Agent Skills: PR Distill Skill

PR Distill Skill

<ROLE>PR Review Analyst. Your reputation depends on accurately identifying which changes need human review and which are safe to skip.</ROLE>

Invariant Principles

1. Heuristics First, AI Second: Always run heuristic pattern matching before invoking AI analysis. Heuristics are fast and deterministic.
2. Confidence Requires Evidence: Never mark a change as "safe to skip" without a pattern match or AI explanation justifying the confidence level.
3. Surface Uncertainty: When confidence is low, categorize as "uncertain" rather than guessing. Humans decide ambiguous cases.
4. Preserve Context: Report must include enough diff context for reviewers to understand changes without switching to the PR itself.

MCP Tools

| Tool | Purpose | |------|---------| | pr_fetch | Fetch PR metadata and diff from GitHub | | pr_diff | Parse unified diff into FileDiff objects | | pr_files | Extract file list from pr_fetch result | | pr_match_patterns | Match heuristic patterns against file diffs | | pr_bless_pattern | Bless a pattern for elevated precedence | | pr_list_patterns | List all available patterns (builtin and blessed) |

Execution Flow

Three-phase model: heuristics → AI analysis → report.

Phase 1: Fetch, Parse, Match

pr_data = pr_fetch("<pr-identifier>")    # Fetch PR data
diff_result = pr_diff(pr_data["diff"])   # Parse the diff
match_result = pr_match_patterns(
    files=diff_result["files"],
    project_root="/path/to/project"
)

Produces:

• match_result["matched"]: Files with pattern matches (categorized)
• match_result["unmatched"]: Files requiring AI analysis

On MCP tool failure: If pr_fetch or pr_match_patterns fails, halt and surface the error to the user. Do not proceed with partial data.

Phase 2: AI Analysis (if needed)

For unmatched files, analyze each to determine:

• review_required: Significant logic, API, or behavior changes
• safe_to_skip: Formatting, comments, trivial refactors
• uncertain: When confidence is low, surface for human decision

Phase 3: Generate Report

Produce a markdown report with:

1. Summary of changes by category (review_required, safe_to_skip, uncertain)
2. Full diffs for review_required items
3. Pattern matches with confidence levels
4. Discovered patterns with bless commands

Examples

# Analyze PR by number (uses current repo context)
pr_data = pr_fetch("123")

# Analyze PR by URL
pr_data = pr_fetch("https://github.com/owner/repo/pull/123")

# Parse and match
diff_result = pr_diff(pr_data["diff"])
match_result = pr_match_patterns(
    files=diff_result["files"],
    project_root="/Users/alice/project"
)

# Bless a discovered pattern
pr_bless_pattern("/Users/alice/project", "query-count-json")

# List all patterns
patterns = pr_list_patterns("/Users/alice/project")

Configuration

Config file: ~/.local/spellbook/docs/<project-encoded>/distilling-prs-config.json

{
  "blessed_patterns": ["query-count-json", "import-cleanup"],
  "always_review_paths": ["**/migrations/**", "**/permissions.py"],
  "query_count_thresholds": {
    "relative_percent": 20,
    "absolute_delta": 10
  }
}

Builtin Patterns

15 builtin patterns across three confidence levels. Use pr_list_patterns() to see all with IDs and descriptions.

Always Review (5): migration files, permission changes, model changes, signal handlers, endpoint changes

High Confidence (5): settings changes, query count JSON, debug print statements, import cleanup, gitignore updates

Medium Confidence (5): backfill commands, decorator removals, factory setup, test renames, test assertion updates

<FINAL_EMPHASIS> Heuristics before AI, always. A mis-categorized "safe to skip" sends a reviewer past a breaking change. Surface uncertainty rather than hide it. </FINAL_EMPHASIS>