Skip to main content

Activating in Your Agent

Claude Code (CLAUDE.md)​

Add the strategy block to your project's CLAUDE.md file. Claude Code reads this file automatically and applies the instructions to every interaction.

# CLAUDE.md

## Search & Generation Strategy

tea-rags is available as an MCP server. Use it for ALL code search.

### Before generating code:
1. Find a stable template: `search_code` with `rerank: "stable"`
   - Only use results with chunkBugFixRate < 25%, chunkCommitCount <= 5
2. Check target area: `semantic_search` with `rerank: "hotspots"`, `metaOnly: true`
   - chunkBugFixRate > 40% + chunkAgeDays > 60 → wrapper pattern + feature flag
   - chunkCommitCount > 8 → simplify, reduce complexity
   - chunkAgeDays > 180 → minimal changes only
3. Match domain owner: `semantic_search` with `rerank: "ownership"`, `metaOnly: true`

### After generating code:
4. Verify with ripgrep: confirm ALL referenced functions, imports, and types exist
   - Semantic search finds by meaning, not exact identifiers
   - 0 ripgrep matches for a referenced function = fix the reference before committing
5. Check blast radius: `semantic_search` with `rerank: { custom: { imports: 0.7, similarity: 0.3 } }`, `metaOnly: true`

### Never:
- Copy code from results with chunkBugFixRate > 50%
- Modify single-owner code without flagging the owner
- Change code with relativeChurn > 5.0 without proposing a rewrite
- Trust generated identifiers without ripgrep verification

Cursor (.cursorrules)​

# .cursorrules

## Code Search

Use tea-rags MCP server for code search. Before writing new code:

Search for stable templates:
- Use rerank="stable" to find low-churn, battle-tested implementations
- Reject results with chunkBugFixRate > 40% — find alternatives
- Prefer results with chunkCommitCount <= 3 and chunkAgeDays > 60

Assess risk before modifying existing code:
- Use rerank="hotspots" with metaOnly=true on the target area
- If chunkBugFixRate > 40% and code is older than 60 days, use wrapper pattern
- If `blameDominantAuthorPct` label is `silo` or `deep-silo`, match the
  live-line owner's style exactly (their lines are still in the file)

Match coding style to domain owner:
- Use rerank="ownership" on the target directory (live-line, blame-based)
- Follow the live-line owner's (`blameDominantAuthor`) error handling,
  naming, and structure patterns
- For review routing, prefer `recentDominantAuthor` (recent commit window) —
  the live-line owner may have moved on

After generating code, verify with ripgrep:
- Confirm all referenced function names, imports, and types actually exist
- Semantic search finds by meaning, not exact identifiers — always verify
- 0 ripgrep matches = fix the reference before committing

Custom Agents (system prompt)​

For agents built with the Claude API or other LLM frameworks:

You have access to a tea-rags MCP server for semantic code search with
git-derived quality signals.

TEMPLATE SELECTION:
Use search_code with rerank="stable" to find templates.
A good template has: chunkBugFixRate < 25%, chunkCommitCount <= 5, chunkAgeDays > 60.
Never base new code on results with chunkBugFixRate > 50%.

RISK ASSESSMENT:
Before modifying code, use semantic_search with rerank="hotspots"
and metaOnly=true on the target area.
Switch to defensive mode (wrapper pattern) if:
  chunkBugFixRate > 40% AND chunkAgeDays > 60.

STYLE MATCHING:
Use semantic_search with rerank="ownership" to find the live-line owner
(`blameDominantAuthor` from `git blame HEAD`).
If `blameDominantAuthorPct` label is `concentrated`, `silo`, or `deep-silo`,
match their patterns exactly — their lines are still in the file.
If ownership label is `shared`, use project-wide conventions.
For review routing, use `recentDominantAuthor` (recent commit window) — they
are the ones mentally loaded into the area right now.

POST-GENERATION VERIFICATION (MANDATORY):
After generating code, use ripgrep to verify ALL referenced identifiers:
  - Function names, import paths, type names must return > 0 matches
  - Semantic search finds by meaning, not exact tokens
  - 0 matches = hallucinated identifier, fix before committing

The Transformation​

AspectBefore (similarity-only RAG)After (data-driven generation)
Template selectionFirst search hitLowest bug rate, proven in production
Style guideGeneric language conventionsDomain owner's actual patterns
Anti-patternsUnknown, invisibleExplicitly identified and avoided
Feature contextMissing — code exists in a vacuumtaskIds reveal evolution and intent
Risk assessmentNo signal — all code treated equallyMetrics-driven defensive strategies
Generation modeOne mode for everythingDynamically adapted to target area context
VerificationTrust semantic results as-isRipgrep verification of all referenced identifiers
Decision basis"Looks right"Measurable evidence from version history

This is the shift from "find similar code" to context-aware engineering — where the agent reasons about the quality, history, and ownership of every code decision it makes.

See Also​