From agents-initializer
Evaluates and improves existing AGENTS.md files in projects. Identifies bloat, contradictions, stale references, and missing scopes — then applies progressive disclosure refactoring based on the ETH Zurich study and context engineering research.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agents-initializer:improve-agentsThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Evaluate existing AGENTS.md files against evidence-based criteria and apply improvements. ETH Zurich (Feb 2026): bloated files reduce success ~3% and increase cost ~20% — operative rubric is the Deletion Test in `evaluation-criteria.md`.
assets/templates/claude-rule.mdassets/templates/domain-doc.mdassets/templates/hook-config.mdassets/templates/root-agents-md.mdassets/templates/scoped-agents-md.mdreferences/automation-mechanism-comparison.mdreferences/automation-token-impact.mdreferences/context-optimization.mdreferences/evaluation-criteria.mdreferences/improvement-card-template.mdreferences/progressive-disclosure-guide.mdreferences/validation-criteria.mdreferences/what-not-to-include.mdEvaluate existing AGENTS.md files against evidence-based criteria and apply improvements. ETH Zurich (Feb 2026): bloated files reduce success ~3% and increase cost ~20% — operative rubric is the Deletion Test in evaluation-criteria.md.
Read ${CLAUDE_SKILL_DIR}/references/evaluation-criteria.md for the scoring rubric, bloat indicators, staleness detection, and Deletion Test.
Delegate to the file-evaluator agent:
Evaluate all AGENTS.md files at the current working directory. Check files >200 lines, bloat indicators, stale references, contradictions, progressive disclosure opportunities, and missing scope-specific files. Return a structured assessment with line numbers and content per issue. Flag automation candidates as
HOOK_CANDIDATE,RULE_CANDIDATE,SKILL_CANDIDATE, orSUBAGENT_CANDIDATE.
The agent runs on Sonnet with read-only tools in an isolated context. Wait for it and parse the structured output.
Delegate to the codebase-analyzer agent:
Analyze the project at the current working directory. Verify tooling commands in AGENTS.md still work; identify scopes with distinct tooling but lacking AGENTS.md; detect new domain areas not covered. Return ONLY actionable findings.
Wait for it and parse the structured output.
Drop any references from Phases 1–2. Read these references:
${CLAUDE_SKILL_DIR}/references/progressive-disclosure-guide.md — hierarchy decisions${CLAUDE_SKILL_DIR}/references/what-not-to-include.md — exclusion criteria + Deletion Test wordingIdentify all Removals (bloat / stale / duplicates / contradictions) and Redundancy Eliminations (apply the Deletion Test to each instruction; document content removed and the cited evidence source).
Drop Phase 3a references. Read this reference:
${CLAUDE_SKILL_DIR}/references/context-optimization.md — token budget guidelinesPlan Refactoring (scope extraction / domain extraction / consolidation) and Additions (only when genuinely missing). Use templates: root-agents-md.md, scoped-agents-md.md, domain-doc.md, claude-rule.md, skill.md, hook-config.md (all under ${CLAUDE_SKILL_DIR}/assets/templates/). In calibrated mode (overall score ≥ 7, no hard-limit violations), keep suggestions proportional to confirmed issues — no speculative restructure or new files.
Drop Phase 3b references. Read these references:
${CLAUDE_SKILL_DIR}/references/automation-mechanism-comparison.md — mechanism selection (always)${CLAUDE_SKILL_DIR}/references/automation-token-impact.md — only when Phase 1 returned automation candidatesApply the decision flowchart in automation-mechanism-comparison.md to each flagged candidate. Merge into the Refactoring category. Skip automation-token-impact.md if Phase 1 returned no automation candidates.
Read ${CLAUDE_SKILL_DIR}/references/validation-criteria.md and execute its Validation Loop Instructions against every improved or newly created file. Apply the "If This Is an IMPROVE Operation" section (information preservation, custom-command retention, structure preserved). In calibrated high-quality cases (score ≥ 7, no hard limits), treat unrelated structural churn as a validation failure — revert and choose the smaller fix. Maximum 3 iterations.
Read ${CLAUDE_SKILL_DIR}/references/improvement-card-template.md and follow it exactly.
npx claudepluginhub rodrigorjsf/agent-engineering-toolkit --plugin agents-initializerProvides UI/UX resources: 50+ styles, color palettes, font pairings, guidelines, charts for web/mobile across React, Next.js, Vue, Svelte, Tailwind, React Native, Flutter. Aids planning, building, reviewing interfaces.
Fetches up-to-date documentation from Context7 for libraries and frameworks like React, Next.js, Prisma. Use for setup questions, API references, and code examples.