mwguerra/plugins-haiku/prd-interview

PRD Interview Skill

Transforms a rough idea into a PRD through a compiled pipeline: fixed question cards in, an on-disk answer log out, the document assembled one template section per turn, then a countable gap checklist and contradiction votes. No step holds the whole PRD in one window; no step relies on conversation memory.

The pipeline (each stage is a phase file in references/)

The four commands are routers over these phases. references/phase-index.md holds the run-state shape, the per-command step tables, the escalation rows, and the report card.

Interview categories

A brief Launch Framing round (thesis + hero flow) comes first, then the 8 categories:

The "When to Use" column is executed as a LOOKUP, never an inference: the subset table in phase-1-intake.md keys on prdType, an existing-PRD Glob hit, and three stated flags (internal_tool, monetized, ui_facing). Per-category depth (which card blocks, how many rounds) is the depth table in phase-2-interview.md.

Question rounds (mechanics)

• Ask 2–4 questions per AskUserQuestion call, maximum 4.
• Questions, headers, options and multiSelect flags come from the card VERBATIM —

references/questions/index.md names exactly one card file per category.

• Provide concrete options; enable multiSelect for non-mutually-exclusive choices;

the tool adds "Other" itself; keep headers under 12 characters.

• Free text only where the card marks "Free text".
• Prefix every call with progress: "Category <i> of <n>: <name>".
• After a category's last round: a ≤5-bullet summary, then move on.
• Every round appends rows to answers/<category>.md BEFORE the run file updates —

a killed session resumes mid-interview from disk, never from memory.

Launch discipline (baked into the cards and the template — not re-decided per run)

1. Thesis & hero flow first. The PRD opens with a one-line product thesis and the

single hero flow that proves the value; technology comes after, in service of it.

1. Lean v1. The launch scope holds only what serves the hero flow; everything else

goes to the sequenced deferred roadmap, each row with the reason it was held back.

1. Decide over defer. Ambiguity becomes a Decisions row with a named owner; the

open-questions table holds only what the escalation rows could not resolve.

1. No recorded "needed now" justification → the heavy piece is deferred, not built.
2. Vet every dependency. Third-party packages must be open source + actively

maintained (verify by search), recorded in the dependency table; first-party/native preferred.

Default stack

No stack stated → present the profile in references/default-stack.md as the pre-selected "(Recommended default)" option; the user opts OUT, never in. That file is the single source for versions and conventions — do not restate them. /feature runs default to the host product's current stack, falling back to the profile only when the host stack is unknown or the product is greenfield. Whatever is chosen lands as an answers/decisions.md row (owner: the user) and surfaces in the PRD Decisions table.

Run-state and resume

State lives at .taskmanager/prd/run.json + .taskmanager/prd/<slug>/ (answers/, sections/, votes/, gap-report.md). Resume = execute the lowest-seq step not done/skipped; prior conversation is never an input. Full shape, step tables, and escalation rows: references/phase-index.md.

Files (load ONE at a time, on the step that names it)