Learn From PRs

Extract patterns from human PR review comments. Shows proposals with citations for you to decide which become AGENTS.md rules.

When to Use

• Setting up AGENTS.md for a new project
• Onboarding new team members
• Documenting implicit team conventions
• "What patterns does this team follow?"
• "What do reviewers keep asking for?"

Critical Rules

1. AGENTS.MD IS FOR AI AGENTS - Rules must be things an AI can follow when writing code
2. SKIP HUMAN PROCESS PATTERNS - These are NOT for AGENTS.md:
   • "keep PRs focused" ❌ (PR workflow, not code)
   • "validate with product specs" ❌ (AI can't consult humans)
   • "add PR description" ❌ (PR workflow)
   • "add tests" ❌ (too generic)
   • "remove comments" ❌ (too generic)
3. FOCUS ON CODE PATTERNS - Things AI can actually do:
   • Specific APIs: "use X instead of Y" ✅
   • Specific architecture: "put logic in A, not B" ✅
   • Specific utils: "use utilX from path/to/utils" ✅
   • Specific style: "use arrow functions for components" ✅
   • Specific organization: "shared types in shared folders" ✅
4. 3+ OCCURRENCES - Threshold for team-specific patterns
5. AUTO-ESCALATE - If < 3 patterns, fetch more PRs. Don't ask user.
6. OUTPUT DIRECTLY - Don't ask "which patterns to add?" - just output the AGENTS.md rules.

Parameters

Workflow

Run Analysis

node scripts/analyze-prs.js <owner/repo> [--limit <n>] [--skip <n>] [--author <username>]

Auto-Escalation with Pagination

DO NOT re-fetch PRs you already have. Use --skip for pagination:

1. First run: --limit 50 → analyze, keep comments in memory
2. If < 3 patterns → --limit 50 --skip 50 → append new comments
3. If still < 3 → --limit 100 --skip 100 → append
4. If still < 3 → --limit 200 --skip 200 → append
5. If still < 3 → --limit 500 --skip 500 → append
6. After 1000 total PRs, report findings

Accumulate comments across runs. Don't discard previous results.

Your Job: Semantic Pattern Finding

The script outputs a flat list of all human review comments. YOU (Claude) must find semantic patterns.

What the script does:

• Fetches PR review comments
• Filters out bots
• Outputs flat JSON list with: body, reviewer, url, pr

What YOU must do:

1. Read ALL comments
2. Find semantic patterns - comments that mean the same thing but use different words
3. SKIP GENERIC PATTERNS - these are useless nitpicks everyone knows:
   • "add tests" ❌
   • "remove comments" ❌
   • "add PR description" ❌
   • "keep PRs focused" ❌
   • "remove redundant code" ❌
   • "justify your changes" ❌
4. FOCUS ON TEAM-SPECIFIC patterns - things unique to THIS codebase:
   • Specific APIs: "use X instead of Y" ✅
   • Specific utils: "use utilX from path/to/utils" ✅
   • Specific architecture: "put logic in A, not B" ✅
   • Specific imports: "import X from specific/path" ✅
   • Specific style: "arrow functions for components" ✅
5. For team-specific patterns, threshold is 3+ occurrences (not 5)
6. Group semantically similar comments and count occurrences

Output Rules

Output AGENTS.md rules directly. Don't ask user to choose.

Format:

## AGENTS.md Rules (from PR review patterns)

### [Category]
- Rule 1 (cite: [PR](url))
- Rule 2 (cite: [PR](url))

Example output:

## AGENTS.md Rules (from PR review patterns)

### API Usage
- Use `newApi()` instead of `legacyApi()` - legacy is deprecated (cite: [PR](url))

### Architecture
- Put feature flags in controllers, NOT services - service flags cause cache issues (cite: [PR](url))
- Use `checkFeature` from `src/utils/features.ts` (cite: [PR](url))

### Components
- Use arrow function syntax: `const Component = (props: Props) => { ... }` (cite: [PR](url))
- Shared types go in `types/` folder, not duplicated per component (cite: [PR](url))

If no code patterns found: "No actionable code patterns found. Team feedback is situational."

Error Handling

Requirements

• gh CLI authenticated
• Node.js 18+