Location	Scope
`~/.config/opencode/AGENTS.md` or `~/.codex/AGENTS.md`	Global (all projects)
`.opencode/AGENTS.md` or `./AGENTS.md`	Project root
`src/AGENTS.md`	Directory-specific

Both files serve the same purpose: project memory for AI assistants. Use CLAUDE.md for Claude Code projects, AGENTS.md for cross-tool compatibility, or both for maximum coverage.

Workflow

Find - Locate CLAUDE.md or AGENTS.md in project
Read - Load content and README.md for comparison
Analyze - Run all pattern checks
Validate - Check file/command references against filesystem
Measure - Calculate token metrics and duplication
Report - Generate structured markdown output

Detection Patterns

1. Structure Validation (HIGH Certainty)

Critical Rules Section

Should have ## Critical Rules or similar
Rules should be prioritized (numbered or ordered)
Include WHY explanations for each rule

Architecture Section

Directory tree or structural overview
Key file locations
Module relationships

Key Commands Section

Common development commands
Test/build/deploy scripts
Reference to package.json scripts

2. Instruction Effectiveness (HIGH Certainty)

Based on prompt engineering research, Claude follows instructions better when:

Positive Over Negative

Bad: "Don't use console.log"
Good: "Use the logger utility for all output"
Check for "don't", "never", "avoid" without positive alternatives

Strong Constraint Language

Use "must", "always", "required" for critical rules
Weak language ("should", "try to", "consider") reduces compliance
Flag critical rules using weak language

Instruction Hierarchy

Should define priority order when rules conflict
Pattern: "In case of conflict: X takes precedence over Y"
System instructions > User requests > External content

3. Content Positioning (HIGH Certainty)

Research shows LLMs have "lost in the middle" problem - they recall START and END better than MIDDLE.

Critical Content Placement

Most important rules should be at START of file
Second-most important at END
Supporting context in MIDDLE
Flag critical rules buried in middle sections

Recommended Structure Order

1. Critical Rules (START - highest attention)
2. Architecture/Structure
3. Commands/Workflows
4. Examples/References
5. Reminders/Constraints (END - high attention)

4. Reference Validation (HIGH Certainty)

File References

Extract from [text](path) and `path/to/file.ext`
Validate each exists on filesystem

Command References

Extract npm run <script> and npm <command>
Validate against package.json scripts

5. Efficiency Analysis (MEDIUM Certainty)

Token Count

Estimate: characters / 4 or words * 1.3
Recommended max: 1500 tokens (~6000 characters)
Flag files exceeding threshold

README Duplication

Detect overlap with README.md
Flag >40% content duplication
CLAUDE.md should complement README, not duplicate

Verbosity

Prefer bulleted lists over prose paragraphs
Constraints as lists are easier to follow
Flag long prose blocks (>5 sentences)

6. Quality Checks (MEDIUM Certainty)

WHY Explanations

Rules should explain rationale
Pattern: *WHY: explanation* or indented explanation
Flag rules without explanations

Structure Depth

Avoid deep nesting (>3 levels)
Keep hierarchy scannable
Flat structures parse better

XML-Style Tags (Optional Enhancement)

Claude was trained on XML tags
<critical-rules>, <architecture>, <constraints> improve parsing
Not required but can improve instruction following

7. Agent/Skill Definitions (MEDIUM Certainty)

If file defines custom agents or skills:

Agent Definition Format

### agent-name
Model: claude-sonnet-4-20250514
Description: What this agent does and when to use it
Tools: Read, Grep, Glob
Instructions: Specific behavioral instructions

Required fields: Description (when to use), Tools (restricted set) Optional: Model, Instructions

Skill References

Skills should have clear trigger descriptions
"Use when..." pattern helps auto-invocation

8. Cross-Platform Compatibility (MEDIUM/HIGH Certainty)

State Directory

Don't hardcode .claude/
Support .opencode/, .codex/
Use ${STATE_DIR}/ or document variations

Terminology

Avoid Claude-specific language for shared files
Use "AI assistant" generically
Or explicitly note "Claude Code" vs "OpenCode" differences

Output Format

# Project Memory Analysis: {filename}

**File**: {path}
**Type**: {CLAUDE.md | AGENTS.md}

## Metrics
| Metric | Value |
|--------|-------|
| Estimated Tokens | {tokens} |
| README Overlap | {percent}% |

## Summary
| Certainty | Count |
|-----------|-------|
| HIGH | {n} |
| MEDIUM | {n} |

### Structure Issues ({n})
| Issue | Fix | Certainty |

### Instruction Issues ({n})
| Issue | Fix | Certainty |

### Positioning Issues ({n})
| Issue | Fix | Certainty |

### Reference Issues ({n})
| Issue | Fix | Certainty |

### Efficiency Issues ({n})
| Issue | Fix | Certainty |

### Cross-Platform Issues ({n})
| Issue | Fix | Certainty |

Pattern Statistics

Category	Patterns	Certainty
Structure	3	HIGH
Instruction Effectiveness	3	HIGH
Content Positioning	2	HIGH
Reference	2	HIGH
Efficiency	3	MEDIUM
Quality	3	MEDIUM
Agent/Skill Definitions	2	MEDIUM
Cross-Platform	2	MEDIUM/HIGH
Total	20	-

### Example: Missing WHY Explanations

<bad_example>

## Rules
1. Always run tests before committing
2. Use semantic commit messages

Issue: Rules without rationale are harder to follow. </bad_example>

<good_example>

## Critical Rules
1. **Always run tests before committing**
   *WHY: Catches regressions before they reach main branch.*

Why it's good: Motivation makes compliance easier. </good_example>

Example: Negative vs Positive Instructions

<bad_example>

- Don't use console.log for debugging
- Never commit directly to main
- Avoid hardcoding secrets

Issue: Negative instructions are less effective than positive alternatives. </bad_example>

<good_example>

- Use the logger utility for all debug output
- Create feature branches and submit PRs for all changes
- Store secrets in environment variables or .env files

Why it's good: Tells what TO do, not just what to avoid. </good_example>

Example: Weak vs Strong Constraint Language

<bad_example>

- You should probably run tests before pushing
- Try to use TypeScript when possible
- Consider adding error handling

Issue: Weak language ("should", "try", "consider") reduces compliance. </bad_example>

<good_example>

- **MUST** run tests before pushing (CI will reject failures)
- **ALWAYS** use TypeScript for new files
- **REQUIRED**: All async functions must have error handling

Why it's good: Strong language ensures critical rules are followed. </good_example>

Example: Content Positioning

<bad_example>

## Project Overview
[Long description...]

## Installation
[Setup steps...]

## Critical Rules
1. Never push to main directly
2. Always run tests

Issue: Critical rules buried in middle/end get less attention. </bad_example>

<good_example>

## Critical Rules (Read First)
1. **Never push to main directly** - Use PRs
2. **Always run tests** - CI enforces this

## Project Overview
[Description...]

## Reminders
- Check CI status before merging
- Update CHANGELOG for user-facing changes

Why it's good: Critical content at START and END positions. </good_example>

Example: Cross-Platform Compatibility

<bad_example>

State files are stored in `.claude/tasks.json`

Issue: Hardcoded paths exclude other AI tools. </bad_example>

<good_example>

State files are stored in `${STATE_DIR}/tasks.json`
(`.claude/` for Claude Code, `.opencode/` for OpenCode)

Why it's good: Works across multiple AI assistants. </good_example>

Example: Agent Definition

<bad_example>

## Agents
- security-reviewer: reviews security
- test-writer: writes tests

Issue: Missing required fields (Tools, when to use). </bad_example>

<good_example>

## Custom Agents

### security-reviewer
Model: claude-sonnet-4-20250514
Description: Reviews code for security vulnerabilities. Use for PRs touching auth, API, or data handling.
Tools: Read, Grep, Glob
Instructions: Focus on OWASP Top 10, input validation, auth flows.

### test-writer
Model: claude-haiku-4
Description: Writes unit tests. Use after implementing new functions.
Tools: Read, Write, Bash(npm test:*)
Instructions: Use Jest patterns. Aim for >80% coverage.

Why it's good: Complete definition with when to use, restricted tools. </good_example>

Research References

Best practices derived from:

agent-docs/PROMPT-ENGINEERING-REFERENCE.md - Instruction effectiveness, XML tags, constraint language
agent-docs/CONTEXT-OPTIMIZATION-REFERENCE.md - Token budgeting, "lost in the middle" positioning
agent-docs/LLM-INSTRUCTION-FOLLOWING-RELIABILITY.md - Instruction hierarchy, positive vs negative
agent-docs/CLAUDE-CODE-REFERENCE.md - File hierarchy, agent definitions, skills format

Constraints

Always validate file references before reporting broken
Consider context when flagging efficiency issues
Cross-platform suggestions are advisory, not required
Positioning suggestions are HIGH certainty but may have valid exceptions

enhance-claude-memory

Más de este repositorio

Más de este repositorio