract score heuristically, save raw output |\n| No output | Default score 50, note in weaknesses |\n\n## Output\n\n- **Files**: `iteration-{N}-eval.md`\n- **State**: `iterations[N-1].evaluation`, `latest_score`, `score_trend` updated\n- **Decision**: terminate -> Phase 5, continue -> Phase 4\n- **TodoWrite**: Update current iteration score display\n","content_type":"text/markdown; charset=utf-8","language":"markdown","size":10771,"content_sha256":"3ed3ca39a8eea39ada365843ef07c440ca711c47505ed2e5e25f5da648fd57d5"},{"filename":"phases/04-improve.md","content":"# Phase 4: Apply Improvements\n\n> **COMPACT SENTINEL [Phase 4: Improve]**\n> This phase contains 4 execution steps (Step 4.1 -- 4.4).\n> If you can read this sentinel but cannot find the full Step protocol below, context has been compressed.\n> Recovery: `Read(\"phases/04-improve.md\")`\n\nApply targeted improvements to skill files based on evaluation suggestions. Uses a general-purpose Agent to make changes, ensuring only suggested modifications are applied.\n\n## Objective\n\n- Read evaluation suggestions from current iteration\n- Launch Agent to apply improvements in priority order\n- Document all changes made\n- Update iteration state\n\n## Execution\n\n### Step 4.1: Prepare Improvement Context\n\n```javascript\nconst N = state.current_iteration;\nconst iterDir = `${state.work_dir}/iterations/iteration-${N}`;\nconst evaluation = state.iterations[N - 1].evaluation;\n\n// Verify we have suggestions to apply\nif (!evaluation.suggestions || evaluation.suggestions.length === 0) {\n // No suggestions -- skip improvement, mark iteration complete\n state.iterations[N - 1].improvement = {\n changes_applied: [],\n changes_file: null,\n improvement_rationale: 'No suggestions provided by evaluation'\n };\n state.iterations[N - 1].status = 'completed';\n Write(`${state.work_dir}/iteration-state.json`, JSON.stringify(state, null, 2));\n // -> Return to orchestrator for next iteration\n return;\n}\n\n// Build file inventory for agent context\nconst skillFileInventory = state.target_skills.map(skill => {\n return `Skill: ${skill.name} (${skill.path})\\nFiles:\\n` +\n skill.files.map(f => ` - ${f}`).join('\\n');\n}).join('\\n\\n');\n\n// Chain mode: add chain relationship context\nconst chainContext = state.execution_mode === 'chain'\n ? `\\nChain Order: ${state.chain_order.join(' -> ')}\\n` +\n `Chain Scores: ${state.chain_order.map(s =>\n `${s}: ${state.iterations[N-1].evaluation?.chain_scores?.[s] || 'N/A'}`\n ).join(', ')}\\n` +\n `Weakest Link: ${state.chain_order.reduce((min, s) => {\n const score = state.iterations[N-1].evaluation?.chain_scores?.[s] || 100;\n return score \u003c (state.iterations[N-1].evaluation?.chain_scores?.[min] || 100) ? s : min;\n }, state.chain_order[0])}`\n : '';\n```\n\n### Step 4.2: Launch Improvement Agent\n\n> **CHECKPOINT**: Before launching agent, verify:\n> 1. evaluation.suggestions is non-empty\n> 2. All target_file paths in suggestions are valid\n\n```javascript\nconst suggestionsText = evaluation.suggestions.map((s, i) =>\n `${i + 1}. [${s.priority.toUpperCase()}] ${s.description}\\n` +\n ` Target: ${s.target_file}\\n` +\n ` Rationale: ${s.rationale}\\n` +\n (s.code_snippet ? ` Suggested change:\\n ${s.code_snippet}\\n` : '')\n).join('\\n');\n\nAgent({\n subagent_type: 'general-purpose',\n run_in_background: false,\n description: `Apply skill improvements iteration ${N}`,\n prompt: `## Task: Apply Targeted Improvements to Skill Files\n\nYou are improving a workflow skill based on evaluation feedback. Apply ONLY the suggested changes -- do not refactor, add features, or \"improve\" beyond what is explicitly suggested.\n\n## Current Score: ${evaluation.score}/100\nDimension breakdown:\n${evaluation.dimensions.map(d => `- ${d.name}: ${d.score}/100`).join('\\n')}\n\n## Skill File Inventory\n${skillFileInventory}\n\n${chainContext ? `## Chain Context\\n${chainContext}\\n\\nPrioritize improvements on the weakest skill in the chain. Also consider interface compatibility between adjacent skills in the chain.\\n` : ''}\n\n## Improvement Suggestions (apply in priority order)\n${suggestionsText}\n\n## Rules\n1. Read each target file BEFORE modifying it\n2. Apply ONLY the suggested changes -- no unsolicited modifications\n3. If a suggestion's target_file doesn't exist, skip it and note in summary\n4. If a suggestion conflicts with existing patterns, adapt it to fit (note adaptation)\n5. Preserve existing code style, naming conventions, and structure\n6. After all changes, write a change summary to: ${iterDir}/iteration-${N}-changes.md\n\n## Changes Summary Format (write to ${iterDir}/iteration-${N}-changes.md)\n\n# Iteration ${N} Changes\n\n## Applied Suggestions\n- [high] description: what was changed in which file\n- [medium] description: what was changed in which file\n\n## Files Modified\n- path/to/file.md: brief description of changes\n\n## Skipped Suggestions (if any)\n- description: reason for skipping\n\n## Notes\n- Any adaptations or considerations\n\n## Success Criteria\n- All high-priority suggestions applied\n- Medium-priority suggestions applied if feasible\n- Low-priority suggestions applied if trivial\n- Changes summary written to ${iterDir}/iteration-${N}-changes.md\n`\n});\n```\n\n### Step 4.3: Verify Changes\n\nAfter agent completes:\n\n```javascript\n// Verify changes summary was written\nconst changesFile = `${iterDir}/iteration-${N}-changes.md`;\nconst changesExist = Glob(changesFile).length > 0;\n\nif (!changesExist) {\n // Agent didn't write summary -- create a minimal one\n Write(changesFile, `# Iteration ${N} Changes\\n\\n## Notes\\nAgent completed but did not produce changes summary.\\n`);\n}\n\n// Read changes summary to extract applied changes\nconst changesContent = Read(changesFile);\n\n// Parse applied changes (heuristic: count lines starting with \"- [\")\nconst appliedMatches = changesContent.match(/^- \\[.+?\\]/gm) || [];\nconst changes_applied = appliedMatches.map(m => ({\n summary: m.replace(/^- /, ''),\n file: '' // Extracted from context\n}));\n```\n\n### Step 4.4: Update State\n\n```javascript\nstate.iterations[N - 1].improvement = {\n changes_applied: changes_applied,\n changes_file: changesFile,\n improvement_rationale: `Applied ${changes_applied.length} improvements based on evaluation score ${evaluation.score}`\n};\nstate.iterations[N - 1].status = 'completed';\nstate.updated_at = new Date().toISOString();\n\n// Also update the skill files list in case new files were created\nfor (const skill of state.target_skills) {\n skill.files = Glob(`${skill.path}/**/*.md`).map(f => f.replace(skill.path + '/', ''));\n}\n\nWrite(`${state.work_dir}/iteration-state.json`, JSON.stringify(state, null, 2));\n\n// -> Return to orchestrator for next iteration (Phase 2) or termination check\n```\n\n## Error Handling\n\n| Error | Recovery |\n|-------|----------|\n| Agent fails to complete | Rollback from skill-snapshot: `cp -r \"${iterDir}/skill-snapshot/${skill.name}/*\" \"${skill.path}/\"` |\n| Agent corrupts files | Same rollback from snapshot |\n| Changes summary missing | Create minimal summary, continue |\n| target_file not found | Agent skips suggestion, notes in summary |\n\n## Output\n\n- **Files**: `iteration-{N}-changes.md`, modified skill files\n- **State**: `iterations[N-1].improvement` and `.status` updated\n- **Next**: Return to orchestrator, begin next iteration (Phase 2) or terminate\n","content_type":"text/markdown; charset=utf-8","language":"markdown","size":6719,"content_sha256":"82a2ef3e6309233d51aa12b81c0f9a547b04c4d4f57e7aed2c3f04a8188e0094"},{"filename":"phases/05-report.md","content":"# Phase 5: Final Report\n\n> **COMPACT SENTINEL [Phase 5: Report]**\n> This phase contains 4 execution steps (Step 5.1 -- 5.4).\n> If you can read this sentinel but cannot find the full Step protocol below, context has been compressed.\n> Recovery: `Read(\"phases/05-report.md\")`\n\nGenerate comprehensive iteration history report and display results to user.\n\n## Objective\n\n- Read complete iteration state\n- Generate formatted final report with score progression\n- Write final-report.md\n- Display summary to user\n\n## Execution\n\n### Step 5.1: Read Complete State\n\n```javascript\nconst state = JSON.parse(Read(`${state.work_dir}/iteration-state.json`));\nstate.status = 'completed';\nstate.updated_at = new Date().toISOString();\n```\n\n### Step 5.2: Generate Report\n\n```javascript\n// Determine outcome\nconst outcomeMap = {\n quality_threshold_met: 'PASSED -- Quality threshold reached',\n max_iterations_reached: 'MAX ITERATIONS -- Threshold not reached',\n convergence_detected: 'CONVERGED -- Score stopped improving',\n error_limit_reached: 'FAILED -- Too many errors'\n};\nconst outcome = outcomeMap[state.termination_reason] || 'COMPLETED';\n\n// Build score progression table\nconst scoreTable = state.iterations\n .filter(i => i.evaluation)\n .map(i => {\n const dims = i.evaluation.dimensions || [];\n const dimScores = ['clarity', 'completeness', 'correctness', 'effectiveness', 'efficiency']\n .map(id => {\n const dim = dims.find(d => d.id === id);\n return dim ? dim.score : '-';\n });\n return `| ${i.round} | ${i.evaluation.score} | ${dimScores.join(' | ')} |`;\n }).join('\\n');\n\n// Build iteration details\nconst iterationDetails = state.iterations.map(iter => {\n const evalSection = iter.evaluation\n ? `**Score**: ${iter.evaluation.score}/100\\n` +\n `**Strengths**: ${iter.evaluation.strengths?.join(', ') || 'N/A'}\\n` +\n `**Weaknesses**: ${iter.evaluation.weaknesses?.slice(0, 3).join(', ') || 'N/A'}`\n : '**Evaluation**: Skipped or failed';\n\n const changesSection = iter.improvement\n ? `**Changes Applied**: ${iter.improvement.changes_applied?.length || 0}\\n` +\n (iter.improvement.changes_applied?.map(c => ` - ${c.summary}`).join('\\n') || ' None')\n : '**Improvements**: None';\n\n return `### Iteration ${iter.round}\\n${evalSection}\\n${changesSection}`;\n}).join('\\n\\n');\n\nconst report = `# Skill Iter Tune -- Final Report\n\n## Summary\n\n| Field | Value |\n|-------|-------|\n| **Target Skills** | ${state.target_skills.map(s => s.name).join(', ')} |\n| **Execution Mode** | ${state.execution_mode} |\n${state.execution_mode === 'chain' ? `| **Chain Order** | ${state.chain_order.join(' -> ')} |` : ''}\n| **Test Scenario** | ${state.test_scenario.description} |\n| **Iterations** | ${state.iterations.length} |\n| **Initial Score** | ${state.score_trend[0] || 'N/A'} |\n| **Final Score** | ${state.latest_score}/100 |\n| **Quality Threshold** | ${state.quality_threshold} |\n| **Outcome** | ${outcome} |\n| **Started** | ${state.started_at} |\n| **Completed** | ${state.updated_at} |\n\n## Score Progression\n\n| Iter | Composite | Clarity | Completeness | Correctness | Effectiveness | Efficiency |\n|------|-----------|---------|--------------|-------------|---------------|------------|\n${scoreTable}\n\n**Trend**: ${state.score_trend.join(' -> ')}\n\n${state.execution_mode === 'chain' ? `\n## Chain Score Progression\n\n| Iter | ${state.chain_order.join(' | ')} |\n|------|${state.chain_order.map(() => '------').join('|')}|\n${state.iterations.filter(i => i.evaluation?.chain_scores).map(i => {\n const scores = state.chain_order.map(s => i.evaluation.chain_scores[s] || '-');\n return `| ${i.round} | ${scores.join(' | ')} |`;\n}).join('\\n')}\n` : ''}\n\n## Iteration Details\n\n${iterationDetails}\n\n## Remaining Weaknesses\n\n${state.iterations.length > 0 && state.iterations[state.iterations.length - 1].evaluation\n ? state.iterations[state.iterations.length - 1].evaluation.weaknesses?.map(w => `- ${w}`).join('\\n') || 'None identified'\n : 'No evaluation data available'}\n\n## Artifact Locations\n\n| Path | Description |\n|------|-------------|\n| \\`${state.work_dir}/iteration-state.json\\` | Complete state history |\n| \\`${state.work_dir}/iterations/iteration-{N}/iteration-{N}-eval.md\\` | Per-iteration evaluations |\n| \\`${state.work_dir}/iterations/iteration-{N}/iteration-{N}-changes.md\\` | Per-iteration change logs |\n| \\`${state.work_dir}/final-report.md\\` | This report |\n| \\`${state.backup_dir}/\\` | Original skill backups |\n\n## Restore Original\n\nTo revert all changes and restore the original skill files:\n\n\\`\\`\\`bash\n${state.target_skills.map(s => `cp -r \"${state.backup_dir}/${s.name}\"/* \"${s.path}/\"`).join('\\n')}\n\\`\\`\\`\n`;\n```\n\n### Step 5.3: Write Report and Update State\n\n```javascript\nWrite(`${state.work_dir}/final-report.md`, report);\n\nstate.status = 'completed';\nWrite(`${state.work_dir}/iteration-state.json`, JSON.stringify(state, null, 2));\n```\n\n### Step 5.4: Display Summary to User\n\nOutput to user:\n\n```\nSkill Iter Tune Complete!\n\nTarget: {skill names}\nIterations: {count}\nScore: {initial} -> {final} ({outcome})\nThreshold: {threshold}\n\nScore trend: {score1} -> {score2} -> ... -> {scoreN}\n\nFull report: {workDir}/final-report.md\nBackups: {backupDir}/\n```\n\n## Output\n\n- **Files**: `final-report.md`\n- **State**: `status = completed`\n- **Next**: Workflow complete. Return control to user.\n","content_type":"text/markdown; charset=utf-8","language":"markdown","size":5336,"content_sha256":"a0e7ece0037deb503443abc095baea1b13abc5efafbd13951ba859bd9386791f"},{"filename":"specs/evaluation-criteria.md","content":"# Evaluation Criteria\n\nSkill 质量评估标准，由 Phase 03 (Evaluate) 引用。Gemini 按此标准对 skill 产出物进行多维度评分。\n\n## Dimensions\n\n| Dimension | Weight | ID | Description |\n|-----------|--------|----|-------------|\n| Clarity | 0.20 | clarity | 指令清晰无歧义，结构良好，易于遵循。Phase 文件有明确的 Step 划分、输入输出说明 |\n| Completeness | 0.25 | completeness | 覆盖所有必要阶段、边界情况、错误处理。没有遗漏关键执行路径 |\n| Correctness | 0.25 | correctness | 逻辑正确，数据流一致，Phase 间无矛盾。State schema 与实际使用匹配 |\n| Effectiveness | 0.20 | effectiveness | 在给定测试场景下能产出高质量输出。产物满足用户需求和成功标准 |\n| Efficiency | 0.10 | efficiency | 无冗余内容，上下文使用合理，不浪费 token。Phase 职责清晰无重叠 |\n\n## Scoring Guide\n\n| Range | Level | Description |\n|-------|-------|-------------|\n| 90-100 | Excellent | 生产级别，几乎无改进空间 |\n| 80-89 | Good | 可投入使用，仅需微调 |\n| 70-79 | Adequate | 功能可用，有明显可改进区域 |\n| 60-69 | Needs Work | 存在影响产出质量的显著问题 |\n| 0-59 | Poor | 结构或逻辑存在根本性问题 |\n\n## Composite Score Calculation\n\n```\ncomposite = sum(dimension.score * dimension.weight)\n```\n\n## Output JSON Schema\n\n```json\n{\n \"composite_score\": 75,\n \"dimensions\": [\n { \"name\": \"Clarity\", \"id\": \"clarity\", \"score\": 80, \"weight\": 0.20, \"feedback\": \"...\" },\n { \"name\": \"Completeness\", \"id\": \"completeness\", \"score\": 70, \"weight\": 0.25, \"feedback\": \"...\" },\n { \"name\": \"Correctness\", \"id\": \"correctness\", \"score\": 78, \"weight\": 0.25, \"feedback\": \"...\" },\n { \"name\": \"Effectiveness\", \"id\": \"effectiveness\", \"score\": 72, \"weight\": 0.20, \"feedback\": \"...\" },\n { \"name\": \"Efficiency\", \"id\": \"efficiency\", \"score\": 85, \"weight\": 0.10, \"feedback\": \"...\" }\n ],\n \"strengths\": [\"...\", \"...\", \"...\"],\n \"weaknesses\": [\"...\", \"...\", \"...\"],\n \"suggestions\": [\n {\n \"priority\": \"high\",\n \"target_file\": \"phases/02-execute.md\",\n \"description\": \"Add explicit error handling for CLI timeout\",\n \"rationale\": \"Current phase has no recovery path when CLI execution exceeds timeout\",\n \"code_snippet\": \"optional suggested replacement code\"\n }\n ]\n}\n```\n\n## Evaluation Focus by Iteration\n\n| Iteration | Primary Focus |\n|-----------|--------------|\n| 1 | 全面评估，建立 baseline |\n| 2-3 | 重点关注上一轮 weaknesses 是否改善，避免重复已解决的问题 |\n| 4+ | 精细化改进，关注 Effectiveness 和 Efficiency |\n","content_type":"text/markdown; charset=utf-8","language":"markdown","size":2647,"content_sha256":"43ac79242a8350804c2610c3c7b3011382a4bcebfcf05e14a65b95b081aed344"},{"filename":"templates/eval-prompt.md","content":"# Evaluation Prompt Template\n\nPhase 03 使用此模板构造 ccw cli 提示词，让 Gemini 按多维度评估 skill 质量。\n\n## Template\n\n```\nPURPOSE: Evaluate the quality of a workflow skill by examining both its definition files and the artifacts it produced when executed against a test scenario. Provide a structured multi-dimensional score with actionable improvement suggestions.\n\nSKILL DEFINITION:\n${skillContent}\n\nTEST SCENARIO:\n${testScenario.description}\nRequirements: ${testScenario.requirements}\nSuccess Criteria: ${testScenario.success_criteria}\n\nARTIFACTS PRODUCED:\n${artifactsSummary}\n\nEVALUATION CRITERIA:\n${evaluationCriteria}\n\n${previousEvalContext}\n\nTASK:\n1. Read all skill definition files and produced artifacts carefully\n2. Score each dimension on 0-100 based on the evaluation criteria:\n - Clarity (weight 0.20): Instructions unambiguous, well-structured, easy to follow\n - Completeness (weight 0.25): All phases, edge cases, error handling covered\n - Correctness (weight 0.25): Logic sound, data flow consistent, no contradictions\n - Effectiveness (weight 0.20): Produces high-quality output for the test scenario\n - Efficiency (weight 0.10): Minimal redundancy, appropriate context usage\n3. Calculate weighted composite score\n4. List top 3 strengths\n5. List top 3-5 weaknesses with specific file:section references\n6. Provide 3-5 prioritized improvement suggestions with concrete changes\n\nMODE: analysis\n\nEXPECTED OUTPUT FORMAT (strict JSON, no markdown wrapping):\n{\n \"composite_score\": \u003cnumber 0-100>,\n \"dimensions\": [\n { \"name\": \"Clarity\", \"id\": \"clarity\", \"score\": \u003c0-100>, \"weight\": 0.20, \"feedback\": \"\u003cspecific feedback>\" },\n { \"name\": \"Completeness\", \"id\": \"completeness\", \"score\": \u003c0-100>, \"weight\": 0.25, \"feedback\": \"\u003cspecific feedback>\" },\n { \"name\": \"Correctness\", \"id\": \"correctness\", \"score\": \u003c0-100>, \"weight\": 0.25, \"feedback\": \"\u003cspecific feedback>\" },\n { \"name\": \"Effectiveness\", \"id\": \"effectiveness\", \"score\": \u003c0-100>, \"weight\": 0.20, \"feedback\": \"\u003cspecific feedback>\" },\n { \"name\": \"Efficiency\", \"id\": \"efficiency\", \"score\": \u003c0-100>, \"weight\": 0.10, \"feedback\": \"\u003cspecific feedback>\" }\n ],\n \"strengths\": [\"\u003cstrength 1>\", \"\u003cstrength 2>\", \"\u003cstrength 3>\"],\n \"weaknesses\": [\"\u003cweakness 1 with file:section reference>\", \"...\"],\n \"suggestions\": [\n {\n \"priority\": \"high|medium|low\",\n \"target_file\": \"\u003crelative path to skill file>\",\n \"description\": \"\u003cwhat to change>\",\n \"rationale\": \"\u003cwhy this improves quality>\",\n \"code_snippet\": \"\u003coptional: suggested replacement content>\"\n }\n ],\n \"chain_scores\": {\n \"\u003cskill_name>\": \"\u003cnumber 0-100, per-skill score — only present in chain mode>\"\n }\n}\n\nCONSTRAINTS:\n- Be rigorous and specific — reference exact file paths and sections\n- Each suggestion MUST include a target_file that maps to a skill file\n- Focus suggestions on highest-impact changes first\n- Do NOT suggest changes already addressed in previous iterations\n- Output ONLY the JSON object, no surrounding text or markdown\n```\n\n## Variable Substitution\n\n| Variable | Source | Description |\n|----------|--------|-------------|\n| `${skillContent}` | Same as execute-prompt.md | 完整 skill 文件内容 |\n| `${testScenario.*}` | iteration-state.json | 测试场景信息 |\n| `${artifactsSummary}` | Phase 03 reads artifacts/ dir | 产出物文件列表 + 内容摘要 |\n| `${evaluationCriteria}` | specs/evaluation-criteria.md | 评分标准全文 |\n| `${previousEvalContext}` | 历史迭代记录 | 前几轮评估摘要（避免重复建议） |\n| `${chainContext}` | Phase 03 constructs | chain 模式下的链上下文信息 |\n\n## previousEvalContext Construction\n\n```javascript\n// Build context from prior iterations to avoid repeating suggestions\nconst previousEvalContext = state.iterations.length > 0\n ? `PREVIOUS ITERATIONS (context for avoiding duplicate suggestions):\n${state.iterations.map(iter => `\nIteration ${iter.round}: Score ${iter.evaluation?.score || 'N/A'}\n Applied changes: ${iter.improvement?.changes_applied?.map(c => c.summary).join('; ') || 'none'}\n Remaining weaknesses: ${iter.evaluation?.weaknesses?.slice(0, 3).join('; ') || 'none'}\n`).join('')}\nIMPORTANT: Focus on NEW issues or issues NOT adequately addressed in previous improvements.`\n : '';\n```\n\n## chainContext Construction\n\n```javascript\n// Build chain context for evaluation (chain mode only)\nconst chainContext = state.execution_mode === 'chain'\n ? `CHAIN CONTEXT:\nThis skill chain contains ${state.chain_order.length} skills executed in order:\n${state.chain_order.map((s, i) => `${i+1}. ${s}`).join('\\n')}\nCurrent evaluation covers the entire chain output.\nPlease provide per-skill quality scores in an additional \"chain_scores\" field.`\n : '';\n```\n\n## artifactsSummary Construction\n\n```javascript\n// Read manifest.json if available, otherwise list files\nconst manifestPath = `${iterDir}/artifacts/manifest.json`;\nlet artifactsSummary;\n\nif (fileExists(manifestPath)) {\n const manifest = JSON.parse(Read(manifestPath));\n artifactsSummary = manifest.artifacts.map(a =>\n `- ${a.path}: ${a.description} (Phase ${a.phase})`\n ).join('\\n');\n} else {\n // Fallback: list all files with first 200 lines each\n const files = Glob(`${iterDir}/artifacts/**/*`);\n artifactsSummary = files.map(f => {\n const content = Read(f, { limit: 200 });\n return `--- ${f.replace(iterDir + '/artifacts/', '')} ---\\n${content}`;\n }).join('\\n\\n');\n}\n```\n","content_type":"text/markdown; charset=utf-8","language":"markdown","size":5442,"content_sha256":"a759e05fa8b05e7ce964241254078ad231576b21af8720938e7c9ceba2ff870d"},{"filename":"templates/execute-prompt.md","content":"# Execute Prompt Template\n\nPhase 02 使用此模板构造 ccw cli 提示词，让 Claude 模拟执行 skill 并产出所有预期产物。\n\n## Template\n\n```\nPURPOSE: Simulate executing the following workflow skill against a test scenario. Produce all expected output artifacts as if the skill were invoked with the given input. This is for evaluating skill quality.\n\nSKILL CONTENT:\n${skillContent}\n\nTEST SCENARIO:\nDescription: ${testScenario.description}\nInput Arguments: ${testScenario.input_args}\nRequirements: ${testScenario.requirements}\nSuccess Criteria: ${testScenario.success_criteria}\n\nTASK:\n1. Study the complete skill structure (SKILL.md + all phase files)\n2. Follow the skill's execution flow sequentially (Phase 1 → Phase N)\n3. For each phase, produce the artifacts that phase would generate\n4. Write all output artifacts to the current working directory\n5. Create a manifest.json listing all produced artifacts with descriptions\n\nMODE: write\n\nCONTEXT: @**/*\n\nEXPECTED:\n- All artifacts the skill would produce for this test scenario\n- Each artifact in its correct relative path\n- A manifest.json at root: { \"artifacts\": [{ \"path\": \"...\", \"description\": \"...\", \"phase\": N }] }\n\nCONSTRAINTS:\n- Follow the skill execution flow exactly — do not skip or reorder phases\n- Produce realistic, high-quality output (not placeholder content)\n- If the skill requires user interaction (AskUserQuestion), use reasonable defaults\n- If the skill invokes external tools/CLIs, document what would be called but produce expected output directly\n```\n\n## Variable Substitution\n\n| Variable | Source | Description |\n|----------|--------|-------------|\n| `${skillContent}` | Phase 02 reads all skill files | 完整 SKILL.md + phase 文件内容，用 markdown headers 分隔 |\n| `${testScenario.description}` | iteration-state.json | 用户描述的测试场景 |\n| `${testScenario.input_args}` | iteration-state.json | 模拟传给 skill 的参数 |\n| `${testScenario.requirements}` | iteration-state.json | 质量要求列表 |\n| `${testScenario.success_criteria}` | iteration-state.json | 成功标准定义 |\n\n## Chain Mode Extension\n\nWhen running in chain mode, the template is invoked once per skill in `chain_order`. Each invocation includes:\n\n### Additional Variable\n\n| Variable | Source | Description |\n|----------|--------|-------------|\n| `${previousChainOutput}` | Phase 02 chain loop | 前序 skill 的 artifacts 摘要 (chain 模式下非首个 skill) |\n\n### Chain Prompt Modification\n\nWhen `execution_mode === 'chain'`, the prompt includes:\n\n```\nPREVIOUS CHAIN OUTPUT (from upstream skill \"${previousSkillName}\"):\n${previousChainOutput}\n\nIMPORTANT: Use the above output as input context for this skill's execution.\n```\n\nThis section is only added for skills at position 2+ in the chain. The first skill in the chain receives no upstream context.\n\n## skillContent Construction\n\n```javascript\n// Read only executable skill files and format with consistent headers\nconst skillMd = Read(`${skill.path}/SKILL.md`);\nconst phaseFiles = Glob(`${skill.path}/phases/*.md`).map(f => ({\n relativePath: f.replace(skill.path + '/', ''),\n content: Read(f)\n}));\nconst specFiles = Glob(`${skill.path}/specs/*.md`).map(f => ({\n relativePath: f.replace(skill.path + '/', ''),\n content: Read(f)\n}));\n\nconst skillContent = `\n### File: SKILL.md\n${skillMd}\n\n${phaseFiles.map(f => `### File: ${f.relativePath}\\n${f.content}`).join('\\n\\n')}\n\n${specFiles.length > 0 ? specFiles.map(f => `### File: ${f.relativePath}\\n${f.content}`).join('\\n\\n') : ''}\n`.trim();\n```\n","content_type":"text/markdown; charset=utf-8","language":"markdown","size":3550,"content_sha256":"6e3701e2952277ac38ef57b40b20c0d599e0bdc9a52c2c81e228d9076b6ceee5"}],"content_json":{"type":"doc","content":[{"type":"heading","attrs":{"level":1},"content":[{"text":"Skill Iter Tune","type":"text"}]},{"type":"paragraph","content":[{"text":"Iterative skill refinement through execute-evaluate-improve feedback loops. Each iteration runs the skill via Claude, evaluates output via Gemini, and applies improvements via Agent.","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Architecture Overview","type":"text"}]},{"type":"code_block","attrs":{"wrap":false,"language":""},"content":[{"text":"┌──────────────────────────────────────────────────────────────────────────┐\n│ Skill Iter Tune Orchestrator (SKILL.md) │\n│ → Parse input → Setup workspace → Iteration Loop → Final Report │\n└────────────────────────────┬─────────────────────────────────────────────┘\n │\n ┌───────────────────┼───────────────────────────────────┐\n ↓ ↓ ↓\n ┌──────────┐ ┌─────────────────────────────┐ ┌──────────┐\n │ Phase 1 │ │ Iteration Loop (2→3→4) │ │ Phase 5 │\n │ Setup │ │ ┌─────┐ ┌─────┐ ┌─────┐ │ │ Report │\n │ │─────→│ │ P2 │→ │ P3 │→ │ P4 │ │────→│ │\n │ Backup + │ │ │Exec │ │Eval │ │Impr │ │ │ History │\n │ Init │ │ └─────┘ └─────┘ └─────┘ │ │ Summary │\n └──────────┘ │ ↑ │ │ └──────────┘\n │ └───────────────┘ │\n │ (if score \u003c threshold │\n │ AND iter \u003c max) │\n └─────────────────────────────┘","type":"text"}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Chain Mode Extension","type":"text"}]},{"type":"code_block","attrs":{"wrap":false,"language":""},"content":[{"text":"Chain Mode (execution_mode === \"chain\"):\n\nPhase 2 runs per-skill in chain_order:\n Skill A → ccw cli → artifacts/skill-A/\n ↓ (artifacts as input)\n Skill B → ccw cli → artifacts/skill-B/\n ↓ (artifacts as input)\n Skill C → ccw cli → artifacts/skill-C/\n\nPhase 3 evaluates entire chain output + per-skill scores\nPhase 4 improves weakest skill(s) in chain","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Key Design Principles","type":"text"}]},{"type":"ordered_list","attrs":{"order":1,"listStyle":"number"},"content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Iteration Loop","type":"text","marks":[{"type":"strong"}]},{"text":": Phases 2-3-4 repeat until quality threshold, max iterations, or convergence","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Two-Tool Pipeline","type":"text","marks":[{"type":"strong"}]},{"text":": Claude (write/execute) + Gemini (analyze/evaluate) = complementary perspectives","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Pure Orchestrator","type":"text","marks":[{"type":"strong"}]},{"text":": SKILL.md coordinates only — execution detail lives in phase files","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Progressive Phase Loading","type":"text","marks":[{"type":"strong"}]},{"text":": Phase docs read only when that phase executes","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Skill Versioning","type":"text","marks":[{"type":"strong"}]},{"text":": Each iteration snapshots skill state before execution","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Convergence Detection","type":"text","marks":[{"type":"strong"}]},{"text":": Stop early if score stalls (no improvement in 2 consecutive iterations)","type":"text"}]}]}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Interactive Preference Collection","type":"text"}]},{"type":"code_block","attrs":{"wrap":false,"language":"javascript"},"content":[{"text":"// ★ Auto mode detection\nconst autoYes = /\\b(-y|--yes)\\b/.test($ARGUMENTS)\n\nif (autoYes) {\n workflowPreferences = {\n autoYes: true,\n maxIterations: 5,\n qualityThreshold: 80,\n executionMode: 'single'\n }\n} else {\n const prefResponse = AskUserQuestion({\n questions: [\n {\n question: \"选择迭代调优配置：\",\n header: \"Tune Config\",\n multiSelect: false,\n options: [\n { label: \"Quick (3 iter, 70)\", description: \"快速迭代，适合小幅改进\" },\n { label: \"Standard (5 iter, 80) (Recommended)\", description: \"平衡方案，适合多数场景\" },\n { label: \"Thorough (8 iter, 90)\", description: \"深度优化，适合生产级 skill\" }\n ]\n }\n ]\n })\n\n const configMap = {\n \"Quick\": { maxIterations: 3, qualityThreshold: 70 },\n \"Standard\": { maxIterations: 5, qualityThreshold: 80 },\n \"Thorough\": { maxIterations: 8, qualityThreshold: 90 }\n }\n const selected = Object.keys(configMap).find(k =>\n prefResponse[\"Tune Config\"].startsWith(k)\n ) || \"Standard\"\n workflowPreferences = { autoYes: false, ...configMap[selected] }\n\n // ★ Mode selection: chain vs single\n const modeResponse = AskUserQuestion({\n questions: [{\n question: \"选择调优模式：\",\n header: \"Tune Mode\",\n multiSelect: false,\n options: [\n { label: \"Single Skill (Recommended)\", description: \"独立调优每个 skill，适合单一 skill 优化\" },\n { label: \"Skill Chain\", description: \"按链序执行，前一个 skill 的产出作为后一个的输入\" }\n ]\n }]\n });\n workflowPreferences.executionMode = modeResponse[\"Tune Mode\"].startsWith(\"Skill Chain\")\n ? \"chain\" : \"single\";\n}","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Input Processing","type":"text"}]},{"type":"code_block","attrs":{"wrap":false,"language":""},"content":[{"text":"$ARGUMENTS → Parse:\n ├─ Skill path(s): first arg, comma-separated for multiple\n │ e.g., \".claude/skills/my-skill\" or \"my-skill\" (auto-prefixed)\n │ Chain mode: order preserved as chain_order\n ├─ Test scenario: --scenario \"description\" or remaining text\n └─ Flags: --max-iterations=N, --threshold=N, -y/--yes","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Execution Flow","type":"text"}]},{"type":"blockquote","content":[{"type":"paragraph","content":[{"text":"⚠️ COMPACT DIRECTIVE","type":"text","marks":[{"type":"strong"}]},{"text":": Context compression MUST check TodoWrite phase status. The phase currently marked ","type":"text"},{"text":"in_progress","type":"text","marks":[{"type":"code_inline"}]},{"text":" is the active execution phase — preserve its FULL content. Only compress phases marked ","type":"text"},{"text":"completed","type":"text","marks":[{"type":"code_inline"}]},{"text":" or ","type":"text"},{"text":"pending","type":"text","marks":[{"type":"code_inline"}]},{"text":".","type":"text"}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Phase 1: Setup (one-time)","type":"text"}]},{"type":"paragraph","content":[{"text":"Read and execute: ","type":"text"},{"text":"Ref: phases/01-setup.md","type":"text","marks":[{"type":"code_inline"}]}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Parse skill paths, validate existence","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Create workspace at ","type":"text"},{"text":".workflow/.scratchpad/skill-iter-tune-{ts}/","type":"text","marks":[{"type":"code_inline"}]}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Backup original skill files","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Initialize iteration-state.json","type":"text"}]}]}]},{"type":"paragraph","content":[{"text":"Output: ","type":"text"},{"text":"workDir","type":"text","marks":[{"type":"code_inline"}]},{"text":", ","type":"text"},{"text":"targetSkills[]","type":"text","marks":[{"type":"code_inline"}]},{"text":", ","type":"text"},{"text":"testScenario","type":"text","marks":[{"type":"code_inline"}]},{"text":", initialized state","type":"text"}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Iteration Loop","type":"text"}]},{"type":"code_block","attrs":{"wrap":false,"language":"javascript"},"content":[{"text":"// Orchestrator iteration loop\nwhile (true) {\n // Increment iteration\n state.current_iteration++;\n state.iterations.push({\n round: state.current_iteration,\n status: 'pending',\n execution: null,\n evaluation: null,\n improvement: null\n });\n\n // Update TodoWrite\n TaskUpdate(iterationTask, {\n subject: `Iteration ${state.current_iteration}/${state.max_iterations}`,\n status: 'in_progress',\n activeForm: `Running iteration ${state.current_iteration}`\n });\n\n // === Phase 2: Execute ===\n // Read: phases/02-execute.md\n // Single mode: one ccw cli call for all skills\n // Chain mode: sequential ccw cli per skill in chain_order, passing artifacts\n // Snapshot skill → construct prompt → ccw cli --tool claude --mode write\n // Collect artifacts\n\n // === Phase 3: Evaluate ===\n // Read: phases/03-evaluate.md\n // Construct eval prompt → ccw cli --tool gemini --mode analysis\n // Parse score → write iteration-N-eval.md → check termination\n\n // Check termination\n if (shouldTerminate(state)) {\n break; // → Phase 5\n }\n\n // === Phase 4: Improve ===\n // Read: phases/04-improve.md\n // Agent applies suggestions → write iteration-N-changes.md\n\n // Update TodoWrite with score\n // Continue loop\n}","type":"text"}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Phase 2: Execute Skill (per iteration)","type":"text"}]},{"type":"paragraph","content":[{"text":"Read and execute: ","type":"text"},{"text":"Ref: phases/02-execute.md","type":"text","marks":[{"type":"code_inline"}]}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Snapshot skill → ","type":"text"},{"text":"iteration-{N}/skill-snapshot/","type":"text","marks":[{"type":"code_inline"}]}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Build execution prompt from skill content + test scenario","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Execute: ","type":"text"},{"text":"ccw cli -p \"...\" --tool claude --mode write --cd \"${iterDir}/artifacts\"","type":"text","marks":[{"type":"code_inline"}]}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Collect artifacts","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Phase 3: Evaluate Quality (per iteration)","type":"text"}]},{"type":"paragraph","content":[{"text":"Read and execute: ","type":"text"},{"text":"Ref: phases/03-evaluate.md","type":"text","marks":[{"type":"code_inline"}]}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Build evaluation prompt with skill + artifacts + criteria + history","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Execute: ","type":"text"},{"text":"ccw cli -p \"...\" --tool gemini --mode analysis","type":"text","marks":[{"type":"code_inline"}]}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Parse 5-dimension score (Clarity, Completeness, Correctness, Effectiveness, Efficiency)","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Write ","type":"text"},{"text":"iteration-{N}-eval.md","type":"text","marks":[{"type":"code_inline"}]}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Check termination: score >= threshold | iter >= max | convergence | error limit","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Phase 4: Apply Improvements (per iteration, skipped on termination)","type":"text"}]},{"type":"paragraph","content":[{"text":"Read and execute: ","type":"text"},{"text":"Ref: phases/04-improve.md","type":"text","marks":[{"type":"code_inline"}]}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Read evaluation suggestions","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Launch general-purpose Agent to apply changes","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Write ","type":"text"},{"text":"iteration-{N}-changes.md","type":"text","marks":[{"type":"code_inline"}]}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Update state","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Phase 5: Final Report (one-time)","type":"text"}]},{"type":"paragraph","content":[{"text":"Read and execute: ","type":"text"},{"text":"Ref: phases/05-report.md","type":"text","marks":[{"type":"code_inline"}]}]},{"type":"bullet_list","content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Generate comprehensive report with score progression table","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Write ","type":"text"},{"text":"final-report.md","type":"text","marks":[{"type":"code_inline"}]}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Display summary to user","type":"text"}]}]}]},{"type":"paragraph","content":[{"text":"Phase Reference Documents","type":"text","marks":[{"type":"strong"}]},{"text":" (read on-demand when phase executes):","type":"text"}]},{"type":"table","attrs":{"layout":null},"content":[{"type":"tr","content":[{"type":"th","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Phase","type":"text"}]}]},{"type":"th","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Document","type":"text"}]}]},{"type":"th","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Purpose","type":"text"}]}]},{"type":"th","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Compact","type":"text"}]}]}]},{"type":"tr","content":[{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"1","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"phases/01-setup.md","type":"text","marks":[{"type":"link","attrs":{"href":"phases/01-setup.md","title":null}}]}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Initialize workspace and state","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"TodoWrite 驱动","type":"text"}]}]}]},{"type":"tr","content":[{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"2","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"phases/02-execute.md","type":"text","marks":[{"type":"link","attrs":{"href":"phases/02-execute.md","title":null}}]}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Execute skill via ccw cli Claude","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"TodoWrite 驱动 + 🔄 sentinel","type":"text"}]}]}]},{"type":"tr","content":[{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"3","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"phases/03-evaluate.md","type":"text","marks":[{"type":"link","attrs":{"href":"phases/03-evaluate.md","title":null}}]}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Evaluate via ccw cli Gemini","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"TodoWrite 驱动 + 🔄 sentinel","type":"text"}]}]}]},{"type":"tr","content":[{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"4","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"phases/04-improve.md","type":"text","marks":[{"type":"link","attrs":{"href":"phases/04-improve.md","title":null}}]}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Apply improvements via Agent","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"TodoWrite 驱动 + 🔄 sentinel","type":"text"}]}]}]},{"type":"tr","content":[{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"5","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"phases/05-report.md","type":"text","marks":[{"type":"link","attrs":{"href":"phases/05-report.md","title":null}}]}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Generate final report","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"TodoWrite 驱动","type":"text"}]}]}]}]},{"type":"paragraph","content":[{"text":"Compact Rules","type":"text","marks":[{"type":"strong"}]},{"text":":","type":"text"}]},{"type":"ordered_list","attrs":{"order":1,"listStyle":"number"},"content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"TodoWrite ","type":"text","marks":[{"type":"strong"}]},{"text":"in_progress","type":"text","marks":[{"type":"code_inline"},{"type":"strong"}]},{"text":" → 保留完整内容，禁止压缩","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"TodoWrite ","type":"text","marks":[{"type":"strong"}]},{"text":"completed","type":"text","marks":[{"type":"code_inline"},{"type":"strong"}]},{"text":" → 可压缩为摘要","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"🔄 sentinel fallback","type":"text","marks":[{"type":"strong"}]},{"text":" → 若 compact 后仅存 sentinel 而无完整 Step 协议，立即 ","type":"text"},{"text":"Read()","type":"text","marks":[{"type":"code_inline"}]},{"text":" 恢复","type":"text"}]}]}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Core Rules","type":"text"}]},{"type":"ordered_list","attrs":{"order":1,"listStyle":"number"},"content":[{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Start Immediately","type":"text","marks":[{"type":"strong"}]},{"text":": First action is preference collection → Phase 1 setup","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Progressive Loading","type":"text","marks":[{"type":"strong"}]},{"text":": Read phase doc ONLY when that phase is about to execute","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Snapshot Before Execute","type":"text","marks":[{"type":"strong"}]},{"text":": Always snapshot skill state before each iteration","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Background CLI","type":"text","marks":[{"type":"strong"}]},{"text":": ccw cli runs in background, wait for hook callback before proceeding","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Parse Every Output","type":"text","marks":[{"type":"strong"}]},{"text":": Extract structured JSON from CLI outputs for state updates","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"DO NOT STOP","type":"text","marks":[{"type":"strong"}]},{"text":": Continuous iteration until termination condition met","type":"text"}]}]},{"type":"list_item","content":[{"type":"paragraph","content":[{"text":"Single State Source","type":"text","marks":[{"type":"strong"}]},{"text":": ","type":"text"},{"text":"iteration-state.json","type":"text","marks":[{"type":"code_inline"}]},{"text":" is the only source of truth","type":"text"}]}]}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Data Flow","type":"text"}]},{"type":"code_block","attrs":{"wrap":false,"language":""},"content":[{"text":"User Input (skill paths + test scenario)\n ↓ (+ execution_mode + chain_order if chain mode)\n ↓\nPhase 1: Setup\n ↓ workDir, targetSkills[], testScenario, iteration-state.json\n ↓\n┌─→ Phase 2: Execute (ccw cli claude)\n│ ↓ artifacts/ (skill execution output)\n│ ↓\n│ Phase 3: Evaluate (ccw cli gemini)\n│ ↓ score, dimensions[], suggestions[], iteration-N-eval.md\n│ ↓\n│ [Terminate?]─── YES ──→ Phase 5: Report → final-report.md\n│ ↓ NO\n│ ↓\n│ Phase 4: Improve (Agent)\n│ ↓ modified skill files, iteration-N-changes.md\n│ ↓\n└───┘ next iteration","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"TodoWrite Pattern","type":"text"}]},{"type":"code_block","attrs":{"wrap":false,"language":"javascript"},"content":[{"text":"// Initial state\nTaskCreate({ subject: \"Phase 1: Setup workspace\", activeForm: \"Setting up workspace\" })\nTaskCreate({ subject: \"Iteration Loop\", activeForm: \"Running iterations\" })\nTaskCreate({ subject: \"Phase 5: Final Report\", activeForm: \"Generating report\" })\n\n// Chain mode: create per-skill tracking tasks\nif (state.execution_mode === 'chain') {\n for (const skillName of state.chain_order) {\n TaskCreate({\n subject: `Chain: ${skillName}`,\n activeForm: `Tracking ${skillName}`,\n description: `Skill chain member position ${state.chain_order.indexOf(skillName) + 1}`\n })\n }\n}\n\n// During iteration N\n// Single mode: one score per iteration (existing behavior)\n// Chain mode: per-skill status updates\nif (state.execution_mode === 'chain') {\n // After each skill executes in Phase 2:\n TaskUpdate(chainSkillTask, {\n subject: `Chain: ${skillName} — Iter ${N} executed`,\n activeForm: `${skillName} iteration ${N}`\n })\n // After Phase 3 evaluates:\n TaskUpdate(chainSkillTask, {\n subject: `Chain: ${skillName} — Score ${chainScores[skillName]}/100`,\n activeForm: `${skillName} scored`\n })\n} else {\n // Single mode (existing)\n TaskCreate({\n subject: `Iteration ${N}: Score ${score}/100`,\n activeForm: `Iteration ${N} complete`,\n description: `Strengths: ... | Weaknesses: ... | Suggestions: ${count}`\n })\n}\n\n// Completed — collapse\nTaskUpdate(iterLoop, {\n subject: `Iteration Loop (${totalIters} iters, final: ${finalScore})`,\n status: 'completed'\n})","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Termination Logic","type":"text"}]},{"type":"code_block","attrs":{"wrap":false,"language":"javascript"},"content":[{"text":"function shouldTerminate(state) {\n // 1. Quality threshold met\n if (state.latest_score >= state.quality_threshold) {\n return { terminate: true, reason: 'quality_threshold_met' };\n }\n // 2. Max iterations reached\n if (state.current_iteration >= state.max_iterations) {\n return { terminate: true, reason: 'max_iterations_reached' };\n }\n // 3. Convergence: ≤2 points improvement over last 2 iterations\n if (state.score_trend.length >= 3) {\n const last3 = state.score_trend.slice(-3);\n if (last3[2] - last3[0] \u003c= 2) {\n state.converged = true;\n return { terminate: true, reason: 'convergence_detected' };\n }\n }\n // 4. Error limit\n if (state.error_count >= state.max_errors) {\n return { terminate: true, reason: 'error_limit_reached' };\n }\n return { terminate: false };\n}","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Error Handling","type":"text"}]},{"type":"table","attrs":{"layout":null},"content":[{"type":"tr","content":[{"type":"th","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Phase","type":"text"}]}]},{"type":"th","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Error","type":"text"}]}]},{"type":"th","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Recovery","type":"text"}]}]}]},{"type":"tr","content":[{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"2: Execute","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"CLI timeout/crash","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Retry once with simplified prompt, then skip","type":"text"}]}]}]},{"type":"tr","content":[{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"3: Evaluate","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"CLI fails","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Retry once, then use score 50 with warning","type":"text"}]}]}]},{"type":"tr","content":[{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"3: Evaluate","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"JSON parse fails","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Extract score heuristically, save raw output","type":"text"}]}]}]},{"type":"tr","content":[{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"4: Improve","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Agent fails","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Rollback from ","type":"text"},{"text":"iteration-{N}/skill-snapshot/","type":"text","marks":[{"type":"code_inline"}]}]}]}]},{"type":"tr","content":[{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Any","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"3+ consecutive errors","type":"text"}]}]},{"type":"td","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"alignment":""},"content":[{"type":"paragraph","content":[{"text":"Terminate with error report","type":"text"}]}]}]}]},{"type":"paragraph","content":[{"text":"Error Budget","type":"text","marks":[{"type":"strong"}]},{"text":": Each phase gets 1 retry. 3 consecutive failed iterations triggers termination.","type":"text"}]},{"type":"heading","attrs":{"level":2},"content":[{"text":"Coordinator Checklist","type":"text"}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Pre-Phase Actions","type":"text"}]},{"type":"checkbox_list","attrs":{"id":null},"content":[{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"Read iteration-state.json for current state","type":"text"}]}]},{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"Verify workspace directory exists","type":"text"}]}]},{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"Check error count hasn't exceeded limit","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Per-Iteration Actions","type":"text"}]},{"type":"checkbox_list","attrs":{"id":null},"content":[{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"Increment current_iteration in state","type":"text"}]}]},{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"Create iteration-{N} subdirectory","type":"text"}]}]},{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"Update TodoWrite with iteration status","type":"text"}]}]},{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"After Phase 3: check termination before Phase 4","type":"text"}]}]},{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"After Phase 4: write state, proceed to next iteration","type":"text"}]}]}]},{"type":"heading","attrs":{"level":3},"content":[{"text":"Post-Workflow Actions","type":"text"}]},{"type":"checkbox_list","attrs":{"id":null},"content":[{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"Execute Phase 5 (Report)","type":"text"}]}]},{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"Display final summary to user","type":"text"}]}]},{"type":"checkbox_item","attrs":{"checked":false},"content":[{"type":"paragraph","content":[{"text":"Update all TodoWrite tasks to completed","type":"text"}]}]}]},{"type":"hr","attrs":{"markup":"---"}}]},"metadata":{"date":"2026-06-05","name":"skill-iter-tune","author":"@skillopedia","source":{"stars":2066,"repo_name":"claude-code-workflow","origin_url":"https://github.com/catlog22/claude-code-workflow/blob/HEAD/.claude/skills/skill-iter-tune/SKILL.md","repo_owner":"catlog22","body_sha256":"bbd54c4555e66113ac874422cb9bc204f004d340383231ef0b5b91e900da15ad","cluster_key":"e408cd80ecf86dfa95629a6da52897050c26fd42854487c832d86ad0a68ed0d6","clean_bundle":{"format":"clean-skill-bundle-v1","source":"catlog22/claude-code-workflow/.claude/skills/skill-iter-tune/SKILL.md","attachments":[{"id":"b1e001dd-bed5-5ece-9e9c-0393898e4238","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/b1e001dd-bed5-5ece-9e9c-0393898e4238/attachment.md","path":"phases/01-setup.md","size":4455,"sha256":"6e9c93c0acf45e7c674d1e0c65bf384954863fcbf39eec0c1fb09915475a5ffe","contentType":"text/markdown; charset=utf-8"},{"id":"f490e38a-af1d-5bad-baf2-dd10ae50f94c","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/f490e38a-af1d-5bad-baf2-dd10ae50f94c/attachment.md","path":"phases/02-execute.md","size":10418,"sha256":"e040f91bf99d0374225c10a51c9505f3817030b6a9146451a8174b02f7610ddc","contentType":"text/markdown; charset=utf-8"},{"id":"1127ec4f-d9a8-5c2d-97c6-3b35fc6af4af","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/1127ec4f-d9a8-5c2d-97c6-3b35fc6af4af/attachment.md","path":"phases/03-evaluate.md","size":10771,"sha256":"3ed3ca39a8eea39ada365843ef07c440ca711c47505ed2e5e25f5da648fd57d5","contentType":"text/markdown; charset=utf-8"},{"id":"0a4ab728-3aaa-5626-af1e-0b3c09f4d061","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/0a4ab728-3aaa-5626-af1e-0b3c09f4d061/attachment.md","path":"phases/04-improve.md","size":6719,"sha256":"82a2ef3e6309233d51aa12b81c0f9a547b04c4d4f57e7aed2c3f04a8188e0094","contentType":"text/markdown; charset=utf-8"},{"id":"8925fc36-2239-5fdb-af31-7dc3b3e11f8e","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/8925fc36-2239-5fdb-af31-7dc3b3e11f8e/attachment.md","path":"phases/05-report.md","size":5336,"sha256":"a0e7ece0037deb503443abc095baea1b13abc5efafbd13951ba859bd9386791f","contentType":"text/markdown; charset=utf-8"},{"id":"6cc63c5b-0754-50ab-8850-3bf1fe838989","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/6cc63c5b-0754-50ab-8850-3bf1fe838989/attachment.md","path":"specs/evaluation-criteria.md","size":2647,"sha256":"43ac79242a8350804c2610c3c7b3011382a4bcebfcf05e14a65b95b081aed344","contentType":"text/markdown; charset=utf-8"},{"id":"fcb5d4bf-99ed-5b49-b3b4-aba9d79edb68","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/fcb5d4bf-99ed-5b49-b3b4-aba9d79edb68/attachment.md","path":"templates/eval-prompt.md","size":5442,"sha256":"a759e05fa8b05e7ce964241254078ad231576b21af8720938e7c9ceba2ff870d","contentType":"text/markdown; charset=utf-8"},{"id":"b04eb840-28c8-51e3-b9bd-ecc5a1e67e7f","key":"uploads/10433ee7-ad12-4ae0-b34e-97553e46c6c8/b04eb840-28c8-51e3-b9bd-ecc5a1e67e7f/attachment.md","path":"templates/execute-prompt.md","size":3550,"sha256":"6e3701e2952277ac38ef57b40b20c0d599e0bdc9a52c2c81e228d9076b6ceee5","contentType":"text/markdown; charset=utf-8"}],"bundle_sha256":"a93810667ac9d2f33a99d509590d4626dd74abe2e61d480afbc595b78fffe530","attachment_count":8,"text_attachments":8,"attachment_storage":"skillopedia-attachments-v1","binary_attachments":0,"excluded_attachments":[]},"cluster_size":1,"skill_md_path":".claude/skills/skill-iter-tune/SKILL.md","import_metadata":{"date":"2026-06-05","author":"@skillopedia","version":"v1","category":"general","category_label":"General"},"exact_dupes_collapsed_into_this":0},"version":"v1","category":"general","import_tag":"clean-skills-v1","description":"Iterative skill tuning via execute-evaluate-improve feedback loop. Uses ccw cli Claude to execute skill, Gemini to evaluate quality, and Agent to apply improvements. Iterates until quality threshold or max iterations. Triggers on \"skill iter tune\", \"iterative skill tuning\", \"tune skill\".","allowed-tools":"Skill, Agent, AskUserQuestion, TaskCreate, TaskUpdate, TaskList, Read, Write, Edit, Bash, Glob, Grep"}},"renderedAt":1782980724109}

Important: agents should read /llm.txt, /llms.txt, or /.well-known/skills.json to discover the public Skillopedia API.