Claude Md Optimizer

Agent Instructions Optimizer

Behavioral Classification: Guided Decision (ASK, THEN EXECUTE)

This skill analyzes CLAUDE.md files and proposes optimizations using progressive disclosure patterns. It NEVER makes destructive changes without explicit user approval. Always present the plan, get confirmation, then execute.

Quick Start

Phase 1: Detect & Analyze

1. Scan the repo for agent instruction files:
   • ./CLAUDE.md or ./.claude/CLAUDE.md (Claude Code)
   • ./AGENTS.md and any nested AGENTS.md in subdirectories (OpenAI Codex / GitHub Copilot)
   • .github/copilot-instructions.md and .github/instructions/*.instructions.md (GitHub Copilot)
2. Report each file found with its line count and the format's optimization threshold (see Supported Formats)
3. If multiple files need optimization, confirm with the user which to address first
4. Parse the selected file into sections and categorize each:
   • Essential - Required for every session (build commands, test patterns, security rules)
   • Reference - Needed occasionally (architecture docs, API references, troubleshooting)
   • Redundant - Duplicates existing docs or overly verbose

Phase 2: Detect Constraints

1. Check .gitattributes for filter=git-crypt diff=git-crypt patterns
2. Identify encrypted paths (docs/, .claude/, etc.)
3. Find safe unencrypted directory for sub-documents (prefer project root or subdirs)
4. Scan for CI machine-readable dependencies (see constraints.md)
5. Identify the format's native sub-doc mechanism (see Supported Formats and constraints.md)

Phase 3: Plan

Present categorization table to user:

| Section | Lines | Category | Target | Rationale |
|---------|-------|----------|--------|-----------|
| Docker Commands | 45 | Reference | docker-guide.md | Used <10% of sessions |
| CI Health | 120 | Reference | ci-reference.md | Detailed reference rarely needed |
| Git-Crypt Setup | 30 | Essential | Keep inline | Unlock instructions (chicken-and-egg) |

For CLAUDE.md: also show whether sections can use @import or move to .claude/rules/. For copilot-instructions.md: show whether sections are candidates for path-scoped .instructions.md files.

Wait for user approval before proceeding.

Phase 4: Review

User confirms:

• Categorization is correct
• Sub-document names make sense
• Essential content remains inline
• Safe directory for sub-docs identified

Phase 5: Execute

1. Create sub-documents in safe directory (verbatim extraction, NO summarization)
2. For CLAUDE.md: use @path/to/file imports or .claude/rules/*.md as the extraction target when possible — these are native to the format and load at session start
3. For AGENTS.md: use nested AGENTS.md files in subdirectories for subsystem-specific content; use .agents/instructions/<ref>.md + "For <reason>, see [.agents/instructions/<ref>.md]" links for general reference content
4. For copilot-instructions.md: extract language/framework rules to .github/instructions/NAME.instructions.md with applyTo: frontmatter; use .github/copilot/<ref>.md + "For <reason>, see [.github/copilot/<ref>.md]" links for reference content that isn't path-specific
5. Rewrite the main file: replace each extracted section with a rich inline synopsis followed by a link (Pattern 2). The synopsis must contain enough concrete specifics (framework names, thresholds, ports, naming conventions) that an agent can answer common questions without opening the sub-doc. The link is a trapdoor for when more detail is needed.
6. Optionally add a reference table at the end of the file if there are many sub-docs and an index adds navigational value (see Pattern 1)
7. Preserve all Essential content inline

Phase 6: Validate

1. Count lines before/after, report reduction vs. format target (see Supported Formats)
2. Diff all extracted content to verify zero information loss
3. Confirm encryption constraints respected
4. Present validation report to user

Supported Formats

CLAUDE.md

Anthropic recommends <200 lines — above this, instruction adherence degrades. Two native extraction paths avoid the overhead of linked sub-docs:

• @path/to/file imports: Add @docs/development/docker-guide.md anywhere in CLAUDE.md; the file loads into context at session start, equivalent to inline content
• .claude/rules/*.md: Topic-specific rule files with optional paths: frontmatter to scope rules to file types (only loads when Claude works with matching files, saving context)

AGENTS.md

Plain Markdown, no required schema. Codex and GitHub Copilot load files hierarchically (global → repo root → subdirectory → CWD); files closer to CWD take precedence. Codex truncates the combined chain at a configurable budget (project_doc_max_bytes, default 32 KiB), so keep the chain within that budget. In Codex, AGENTS.override.md at any level takes precedence over AGENTS.md at the same level (Codex-specific, not part of the cross-tool agents.md spec).

Two progressive disclosure strategies:

• Nested AGENTS.md files: move subsystem-specific rules into services/payments/AGENTS.md etc. — auto-discovered when working in that directory
• .agents/instructions/<ref>.md + contextual link: for reference content that doesn't belong to a specific subsystem, extract to .agents/instructions/<ref>.md and link with "For <reason>, see .agents/instructions/<ref>.md." — the link text carries intent so the agent knows when to follow it. Each sub-doc should include a description: frontmatter field (a one-line summary).

copilot-instructions.md

Repository-wide instructions in .github/copilot-instructions.md (no frontmatter required). Two progressive disclosure strategies:

• Path-scoped .instructions.md files: extract language/framework rules to .github/instructions/NAME.instructions.md with applyTo: frontmatter — load automatically for matching files, no link-following. applyTo: is required; also include a description: field. If the rules are also scoped to specific directories (not just file types), add directory globs to applyTo: or include a paths: field alongside it.
• .github/copilot/<ref>.md + contextual link: for reference content that isn't path-specific, extract to .github/copilot/<ref>.md and link with "For <reason>, see .github/copilot/<ref>.md."

Path-scoped files supplement the repo-wide file — both apply when a file matches.

Sub-Documentation

Essential Rules

Safety First

• ALWAYS present the plan before making changes
• NEVER make destructive edits without explicit user approval
• ALWAYS validate zero information loss after optimization

Format Awareness

• ALWAYS identify the file format before choosing an extraction strategy
• PREFER the format's native sub-doc mechanism (imports, nested files, path-scoped instructions) over generic markdown links
• NEVER move content that is consulted in >50% of sessions

Content Preservation

• NEVER summarize or paraphrase content — verbatim extraction only
• ALWAYS keep encryption unlock instructions in main file (chicken-and-egg problem)
• ALWAYS keep security rules, test file patterns, behavioral reminders inline
• NEVER move content that is consulted in >50% of sessions
• ALWAYS keep content that CI scripts parse via regex (see CI Machine-Readable Content below)

Encryption Safety

• Sub-documents MUST be in unencrypted directories
• Check .gitattributes for filter=git-crypt patterns
• If entire project is encrypted, warn user and recommend decrypted subdirectory

CI Machine-Readable Content

• CI audit scripts may use regex to scan CLAUDE.md for specific literal strings
• ALWAYS scan for scripts that readFileSync('CLAUDE.md') or grep CLAUDE.md
• Common patterns: deploy commands, function lists, config references
• Content matched by CI regex MUST remain in CLAUDE.md — extraction breaks CI
• See constraints.md for detection methodology

Progressive Disclosure

• Replace extracted sections with a rich inline synopsis + link (Pattern 2): 3–5 sentences with concrete specifics (names, thresholds, commands) followed by the link
• The synopsis is the primary value — most agent questions should be answerable from it without opening the sub-doc
• The link is a trapdoor for detail: include what extra information the sub-doc contains so the agent can judge whether to follow it
• Avoid a table of all sub-docs at the top; an upfront index doesn't convey when each file is needed
• Target: >50% line reduction with zero information loss

Quick Reference

Content Tiers

Sub-Document Naming

• Use kebab-case domain names: docker-guide.md, ci-reference.md, api-docs.md
• Group related content: troubleshooting/native-modules.md
• Avoid generic names: reference.md, docs.md

Progressive Disclosure Patterns

1. Sub-Documentation Table - Optional reference index (place at end, not top)
2. Rich Abstract+Link - 3–5 sentence synopsis with concrete specifics + link; agent answers most questions from synopsis alone
3. Compressed Table+Link - Condensed table with full version in sub-doc
4. Summary+Link - Brief paragraph + link to comprehensive guide

Target Metrics

• Essential content remains accessible (inline or via format-native import/reference)
• Session impact: minimal friction for common tasks

Example Output

CLAUDE.md: 850 lines → 175 lines (79% reduction)

• Build commands, security rules, test patterns: kept inline
• Docker guide: extracted to docs/development/docker-guide.md, imported with @docs/development/docker-guide.md
• CI rules: moved to .claude/rules/ci.md (path-scoped to **/.github/**)
• Architecture links: replaced with single index link

AGENTS.md: 420 lines → 110 lines root + nested subdirectory files (74% reduction)

• Root: core conventions, build commands, PR rules
• services/payments/AGENTS.md: payments-specific rules (65 lines)
• services/auth/AGENTS.md: auth-specific rules (45 lines)

copilot-instructions.md: 300 lines → 85 lines root + path-scoped files (72% reduction)

• Root: project overview, build commands, PR conventions
• .github/instructions/typescript.instructions.md (applyTo: **/*.ts,**/*.tsx)
• .github/instructions/python.instructions.md (applyTo: **/*.py)

Validation Checklist

• All extracted content is byte-for-byte identical (verified via diff)
• Sub-documents are in unencrypted directory (or use format's native mechanism)
• Essential content remains accessible (inline or imported/linked)
• Extracted sections replaced with inline contextual references ("For <reason>, see [file.md]") at the point of relevance
• All relative links and @import paths work
• User approved the plan before execution
• Line count is within format target (see Supported Formats)
• No chicken-and-egg problems (encryption, authentication)
• CI scripts that regex-scan the instruction file still pass (see CI Machine-Readable Content)

Error Handling

Entire project is encrypted: Warn user, suggest creating docs/claude/ or similar unencrypted subdirectory.

User rejects plan: Ask for feedback, iterate on categorization.

Information loss detected: Abort optimization, present diff to user, do not commit changes.

Ambiguous categorization: Default to Essential (keep inline), ask user to review.

File already at target size: Report analysis but decline to optimize. Thresholds: CLAUDE.md <200 lines, AGENTS.md <200 lines, copilot-instructions.md <150 lines.

Notes

• This skill is project-agnostic and works with any of the three supported instruction file formats
• Always read the full file before categorizing (context matters)
• Encryption constraints are project-specific, always check .gitattributes
• Validation is mandatory, never skip Phase 6
• For CLAUDE.md: prefer @import and .claude/rules/ over generic markdown sub-docs — they're the format's native mechanism
• For AGENTS.md: nested AGENTS.md files in subdirectories keep subsystem context close to the code
• For copilot-instructions.md: path-scoped .instructions.md files reduce noise by only loading rules relevant to the file being edited