Methodology picker

Methodology Picker

Help the user pick the right Spec-Driven Development methodology for their project context. This is conversational guidance, not scaffolding. The user is making a decision; you are helping them frame it.

When this skill auto-triggers

The user mentions:

• Starting / setting up / scaffolding a new project
• Choosing / picking / selecting an SDD framework or methodology
• Configuring Claude Code for a specific project
• "What methodology should I use for …"
• Comparison of Superpowers vs Spec-Kit vs MUSUBI vs OpenSpec
• Piskala / Wasowski by name

What to NOT do

• Do not write files. If the user wants files written, route them to /mck-scaffold:init.
• Do not install plugins. Same rule.
• Do not present all options as equivalent — the whole point is to choose.

Conceptual frame

Piskala's three-level rigor taxonomy

Level	Promise	Cost	When to pick it
L1 Spec-First	Flexibility	Low	~80% of projects. Spec lives in the feature branch and becomes historical after merge.
L2 Spec-Anchored	Auditability + traceability	Medium-high	Regulated work. Spec lives the whole lifecycle; every change requires an addendum.
L3 Spec-as-Source	Guarantee by construction	Aspirational in 2026	Watch / not-yet. Tessl (closed beta) and CSDD (academic).

Higher rigor is not better. It is more expensive, and only pays off when the project genuinely needs the promise.

Wasowski's four decision contexts

Context	Profile	Default path
A	Solo / small team, greenfield	Spec-Kit lean ± Superpowers HARD-GATE
B	10+ team, regulated, compliance	MUSUBI + CSDD security overlay
C	Brownfield / legacy	OpenSpec + Shotgun CLI
D	Security-critical (fintech, healthcare, embedded)	CSDD + MUSUBI + OWASP Skill

How to talk the user through it

Ask questions one at a time. Don't batch. Reference the question tree in interview/questions.md but adapt the wording conversationally.

The single most-discriminating question is: "Will an auditor ever ask about REQ-to-Test traceability for this code?" Answer determines L1 vs L2 cleanly.

Second-most-discriminating: "Greenfield or brownfield?" Brownfield routes to OpenSpec almost regardless of other answers — once you've identified brownfield, do not ask Q3 or Q4 (they're pruned by interview/questions.md; the routing table's brownfield-any rule short-circuits everything else).

Third (greenfield only): "Solo / small team / 10+?" Solo + greenfield + no audit → L1-lean or L1-superpowers. Team 10+ + audit → L2.

Honest caveats to surface

When recommending Superpowers: note the HN consensus that "I stripped it back to ~30%" — it's opinionated; that's a feature for serious work, friction for a 20-minute fix.

When recommending Spec-Kit: note the measured ~10x agent-time overhead on the first cycle (33.5 min vs 8 min baseline). Worth it on real product work, painful on throwaways.

When recommending MUSUBI: note that 28 GitHub stars vs Superpowers' 166k is not a sign of MUSUBI being bad — it's a sign that most teams don't need its rigor. If an auditor will ask, the 28 stars are irrelevant.

When the user's context suggests L3 might matter in 18+ months (security-critical financial/healthcare): mention Tessl and CSDD as "watch, not deploy" — don't oversell them. Tessl is closed beta; CSDD is a paper with one empirical study.

Handoff to the scaffolder

When the user agrees to install (any of "okay, do it" / "set this up" / "scaffold it" / "let's go"):

1. Capture the answer letters from the conversation. You've been having a guided discussion — now condense it into the canonical interview tuple. Map your conversational understanding to interview/questions.md letter options (a/b/c/d). If you're unsure of any letter, ask the user one clarifying question before proceeding.

2. Run scripts/route.py with the captured tuple (see commands/init.md step 3 for the JSON shape) to get the methodologies + profile deterministically. Do not interpret the routing table yourself.

3. Produce a ready-to-paste handoff. Output a single fenced block the user can paste verbatim:

   /mck-scaffold:init --decision={"version":1,"methodologies":[...],"profile":"...","stack":"...","project_name":"...","rationale":"<one paragraph>","interview_answers":{"q1":"...","q2":"...","q3":"...","q4":"...","q5":"...","q6":"...","q7":"..."}}
   

   Required fields in the JSON: version (always 1), methodologies (from route.py), profile (from route.py). Strongly recommended fields: stack (from Q5 — without this the installer cannot drop stack strictness files), rationale (one paragraph for the audit log), interview_answers (the captured letters, for reproducibility).

   This is JSON-on-one-line; the JSON content must match schema/scaffold-decision.schema.json. When the user pastes that command, commands/init.md step 1a recognizes the --decision= arg and skips the interview — your conversation is the interview.

   If the rationale paragraph would contain quote characters that break one-line JSON encoding, keep the rationale short for the handoff and write the longer version into docs/decisions/METHODOLOGY.md post-install. Don't fight JSON escaping.

4. Tell the user what just happened. One short paragraph: "I'd recommend X + Y at profile Z because <rationale>. Paste the command above to install. The installer is deterministic — same input always produces the same output."

The picker remains read-only — it does not write files. The handoff is text the user explicitly pastes. The slash command does the write.

Status

Phase 2 complete. Conceptual frame, conversational guidance flow, and routing references are all functional. The interview executor (/mck-scaffold:init) consumes the same interview/questions.md and interview/routing.md files this skill references — they are the single source of truth for both the conversational and the deterministic paths.