Route any task, question, or situation to the right agent skill in the library.
You do not need to memorize every skill. When a problem arrives—or when you or the user are unsure how to proceed—use this routing matrix to find the high-leverage skill that fits.
The Routing Matrix
1. Conception, Ideas & Planning
| Situation | Recommended Skill | Why It Fits |
|---|---|---|
| Idea is fuzzy; requirements not settled | grill-me or grill-with-docs | Relentless 1-question-at-a-time interview to pin down real user intent. |
| Requirements feel bloated, slow, or arbitrary | the-algorithm | Elon Musk’s 5-step process: question requirement, delete parts, optimize, accelerate, automate. |
| Challenging the core premise before building | attack-the-premise | Tests whether the problem is worth solving before writing code. |
| Designing a complex system or service | architect | Creates structural blueprints and checks against architectural red flags. |
| High-risk change that could break dependencies | blast-radius | Maps upstream/downstream callers and calculates blast radius before editing. |
| Greenfield project / massive multi-session build | to-spec → to-tickets | Converts ideas into specs, then splits into self-contained blocker-linked tickets. |
2. Forensics & Understanding Code
| Situation | Recommended Skill | Why It Fits |
|---|---|---|
| “How does this code / repo work?” | how | Traces code execution, component ownership, and boundaries without altering state. |
| “Why was this written this way?” | why | Deep archaeology through commits, PRs, and history to uncover original design rationale. |
| Need a plain-human, zero-jargon explanation | bro | Translates complex technical concepts into direct, friendly, jargon-free English. |
| Verifying a claim against empirical evidence | verify-this | Captures baseline vs treatment artifacts to prove or disprove a technical claim. |
3. Debugging & Remediation
| Situation | Recommended Skill | Why It Fits |
|---|---|---|
| Intermittent, hard, or flaky bug | diagnosing-bugs | Establishes a tight red-green reproduction loop before hypothesizing fixes. |
| Bug keeps returning / symptom patching | fix-root-causes | Refuses superficial patches; fixes underlying structural causes. |
| GitHub Actions CI failed | fix-ci + loop-on-ci | Inspects failure logs, isolates root cause, and loops until green. |
| In-progress merge or rebase conflict | fix-merge-conflicts | Resolves conflicts hunk-by-hunk based on intent traced to primary sources. |
4. Implementation & Craft
| Situation | Recommended Skill | Why It Fits |
|---|---|---|
| Building a feature or behavior | tdd | Red-green-refactor loop. Writes the failing test first, then the minimal code. |
| Testing behavior rather than implementation | test-behavior-not-implementation | Calls code like users do; stops tests breaking on benign internal refactors. |
| Designing functions, APIs, or data models | type-system-discipline | Uses types to make illegal states unrepresentable. |
| Writing stateful business logic | model-the-domain | Maps domain concepts cleanly into the type system before writing procedures. |
| Building repetitive tasks into tools | build-the-lever | Builds a script or lever when doing a task for the second or third time. |
| Concurrent state or distributed writes | separate-before-serializing | Eliminates shared write locks before serializing state. |
5. Refactoring & Codebase Health
| Situation | Recommended Skill | Why It Fits |
|---|---|---|
| Too much code / dead weight / bloat | subtract-before-you-add | Deletes dead code and simplifies interfaces before adding new features. |
| Code full of AI boilerplate and slop | deslop | Strips generic AI coding patterns, redundant comments, and defensive clutter. |
| Code littered with obvious/noisy comments | no-comments | Removes comments that repeat what the code says; encodes constraints in types. |
| Sunsetting old APIs or migrating callers | migrate-callers-then-delete-legacy | Migrates all call sites first, then completely deletes the legacy code. |
6. Review, Verification & Shipping
| Situation | Recommended Skill | Why It Fits |
|---|---|---|
| Aggressive security & correctness audit | thermo-nuclear-review | Deep multi-axis security audit with harsh rubrics for production code. |
| Architectural debt & giant file review | thermo-nuclear-code-quality-review | Flags sprawling files, leaky abstractions, and tangled modules. |
| Preparing a PR to be merged safely | review-and-ship | Checks intent fit, runs tests, creates conventional commit, and opens PR. |
| Proving changes work before declaring done | prove-it-works | Verifies against real runtime artifacts, logs, or outputs before claiming success. |
| Tidying a PR so human reviewers love it | make-pr-easy-to-review | Cleans commit noise, improves descriptions, and groups diffs logically. |
7. Documentation & Writing
| Situation | Recommended Skill | Why It Fits |
|---|---|---|
| Prose, docs, or PR descriptions sound like AI | antislop or unslop | Strips negative parallelisms, inflated significance, em-dash addiction, and AI buzzwords. |
| Writing structured technical documentation | technical-writing | Diátaxis framework (tutorials, how-to guides, reference, explanation) with clear prose. |
8. Scaling, Agents & Meta
| Situation | Recommended Skill | Why It Fits |
|---|---|---|
| Large task that can be split into parallel workers | swarm | Fans out N independent subagent workers and synthesizes results. |
| Long-running work needing an audit log | show-me-your-work | Maintains a timestamped decision trail in TSV for unattended execution. |
| Opinionated, disciplined agent execution | poteto-mode or dyl-mode | High-signal, unslopped responses with deliberate subagent routing. |
| Need a second opinion from a stronger model | advisor | Consults a stronger model before critical architectural decisions. |
| Designing a CLI that AI agents can run | cli-for-agents | Non-interactive flags, structured errors, idempotency, and clean --help. |
Operating Instructions for the Agent
When which-skill is triggered (either explicitly by the user or automatically when intent is ambiguous):
- Identify the Core Need: Distill the user’s situation into one primary category (Planning, Understanding, Fixing, Building, Refactoring, Reviewing, Writing, Scaling).
- Select the Top 1–2 Skills: Pick the exact match from the matrix. Do not dump all 72 skills on the user.
- Present the Recommendation in 3 Parts:
- Recommended Skill: Name and category.
- Why It Fits: 1–2 sentences connecting the skill’s strength directly to the user’s problem.
- Immediate Next Step: The exact prompt or action to execute right now.
- Offer Execution: Ask the user if they’d like you to invoke that skill immediately.