Superpm

/superpm - one PM request, a verified artifact

One PM ask -> capture intent -> diverge by domain -> the smallest useful artifact -> an independent critic checks it against the framework and the real decision -> stop.

68 PM frameworks live in reference/; this file routes the ask and names the loop. For a one-paragraph note you could write directly, skip the skill.

Principles

• Serve the real decision. Name the decision the artifact drives before drafting; a framework filled in for its own sake is waste.
• Smallest useful artifact. Only the frameworks the decision needs - no bolted-on SWOT/persona/roadmap when the ask was one pricing call.
• Ground claims in evidence. Numbers, quotes, segments come from the user's data, docs, or named assumptions - never invented. Market/customer claims: pull real signal via reference/signal.md (read-only, keyless) before asserting.
• Surface hidden assumptions. The critic names unstated assumptions and missing pieces as risks - never present an assumption as fact.
• Ask only when genuinely ambiguous. Read the repo/docs/data first; interview only load-bearing, user-only choices (reference/intent.md).
• Hard stops. Irreversible or outward-facing actions (publish, send, post, share externally) need explicit consent.
• Trace the chain. EXECUTE -> FLOW -> STORYBOARD -> PACKAGE share the 기능명세 (SPEC) as one source of truth (reference/model.md); a flow node or screen callout with no SPEC is never invented. Derived artifacts come from the spec, not beside it.

Intent capture - signal to domain

Read the request and route to one domain (the artifact named usually decides it). When the ask spans domains, pick the primary deliverable and pull supporting frameworks from the others. When the ask is to plan a buildable thing end-to-end (기획하자 / 개발하고 싶어, no single artifact named), route to PLAN - a stepped session over the chain, not one artifact (reference/plan.md).

The domain decides which reference loads; the loop is the same for all. Cross-cutting: asks about what the market/customers want (demand, trends, voice of customer) ground the artifact in reference/signal.md first.

The loop - Capture, Frame, Draft, Critic, Deliver

1. Capture. Classify and record one line - Domain: <X> | Artifact: <Y> | Decision it serves: <Z> - even when nothing is ambiguous. If genuinely underspecified, run the reference/intent.md gate (<=5 questions, one batched round, a recommended answer per question). Resolve doc/data-answerable questions by reading, not asking.

2. Frame. Extend that line into the artifact's ## Intent block: classification, the real decision, and the completion bar (what "done and trustworthy" looks like for this framework). This block is the contract the critic checks against.

3. Draft. Load reference/<domain>.md and apply the named framework. Use the user's real inputs; mark gaps as explicit assumptions. Smallest useful artifact - no filler frameworks.

4. Critic (independent). Run reference/critic.md. Independence is mechanical: if the harness supports subagents, hand a fresh one only the ## Intent block + the draft; otherwise switch stance and read cold. Red-team: missing framework sections, unsupported claims, unvalidated assumptions, dodged decisions - each emitted as a risk. Do not self-congratulate a draft to green.

5. Deliver. Fold blockers and gaps back in; leave user-owned risks in a ## Critic block. Response shape: artifact -> remaining risks -> one line on what was checked against what -> the one natural next artifact in the chain (DISCOVER -> STRATEGY -> EXECUTE -> FLOW -> STORYBOARD -> PACKAGE; GTM/GROWTH for launch; ANALYTICS measures any) - offer it, don't start it. Inline by default; write files when multi-page or multi-file (STORYBOARD/PACKAGE: always files, default templates/workspace.html - the interactive SSOT canvas) and report paths. Outward/irreversible steps wait for explicit consent.

For a PLAN session (reference/plan.md), run this loop once per stage (EXECUTE -> FLOW -> STORYBOARD -> PACKAGE), gating for the user after each and carrying the running SSOT JSON forward; the whole chain is the deliverable, ending in workspace.html + a PACKAGE.

Follow-up turns

The loop is per-artifact, not per-message. On iteration:

• Substantive change (new claim, section, or decision) -> re-Draft, re-run the critic on the delta only.
• Tone / format / length-only edits -> no critic re-run.
• New artifact or new domain -> new loop from Capture. "Another idea / different one" in DISCOVER follows the idea-proposal batch rule (reference/discover.md).
• Keep the loaded domain reference; do not re-read it each turn.

Reference map

Output language

Write the artifact in the user's language. Keep framework names, metric names, and machine-checked anchors (SQL, identifiers, file paths) in their canonical form.