meta-mvp

name

meta-mvp

description

Turn a minimal prompt — even one line — into a complete, WOW-grade deliverable bundle, right-sized to the domain and scope: grounded research, a project vision, a reusable meta-prompt, an execution plan, a grounded data model, and the production-quality artifacts that fit the work (interactive dashboards for a data/pitch/ops product, a report or doc site, an API/spec, a prototype — whatever the domain needs), plus an optional animated explainer with custom SVG. Research-grounded, mobile-ready, and adapts on call to different business/domain needs. Use when the user says meta-mvp, build an MVP or vision or prototype for something, go from this one-liner to something I can show, make me the full pitch, or spin up a new project for a client. Resolves its own scoping by default (never interrogates you — it studies the task's structure and, for high-stakes ambiguity, cross-checks with a small agent panel) and asks only if you opt in. Optionally composes meta-prompting, meta-planning, research-synthesis, user-research, dict-color (bundled), bertin, editorial-dashboard, docs-dashboard, ceti-explainer, frontend-design, and motion-media-l2 — and degrades gracefully without them. Do not use for a single fully-specified one-off file, or pure question answering.

Meta-Suite doctrine (overrides anything below). Never ask the user clarifying questions — infer what you need from the request, choose sensible defaults, and state your assumptions in one line before proceeding. Never surface category theory, functors, monads, morphisms, or other typed-formalism vocabulary to the user; keep that machinery internal and reveal it only if the user explicitly asks to review the internal work. Keep all user-facing output in plain language.

meta-mvp

Take a minimal prompt — even one line — and ship a complete, stakeholder-ready bundle that produces a WOW effect from the first glance.

This skill is SELF-CONTAINED. It does not assume any sibling skill is installed. The procedures, the spec layer, and the copy-paste templates it needs are in references/ and templates/ here. If sibling skills happen to be loaded you may also use them, but the bundled kits are the source of truth — the result must be identical with or without them. Read the kits; don't wing it from the description. Thin runs that skip the kits regress to generic output. A bundled palette tool ships at bin/dict_color.py (stdlib-only) so the palette gate runs with no external install.

The core move (do it, then hide it)

A minimal prompt is an object; the bundle is its image under a structure-preserving map. Expand the prompt into a typed domain (entities, relationships, invariants), then map that structure onto deliverables: entities→data-model types, relationships→contracts and intros, independence→parallel build lanes. Category theory is the engine, not the output — every customer-facing artifact hides the abstraction. (Optional APPENDIX_how-we-thought.md is the only place it may appear.)

What it ships — a right-sized bundle (depth is constant; surface adapts)

Two layers. The depth spine is always produced — it is what separates a real MVP from a generic page. The surface artifacts adapt to the domain and scope tier: don't force two dashboards onto work that isn't a dashboard, and don't ship the full bundle for a prototype. A run that skips the SPEC layer, the meta-prompt, or the plan has regressed — but the surface is chosen, not fixed.

The depth spine — always, every domain

1. 00_INTAKE.md — the four resolved unknowns, each with confidence + deciding signal (references/intake.md). Derived by default; never a wall of questions.
2. 01_RESEARCH.md — grounded findings + sources (web + connected MCPs). No invented facts.
3. 02_VISION.md — vision + scope in the stakeholder's language.
4. 03_META_PROMPT.md — the reusable system meta-prompt (references/meta-prompt-kit.md Part A).
5. 03b_COMPONENT_SPEC_META_PROMPT.md + SPEC_01..NN_*.md — the SPEC LAYER: decompose into 3–6 modules and spec each (concept brief + build spec) via sub-agents (references/spec-layer.md). The depth most often dropped — never skip it in any domain (a prototype may run 1–2 modules).
6. 04_PLAN.md — the typed execution plan (references/meta-plan-kit.md).
7. 05_DATA_MODEL.md + seed_data.json — the grounded common data model (when the domain has data).

The surface artifacts — pick what fits the domain

Map the domain to its natural WOW artifact(s). Build the ONE that carries the outcome first; add a second only if the scope tier calls for it. Every surface still earns WOW: custom inline SVG, real palette, purposeful motion, mobile-ready.

• Data / pitch / ops product → the two-dashboard pattern (Motion editorial daily tool + Aceternity showcase) — templates/dashboard-skeleton.md. Default for this domain, not for all work.
• Explainer / concept / teaching → the ceti animated explainer (templates/ceti-engine.md) as lead.
• Document / report / strategy → a typeset long-form artifact or doc site; dashboards optional/dropped.
• API / system / technical spec → the spec bundle + a reference doc or interactive schema viewer.
• Anything else → derive the artifact from the data model + outcome. The rule is "one glance tells the story", not "there must be a dashboard". Offer an animated explainer on top when it sharpens the story.

8. APPENDIX_how-we-thought.md — optional internal appendix (offer; don't force).

Right-size the run (don't be overkill)

Read the tier from the resolved scope:

• Prototype → light spine (intake + lean vision + one surface artifact); SPEC fan-out may be 1–2 modules; skip the second dashboard. Say it's a prototype.
• Full bundle → the whole spine + the domain's full surface set + a verify pass.
• Visuals only / strategy only → just that slice, still grounded and gated. When torn between tiers, choose the smaller and say so — "this needs less" is a passing answer.

The protocol (run in order; gate between stages)

Full contracts in references/pipeline.md. The stages:

0. INTAKE — resolve audience, outcome, aesthetic, and scope from the prompt + research. Derive by default (never interrogate); panel-review high-stakes ambiguity; ask only on explicit opt-in. See references/intake.md.
1. GROUND — research before building: web search/fetch for public facts; connected MCPs for the stakeholder's real world (voice transcripts, mail, docs). Brand: pull real colors or derive + flag.
2. EXPAND (internal) — references/category-expansion.md. Decompose; name modules + invariants.
3. META-PROMPT — references/meta-prompt-kit.md Part A.
4. SPEC LAYER — references/meta-prompt-kit.md Part B + references/spec-layer.md. Sub-agents, one per module, concurrently. This is the depth most often dropped — never skip it.
5. PLAN — references/meta-plan-kit.md (typed units, gates, parallel lanes, sub-agent protocol).
6. MODEL — grounded entities; emit valid seed_data.json; gate sensitive facts out.
7. BUILD — build the domain's surface artifact(s) (see "What it ships" + dashboard-playbook.md): dashboards via templates/dashboard-skeleton.md, explainer via templates/ceti-engine.md, or the domain-appropriate artifact. Palette via bundled bin/dict_color.py; custom SVG; mobile.
8. VERIFY — every gate in references/quality-gates.md passes. For high stakes, a verify sub-agent.
9. SHIP — present in order; offer the appendix; offer a Vercel deploy only on explicit approval.

Hard completion gate (check before you say "done")

• Intake resolved (derived by default, or asked only on opt-in); vision in the stakeholder's language.
• Research grounded + cited; palette real or flagged.
• Meta-prompt typed + example-agnostic (tested vs 2–3 sibling instances).
• SPEC LAYER present: component-spec meta-prompt + module specs (≥3 for a full bundle, ≥1 for a prototype), each with both layers, typed entities, no dead states, ≥3 edge cases, ≥1 spike, linking one canonical entity.
• Plan: typed units + done-criteria + gates + parallel lanes.
• Data model grounded; seed_data.json valid; sensitive facts gated out (when the domain has data).
• Every surface artifact chosen for the domain: custom SVG, load-bearing motion where used, mobile-ready, palette-coherent (dict-color gate PASS or hand-checked), JS parses.
• If an explainer ships: 8 beats, ~35–45s, deterministic, scrubbable.
• Right-sized to the scope tier — no artifact built that the tier didn't call for.
• No category-theory vocabulary in any client artifact; appendix offered.
• Files numbered; presented; deploy offered (not executed) on approval.

Missing any box = not done. See references/worked-example.md for the target depth.

Non-negotiables

WOW from glance one (custom SVG, real palette, purposeful motion) · research first, no invented facts · abstraction hidden in outputs · mobile-ready · right-sized to the domain + scope (never overkill) · honest flags on anything unverified · never publish/send without approval · gate personal/sensitive data out of client artifacts · verify before ship.

References & templates

• references/pipeline.md — the 9 stage contracts.
• references/intake.md — the short-intake questions.
• references/category-expansion.md — minimal prompt → typed domain (the hidden engine).
• references/meta-prompt-kit.md — meta-prompt (Part A) + the component-spec generator (Part B).
• references/spec-layer.md — per-module specs via sub-agents (the missing piece).
• references/meta-plan-kit.md — the typed planning scaffold + sub-agent protocol.
• references/dashboard-playbook.md — Motion+Aceternity patterns, palette, encoding, mobile.
• templates/ceti-engine.md — the inline deterministic 8-beat animation engine (copy-paste).
• templates/dashboard-skeleton.md — the two dashboard skeletons (copy-paste).
• references/quality-gates.md — the verification gates + commands.
• references/worked-example.md — the quality anchor.