Prompt Examples

All blocks below implement the same Five Strategies with different thresholds and emphasis. Each is self-contained — copy the one that matches your context, then append the Generation Mode Auto-Switch block.

For where to place these blocks (CLAUDE.md, .cursorrules, system prompt), see Activating in Your Agent.

Shared vocabulary across all strategies:

Step label	Maps to strategy	Tool
TEMPLATE	1. Template Selection	search_code with rerank: "stable"
RISK CHECK	5. Risk Assessment	semantic_search with rerank: "hotspots"
STYLE MATCH	3. Style Consistency	semantic_search with rerank: "ownership"
ANTI-PATTERN	2. Anti-Pattern Detection	search_code with rerank: "hotspots"
CONTEXT	4. Historical Context	check taskIds in search results
VERIFY	Post-generation validation	ripgrep for every function, import, type
BLAST RADIUS	Post-generation impact	semantic_search with rerank: "impactAnalysis"

Signal names used in all blocks:

Level	Signals
Chunk (function/block)	chunkBugFixRate, chunkCommitCount, chunkAgeDays, chunkChurnRatio
File	bugFixRate, commitCount, ageDays, dominantAuthorPct, relativeChurn, churnVolatility, imports

Chunk-level signals are more granular — use them in threshold tables. File-level signals are used for ownership, blast radius, and filters.

Balanced Strategy (Recommended)

All five strategies with sensible defaults. Recommended starting point for most projects.

## Code Generation Strategy

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

### Before generating code:

1. TEMPLATE: Find stable template — `search_code` with `rerank: "stable"`
   - Only use results with chunkBugFixRate < 25%, chunkCommitCount <= 5
2. RISK CHECK: Assess target area — `semantic_search` with `rerank: "hotspots"`, `metaOnly: true`
   - See threshold table → select generation mode
3. STYLE MATCH: Match domain owner — `semantic_search` with `rerank: "ownership"`, `metaOnly: true`

### After generating code:

4. VERIFY: Confirm identifiers with ripgrep — every function, import, type must exist
   - Semantic search finds by meaning, not exact names
   - 0 ripgrep matches = fix the reference before committing
5. BLAST RADIUS: Check impact — `semantic_search` with `rerank: "impactAnalysis"`, `metaOnly: true`

### Threshold Decision Table

| Signal              | Safe   | Caution  | Stop                            |
|---------------------|--------|----------|---------------------------------|
| chunkBugFixRate     | < 25%  | 25–40%   | > 40% (wrapper + feature flag)  |
| chunkCommitCount    | 1–5    | 6–8      | > 8 (simplify, reduce)          |
| chunkAgeDays        | < 90   | 90–180   | > 180 (minimal changes only)    |
| dominantAuthorPct   | < 70%  | 70–85%   | > 85% (flag owner for review)   |
| relativeChurn       | < 2.0  | 2.0–5.0  | > 5.0 (propose rewrite)         |

### 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

Generation Mode Auto-Switch

Append after any strategy block. The agent auto-selects a generation mode based on the RISK CHECK step.

## Generation Mode

After the RISK CHECK, auto-select the generation mode:

### Mode Decision Table

| Condition                                 | Mode          | Behavior                                    |
|-------------------------------------------|---------------|---------------------------------------------|
| chunkBugFixRate > 40% AND chunkAgeDays > 60       | Defensive     | Wrapper pattern, feature flag, keep old path |
| chunkCommitCount > 8 AND churnVolatility > 15     | Stabilization | Simplify branching, extract methods, log     |
| chunkAgeDays > 180 AND chunkCommitCount <= 2       | Conservative  | Minimal changes, preserve signatures         |
| otherwise                                 | Standard      | Clean idiomatic code, follow templates       |

Evaluate top-to-bottom. First match wins.
Multiple red flags → escalate to user before proceeding.

### Mode-Specific Rules

| Mode          | Do                                           | Don't                                    |
|---------------|----------------------------------------------|------------------------------------------|
| Defensive     | Wrap, flag, keep fallback, plan cleanup date | Delete old code, refactor inline          |
| Stabilization | Extract methods, reduce paths, add logging   | Add features on top, increase branching   |
| Conservative  | Add alongside, duplicate over abstract       | Move code, rename, change signatures      |
| Standard      | Follow template, match owner style           | Ignore metrics — still verify with ripgrep |

Safety-First Strategy

Same five strategies, stricter thresholds. RISK CHECK runs first (mandatory danger zone gate before any generation). For legacy systems or regulated environments.

## Code Generation Strategy: Safety-First

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

### Before generating code:

1. RISK CHECK (mandatory first): Assess danger zone — `semantic_search` with `rerank: "hotspots"`, `metaOnly: true`
   - If any signal hits Stop → STOP and warn user before proceeding
   - If modifying auth/payment/crypto: also run `rerank: "securityAudit"`
2. TEMPLATE: Find stable template — `search_code` with `rerank: "stable"`
   - Only use results passing the quality gate
3. STYLE MATCH: Match domain owner — `semantic_search` with `rerank: "ownership"`, `metaOnly: true`
   - Default to wrapper pattern for any code with chunkAgeDays > 90

### After generating code:

4. VERIFY: Confirm identifiers with ripgrep — every function, import, type must exist
   - 0 ripgrep matches = fix the reference
5. BLAST RADIUS: Check impact — `semantic_search` with `rerank: "impactAnalysis"`, `metaOnly: true`

### Threshold Decision Table

| Signal              | Safe   | Caution  | Stop                     |
|---------------------|--------|----------|--------------------------|
| chunkBugFixRate     | < 15%  | 15–30%   | > 30% (stop, warn user)  |
| chunkCommitCount    | 1–3    | 4–5      | > 5                      |
| dominantAuthorPct   | < 80%  | 80–90%   | > 90% (knowledge silo)   |
| imports             | < 5    | 5–10     | > 10 (blast radius risk) |

### Template Quality Gate

| Signal           | Required             |
|------------------|----------------------|
| chunkBugFixRate  | < 15%                |
| chunkCommitCount | <= 3                 |
| chunkAgeDays     | > 60 (battle-tested) |

### Never:

- Copy code from results with chunkBugFixRate > 50%
- Modify code with chunkBugFixRate > 50% directly — propose rewrite plan
- Trust generated identifiers without ripgrep verification

Incident Response Strategy

Speed over thoroughness. TEMPLATE step is skipped — recent changes are checked first as the most likely cause. For emergency patches during production incidents.

## Incident Response Strategy

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

When fixing a production issue:
1. FIND RECENT: Search recent changes — `search_code` with `rerank: "recent"`, `maxAgeDays: 7`
   - Recent changes are the most likely cause
2. RISK CHECK (fallback): If no recent results — `search_code` with `rerank: "hotspots"`
   - Historically fragile code is the next suspect
3. STYLE MATCH: Identify owner — `semantic_search` with `rerank: "ownership"`, `metaOnly: true`
4. Generate: Fix minimally — smallest possible change, no refactoring
5. VERIFY: Confirm identifiers with ripgrep — every function, import, type must exist
6. Add regression test for the specific failure

### Escalation Decision Table

| Recent changes found? | Hotspot? | Action                          |
|-----------------------|----------|---------------------------------|
| Yes                   | —        | Revert or patch the recent code |
| No                    | Yes      | Focus on historically fragile   |
| No                    | No       | Widen search, check logs        |

New Feature Strategy

No legacy constraints — TEMPLATE step emphasized, ANTI-PATTERN step added to learn from mistakes in adjacent code. For greenfield development.

## New Feature Strategy

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

When creating new features:
1. TEMPLATE: Find proven patterns — `search_code` with `rerank: "stable"` in similar features
   - See quality gate below
2. STYLE MATCH: Check directory owner — `semantic_search` with `rerank: "ownership"`, `metaOnly: true`
   - Match their error handling, naming, and structure
3. ANTI-PATTERN: Identify what to avoid — `search_code` with `rerank: "hotspots"` in the same area
   - Note: complex branching, nested flags, mixed responsibilities
4. CONTEXT: Check related taskIds for feature history
5. Generate: Clean code following the best template found
6. VERIFY: Confirm identifiers with ripgrep — every function, import, type must exist

### Template Quality Gate

| Signal           | Ideal  | Acceptable | Reject              |
|------------------|--------|------------|---------------------|
| chunkBugFixRate  | 0%     | < 20%      | > 20%               |
| chunkAgeDays     | > 60   | > 30       | < 30 (too fresh)    |
| chunkCommitCount | 1–2    | 3–5        | > 5 (too volatile)  |

tip

Start with Balanced + Mode Auto-Switch and customize from there. For Cursor and custom agent formats, see Activating in Your Agent.