methodology-selector

Methodology Selector

Choose the thinking method for the current stage of work. Do not perform the work, create a skill, or write a loop goal.

Input contract

Read these inputs from the user's request and current artifacts:

• Task: the work that needs a method.
• Audience: who will use or review the result.
• Context: facts and constraints that change the method choice.
• Output format: the requested artifact or decision shape.
• Validation: checks and evidence lists describing observable acceptance.
• Human review / stop: stop_conditions, human_approval_required, and approval_actions.

Do not invent a missing input if it could change the selected method. Name the missing field and return none when selection would otherwise be guesswork.

Selection procedure

1. Read references/manifest.json and references/catalog.md.
2. Read the Fit sections of all twelve canonical references. Do not route from a hard-coded method summary in this skill.
3. Identify the task's dominant need for this stage, not every method that could eventually be useful.
4. Test each plausible candidate's positive fit against task evidence.
5. Test its negative fit. Negative fit vetoes keyword similarity.
6. When two methods could help, select the one that owns the current stage and list the other as deferred. Never blend two methods into an invented hybrid.
7. Select none when all candidates are vetoed, required evidence is missing, or the task is already narrow direct execution.
8. Read the selected reference in full and use its canonical quality questions.

Output contract

Return exactly this structure. contract is the canonical eight-field handoff to goal-engineering; copy the six input fields faithfully and add the selected methodology plus method-fit evidence:

contract:
  methodology: <exact slug from manifest.json or none>
  task: <input task, unchanged in meaning>
  audience: <input audience, unchanged in meaning>
  context: <input context, unchanged in meaning>
  output_format: <input output format, unchanged in meaning>
  validation:
    checks: <input checks, unchanged in meaning>
    evidence: <input evidence, unchanged in meaning>
  method_fit:
    best_when: <catalog fit or "Not applicable">
    avoid_when: <catalog negative fit or "Not applicable">
    reason: <task facts that establish the choice>
  human_review_stop:
    stop_conditions: <input stop conditions, unchanged in meaning>
    human_approval_required: <input boolean>
    approval_actions: <input actions, unchanged in meaning>
selection:
  methodology_name: <display name or "None">
  reason: <why this method owns the current stage>
  deferred_method: <methodology slug or none>
  method_ref: <relative reference path or none>
  quality_questions:
    - <question copied or faithfully adapted from the selected method reference>

When no secondary method owns a later stage, always return deferred_method: none. For none, keep quality_questions empty and explain which missing fact or negative-fit rule prevented selection. A none contract must stop before goal-engineering.

Boundaries

• Do not route by a methodology name appearing in the request.
• Do not choose a method because it is generally useful.
• Do not reproduce canonical method definitions or fit rules in this file.
• Do not convert method selection into task execution.
• Do not create or edit a methodology skill.
• Do not write a loop goal or readiness verdict.
• Do not select a slug absent from the manifest.

Fixture validation

Resolve the absolute directory containing this SKILL.md, then run:

node /absolute/path/to/methodology-selector/scripts/self-test.mjs

This verifies manifest integrity, the six incoming contract fields, exact slugs, must-not choices, deferred-method cases, and positive plus negative-fit coverage for all twelve methods. The same check validates the persisted per-case observations in evals/cold-routing-results.json against the fixtures after expectations were revealed. The persisted matrix records, but does not recreate, the independent blinded routing review.

Stop

Stop after one selection. If evidence is insufficient, return none with the smallest missing input; do not interview indefinitely or guess.