mm-feature-rnd

name	mm-feature-rnd
description	Feature R&D via a dynamic workflow. Mine a codebase's roadmap, scaffolded/gated features, ADRs, and novel ideas for net-new feature candidates; generate full design specs; score them with a judge panel on value/effort/risk/differentiation; synthesize a ranked HTML feature catalog. Use when the user asks "what should we build next", "feature ideas", "R&D", "find net-new features", or wants a decision-ready catalog of buildable features (not bug fixes). Read-only — produces designs + a catalog, writes no feature code.
version	1.0.0

mm-feature-rnd — Feature R&D Catalog (dynamic workflow embedded in a skill)

Runs a multi-agent dynamic workflow that discovers net-new feature candidates from real project evidence, designs each one, scores them with a judge panel, and renders a ranked HTML catalog you pick from. This is the harness that produced a 17-feature scored catalog for MemoryMaster (2026-06-01).

When to use

• "what should we build next", "feature ideas", "R&D session", "roadmap mining"
• You want grounded, buildable feature proposals (scored, with full specs) — NOT generic brainstorming, NOT bug fixes/refactors (those are out of scope).

How it works (phases)

1. Scout — parallel agents mine: roadmap/backlog files, scaffolded/env-gated half-built features, ADRs/architecture for under-exploited primitives, and novel differentiators vs competitors. Curate to a deduped slate.
2. Design — one full design spec per candidate (problem, design on the real stack, surface, effort S/M/L/XL, risk, differentiation, test plan, deps).
3. Judge — 2 judges per design (product-value + build-pragmatics lenses) score value/effort/risk/differentiation; deterministic in-code tally.
4. Synthesize — executive "build-first" list + themes.

Run it

1. Confirm scope — the repo + that this is features/improvements only (no bugs/cleanup). The embedded SCOUT_SLICES/CONTEXT/GROUPS are MemoryMaster-tuned; for another project edit the copy first.
2. Launch:

   Workflow({ scriptPath: ".claude/skills/mm-feature-rnd/feature-catalog.workflow.js" })
   

   Returns { stats, synthesis{executive_summary, build_first[], themes[]}, catalog[] }.
3. Render the HTML catalog from the .output file via a small Python script (severity/score-ranked cards, filterable by verdict) to C:\Users\pauol\artifacts\YYYY-MM-DD-<slug>.html, then open it. Don't pull the full result into context.
4. Present the top "build-first" features inline; let the USER pick which to prototype. Prototyping is a SEPARATE gated step (worktree builds + tests) — never auto-build features.

Adapting to another repo

Edit REPO, CONTEXT (project + competitors), and SCOUT_SLICES (which files to mine) at the top of the workflow .js. Keep the judge-panel + tally intact.

Guardrails

• Read-only / design-only. Produces specs + a catalog; writes no feature code.
• Features only. Drop anything that is a bug fix, test, refactor, or cleanup.
• Trust scores, not verdict prose. The 2-judge median can skew labels pessimistic; rank by the numeric score + the synthesis build-first list.
• Verify "already exists" claims before building later. Scout/design agents hallucinate both directions about what's already built — any follow-up build MUST grep/read the real code first (see ref mm-712f).