startup-skills

startup-skills

Claude Code skills that take a vague startup idea and produce a complete, research-grounded founder artifact pack — the kind of package you'd walk into YC or a deep-tech VC meeting with.

See a finished pack, live

→ dlmastery.github.io/startup-skills

One run, generated end to end: 64 documents · 113 rendered visuals · 12 HTML infographics, every claim traced to a source tag.

The system runs in five modes, in order:

1. Grill (grill-me) — a YC-partner-grade interrogation that turns a vague idea into a sharp founder brief. Nothing else runs until the brief exists.
2. Research (startup-research) — deep web research: landscape, competitors, capability table, survey. Every later artifact must cite this layer; no invented market numbers.
3. Generate + critique — seven generator skills, each with a strict output contract: startup-strategy, startup-product, startup-tech, startup-narrative, startup-validation, startup-financials, startup-visuals. Every generated phase passes through startup-critic — a three-persona red-team panel (skeptical deep-tech VC, domain PhD, elite operator) with up to three revise rounds — before its gate counts as passed.
4. Audit (startup-audit) — coverage audit against the artifact manifest, generation of anything missing, and a final pack index. This replaces the "I don't see it, regenerate" loop with a deterministic checklist.
5. Ship (startup-website) — turns the finished pack into a deployable landing site: hero, problem/solution, feature showcase, social proof scaffold, waitlist CTA — copy drawn from the pack, never invented fresh.

startup-forge orchestrates all of the above end-to-end. startup-ingest optionally folds in external sources (YouTube talks, PDFs, podcasts) keeping only novel angles.

Run it end to end (any coding agent)

Point any file-reading, web-searching, file-writing agent at this repo. Nothing here is Claude-specific.

Step 0 — clone and orient.

git clone https://github.com/dlmastery/startup-skills && cd startup-skills

Tell your agent: "Read AGENTS.md, then drive the pipeline in skills/startup-forge/SKILL.md for this idea: ." If your harness loads skills natively, /startup-forge <idea> is equivalent.

Step 1 — grill the idea into a brief. grill-me interrogates the one-liner YC-partner style and writes runs/<slug>/BRIEF.md plus ASSUMPTIONS.md. Nothing else may run until the brief exists. If nobody is available to answer, the skill grills the material instead and logs every open question.

Step 2 — research. startup-research produces the six-file evidence layer with dated citations. This phase needs real web search; if your harness cannot search, stop here rather than fabricate it. Everything downstream cites this layer.

Step 3 — generate the seven layers. strategy → product → tech → narrative → validation → financials, each against its skill's output contract. Phases 3–7 depend on 0–2 but not on each other, so run them in parallel if your harness supports subagents.

Step 4 — critique each phase before it counts. startup-critic runs a three-persona red team (skeptical deep-tech VC, domain PhD, elite operator), up to three revise rounds. A phase is not done because it exists; it is done when it survives this.

Step 5 — visuals. startup-visuals writes the manifest first, then HTML infographics, then images if an image tool exists. Copy the builders rather than writing your own:

cp templates/build_docmanifest.js templates/build_docimages.js runs/<slug>/visuals/
cd runs/<slug> && node visuals/build_docmanifest.js && node visuals/build_docimages.js

Re-run both after every batch — they are the reconciliation step and they print what is still unillustrated.

Step 6 — audit. startup-audit diffs the run against references/artifact-manifest.md, generates what is missing, and writes runs/<slug>/README.md — the pack's front door. Done means the manifest is satisfied, never memory.

Step 7 — ship the site.

cp templates/site.css templates/build_site.js runs/<slug>/
cd runs/<slug> && node build_site.js        # builds index/product/evidence/pricing/about

Then publish (this makes the repo's contents public — confirm first):

touch .nojekyll
gh api -X POST repos/<owner>/<repo>/pages -f "source[branch]=main" -f "source[path]=/"
gh api repos/<owner>/<repo>/pages --jq .html_url

Record the live URL in the run README, in audit/COVERAGE.md, and in this repo's root README — the last one is the one that gets forgotten.

Commit at every step, not at the end. These runs are long and get cut off mid-phase; uncommitted work is indistinguishable from work never done.

What you get

What the pack enforces

• Every number cites research/sources.md or carries an (assumption: basis) tag.
• Every artifact opens with an orientation block: what it is, why it exists, how to read it, what it depends on.
• Mechanism over adjective — no claim rests on "powerful" or "seamless".
• ASD-STE100 prose — every sentence follows Simplified Technical English: at most 25 words, active voice, approved words, no dash or semicolon inside a sentence. templates/check_ste.js --gate lints it at every gate, and this is what removes the generated-text signature. The pack binds a declared subset of the standard, named in section 0 of references/writing-style.md, rather than a claim of full conformance.
• Gaps are named, not hidden — the run README states an honest completion count.
• No fabricated traction — no invented logos, testimonials or metrics, ever.

Install (harness-agnostic)

The skills use the standard SKILL.md format (agentskills.io spec) and are written to run under any coding agent:

• Claude Code — install the repo as a plugin (it ships .claude-plugin/plugin.json), or copy/symlink skills/* into ~/.claude/skills/ or <project>/.claude/skills/.
• Codex / Copilot CLI / Gemini CLI — copy/symlink skills/* into ~/.agents/skills/ (the cross-runtime skills directory they all recognize).
• Grok Build / Antigravity / anything else — clone the repo and point the agent at it; AGENTS.md tells any agent how to drive the pipeline by reading the skill files directly, no skill loader required.

Whatever the harness: keep the repo intact rather than cherry-picking single skill folders — references/ (quality bar, writing style, artifact manifest, grill question bank) is the shared contract every skill cites.

Output contract

Each run writes to runs/<slug>/ in the current project:

runs/<slug>/
  README.md                        # the front door: pitch, reading paths, embedded visuals,
                                   #   honest completion count (written phase 0, refreshed every phase)
  BRIEF.md  ASSUMPTIONS.md          # grill-me
  research/                        # startup-research
  strategy/                        # startup-strategy
  product/                         # startup-product
  tech/                            # startup-tech (incl. Mermaid architecture diagrams)
  narrative/                       # startup-narrative
  validation/                      # startup-validation (Steve Blank board set)
  financials/                      # startup-financials
  visuals/                         # startup-visuals (HTML infographics + image prompts)
  ingest/                          # startup-ingest (optional)
  audit/COVERAGE.md                # startup-audit
  website/                         # startup-website (deployable landing site)

The full file-by-file manifest lives in references/artifact-manifest.md — it is the audit skill's checklist.

Design principles (why this pack is shaped this way)

• Grill before generate. Generic inputs produce generic packs. The brief is the single source of truth every generator reads.
• Research before claims. Market sizes, competitor facts, and capability statements must trace to research/sources.md or be marked (assumption).
• Contracts, not vibes. Each generator specifies its artifact's sections, numbering, and quality bar. Artifacts are numbered, priority-ranked, span the full user spectrum, and end with a "recommended next" decision.
• Controlled language. Prose is ASD-STE100, the aerospace industry's Simplified Technical English. Its mechanical rules (sentence cap, active voice, one meaning per word, no mid-sentence dashes) are exactly the rules that generated text breaks, and a linter enforces them.
• Deterministic visuals. Diagrams are Mermaid/HTML (verifiable text), and raster-image needs become prompt files for external generators — never a flaky "hope the image rendered" loop.
• Audit closes the loop. Done is defined by the manifest, not by memory.