kata

Router skill: one entry, 30 agentic-engineering playbooks. Resolve topic, Read EXACTLY ONE playbook from routes:, apply it. [DOC]

When to use

Trigger when the request maps to a known agentic pattern — building a Claude agent loop, MCP server, hook, structured extraction, code-review harness, memory hierarchy, multi-agent topology, or prompt/sampling/context tactic — and you want the proven JM Labs recipe instead of improvising. [INFERENCIA] Do NOT use as a generic chat or to answer questions that no playbook covers; if no topic fits, say so and route the user elsewhere rather than guessing. [DOC]

Inputs

• topic (required): one of the 30 routes: keys. Infer from the request; ask only when two topics are genuinely plausible. [DOC]
• depth (default quick): quick → essentials + the validation gate only; deep → apply exhaustively, verifying at each step. [DOC]

Procedure

1. Map request → topic. If ambiguous, present the 2 closest keys and ask. [DOC]
2. Read EXACTLY ONE playbook (its routes: path). Never load the whole cluster — that dilutes context and defeats hub-and-spoke isolation. [INFERENCIA]
3. Execute along the spine: Discover → Analyze → Execute → Validate. [DOC]
4. Apply the playbook's own acceptance criteria before declaring done. [DOC]

Validation gate (acceptance)

• Exactly one playbook was read; topic matches the user's actual intent. [DOC]
• Output follows the chosen playbook's structure, not improvised prose. [DOC]
• Every non-obvious claim carries an evidence tag from the kit (Alfa/bracket) family per ../../references/verification-tags.md — never mix tag families. [DOC]
• Constitution v6.0.0 gates honored: enforcement, evidence tags, script-first (prefer a script over manual steps when one exists). [DOC]
• Score the result with assets/quality-rubric.json and run assets/routing-checklist.md before declaring done (see assets/README.md). [DOC]

Anti-patterns

• Loading several playbooks "to compare" — pick one; re-route if wrong. [DOC]
• Guessing a topic silently when the request is ambiguous. [DOC]
• Answering from memory of a pattern instead of reading its current playbook — recipes drift; the file is the source of truth. [INFERENCIA]
• Emitting quick depth but skipping the validation gate. [DOC]

Self-correction

If mid-task the evidence contradicts the chosen topic (wrong failure mode, the playbook's preconditions don't hold), STOP, name the mismatch, and re-resolve topic — do not force-fit the original playbook. [INFERENCIA]