/learn-from-prs

argument-hint: "[--count <n>] [--state merged|closed|all]"
description: Analyze PR feedback patterns and suggest config updates to catch issues locally
allowed-tools: Bash, Glob, Read, Grep

Learn From PR Feedback

Analyze feedback from recent PRs (CodeRabbit, SonarQube, developer comments) and suggest updates to project configuration to catch these issues before PR review.

Arguments

ARGUMENTS: $ARGUMENTS

• `--count <n>`: Number of PRs to analyze (default: 5)
• `--state <state>`: PR state to fetch - merged, closed, or all (default: merged)

Procedure

Step 1: Verify Prerequisites

Check that `gh` CLI is authenticated:

gh auth status

If not authenticated, stop and inform user to run `gh auth login`.

Step 2: Fetch Recent PRs

gh pr list --state <state> --limit <count> --json number,title,url,mergedAt,closedAt

If no PRs found, report "No PRs found matching criteria" and stop.

Report:

Found X PRs to analyze:
1. #123: PR title
2. #124: PR title
...

Step 3: Collect All Feedback

For each PR, fetch ALL comments and reviews:

**PR review comments (inline code comments):**

gh api repos/{owner}/{repo}/pulls/{pr_number}/comments --paginate

**PR reviews (approval/request changes with body):**

gh api repos/{owner}/{repo}/pulls/{pr_number}/reviews --paginate

**Issue-style comments (general discussion):**

gh api repos/{owner}/{repo}/issues/{pr_number}/comments --paginate

Extract from each comment:

• `body`: The actual feedback text
• `user.login`: Who wrote it (helps identify bots like CodeRabbit, SonarQube)
• `path` and `line`: Where in code (if applicable)
• `created_at`: When

**Categorize by source:**

• **Automated tools**: Identify by user.login or body markers (CodeRabbit, SonarQube, Codacy, DeepSource, etc.)
• **Human reviewers**: Everything else

Report:

Collected feedback from PR #X:
- [Tool name]: Y comments
- Human reviewers: Z comments

Step 4: Analyze Patterns

Group all collected feedback and identify patterns:

**Look for recurring themes such as:**

• Code style and formatting issues
• Logic errors and edge cases
• Security concerns
• Performance issues
• Testing gaps
• Architecture and design feedback
• Documentation issues
• Language-specific best practices

For each pattern found:

• Count occurrences across all PRs
• Note specific examples (file, line, comment text)
• Identify the source (which tool/reviewer caught it)

**Prioritize by:** 1. Frequency (appears in multiple PRs) 2. Source consistency (caught by multiple reviewers/tools) 3. Severity (security > bugs > style)

Report top patterns:

## Feedback Patterns Found

### High Frequency (appeared in 3+ PRs)
1. [Pattern]: [X occurrences] - caught by [sources]
   Example: "[actual comment snippet]"

### Medium Frequency (appeared in 2 PRs)
...

### Single Occurrence (but notable)
...

Step 5: Discover Current Config

Search the project for ANY configuration files that might be relevant:

find . -maxdepth 3 -type f \( \
  -name "CLAUDE.md" -o \
  -name ".claude" -o \
  -name ".*rc" -o \
  -name ".*rc.json" -o \
  -name ".*rc.yaml" -o \
  -name ".*rc.yml" -o \
  -name "*.config.js" -o \
  -name "*.config.ts" -o \
  -name "*.config.json" -o \
  -name ".pre-commit-config.yaml" -o \
  -name ".editorconfig" -o \
  -name "Makefile" -o \
  -name "pyproject.toml" -o \
  -name "setup.cfg" -o \
  -name "tox.ini" -o \
  -name ".rubocop.yml" -o \
  -name "Gemfile" -o \
  -name "go.mod" -o \
  -name "Cargo.toml" -o \
  -name "composer.json" \
\) 2>/dev/null | head -50

Also check for `.claude/` directory contents.

For each relevant config, note:

• What it currently enforces
• What gaps exist relative to PR feedback patterns

Report:

## Current Configuration

Found these config files:
- [file]: [brief summary of what it controls]
...

Step 6: Generate Recommendations

For each high-frequency pattern, propose specific config updates:

**Format for each recommendation:**

### [Pattern Name]

**Problem:** [What the PR feedback consistently caught]

**Frequency:** [X occurrences across Y PRs]

**Caught by:** [Tool names, specific reviewers]

**Examples from PRs:**
- PR #123: "[comment snippet]"
- PR #456: "[comment snippet]"

**Suggested fixes:**

**Option A: CLAUDE.md** (catches during Claude Code sessions)
```markdown
[Exact text to add to CLAUDE.md]

**Option B: [Relevant tool/config for this project]**

[Exact config to add, appropriate to the project's language/tooling]

**Option C: Pre-commit hook** (catches before commit)

[Hook configuration if applicable]

**Recommended:** [Which option and why]

### Step 7: Present Summary

```markdown
# PR Feedback Analysis Complete

## PRs Analyzed
- Count: X
- Date range: [oldest] to [newest]
- Feedback sources: [list tools and reviewer count]

## Key Findings

### Patterns That Should Be Caught Locally

| Pattern | Frequency | Top Source | Recommended Fix |
|---------|-----------|------------|-----------------|
| [name]  | X times   | [source]   | [config type] |

## Detailed Recommendations

[Full recommendations from Step 6]

## Quick Wins (Copy-Paste Ready)

### Add to CLAUDE.md:
```markdown
[All CLAUDE.md additions consolidated]

Add to [other relevant config]:

[Additions consolidated by config file]

--- *Analysis based on PRs: #X, #Y, #Z...* *Run `/learn-from-prs` again after implementing changes to verify improvement*

## Notes

- This command only suggests changes, never modifies files
- Focus recommendations on issues that appeared multiple times
- For single-occurrence issues, mention them but deprioritize
- Recommend fixes appropriate to the project's actual language and tooling
- When a pattern could be caught by multiple tools, recommend the earliest in the pipeline
- If feedback is vague or unclear, note it but don't force a recommendation