Incipit

"here begins" — turn a rough idea into a structured Implementation Brief before any work starts. Instead of hoping you remembered to specify everything, Incipit interrogates the idea for you: it researches what already exists, infers the stakes and form factor, asks only the questions research couldn't answer, drafts the spec a section at a time, breaks it into ordered tasks, and verifies the result with a script before handing it over.

You get files an agent can execute against — brief.md, plus research.md and tasks.md once the stakes justify them — rather than a wall of chat to copy out. Findings carry sources and dates, and the brief cites them, so a version number or a rate limit in the spec is evidence rather than recall. It defaults to software, but the section schema is swappable, so the same flow works for a process, a policy, or a research plan.

skills/    portable agent skills — the flow, for any coding agent   (start here)
webapp/    a local FastAPI + HTMX wizard driving your own model endpoint

Agent skills

Three skills that any skill-capable agent can run — Claude Code, Cursor, Codex, OpenClaw, Hermes, and anything that reads AGENTS.md. No server, no model endpoint, no dependencies; the agent runs the flow itself in the conversation.

Skill	What it does
incipit	Full flow: research → calibrate → clarify → draft → write the three files → verify
incipit-clarify	Just the decision-changing questions, each with an [ASSUMPTION] default
incipit-review	Audit an existing spec, PRD, or brief for gaps, risks, and scope creep

Install

The easiest path is to let your agent do it. From a checkout, ask it:

Install the Incipit skills for this agent, following AGENTS.md.

AGENTS.md gives it the host-to-path table, the build check, and the verification step.

To do it yourself, copy or symlink the adapter for your host out of skills/dist/:

# Cursor, globally — swap `cursor` for claude / codex / openclaw
mkdir -p ~/.cursor/skills
for s in incipit incipit-clarify incipit-review; do
  ln -sfn "$PWD/skills/dist/cursor/$s" ~/.cursor/skills/"$s"
done

Host	Install to
Claude Code	~/.claude/skills/<skill> or .claude/skills/<skill> per project
Cursor	~/.cursor/skills/<skill> or .cursor/skills/<skill> per project
Codex	~/.codex/skills/<skill>
OpenClaw	~/.openclaw/skills/<skill>
Hermes Agent	~/.hermes/skills/software-development/<skill>
Anything reading AGENTS.md	copy skills/dist/agents/AGENTS.md to your project root

Most hosts need a new session to discover the skills. Full details, including per-host cautions and uninstall, are in skills/README.md.

Use

Say "spec this out with incipit" and describe your idea. It infers the stakes and form factor, researches prior art and what's already been tried, asks the handful of questions research couldn't settle (each with a default you can ignore), drafts the sections one at a time, and writes the artifacts under docs/specs/<slug>/. For a hands-off run, say "shoot the moon". Point it at a repo and it reads the real stack, the tracker, and the git history first, so the constraints match what you already have.

A weekend script gets one file; a production service gets three with full requirement traceability. The stakes it infers decide, so the ceremony matches what you're actually building.

If research finds something that already does the job, you get told that instead of a brief for a redundant build.

Nothing is handed over until scripts/brief_check.py passes, which verifies structure, requirement syntax, citation format, requirement-to-criterion coverage, and task ordering — the checks an agent is least reliable at doing by rereading its own output.

skills/ is the authoritative definition of the flow and is standalone — standard library only, no dependency on the web app.

Web app

A local wizard that runs the same flow against any OpenAI-compatible endpoint (Ollama, LM Studio, llama.cpp, vLLM, OpenAI), with SSE progress and a downloadable brief. It is fully self-contained under webapp/.

cd webapp
pip install -r requirements.txt
uvicorn app.main:app --host 0.0.0.0 --port 8911

See webapp/README.md for configuration, the round-table review, and the optional GPU DiffusionGemma backend.

Repo layout

Path	What
skills/src/	canonical skill sources — edit here
skills/dist/	generated per-harness adapters — do not edit
skills/build.py	regenerates dist/; --check fails when stale
webapp/	the FastAPI app, its tests, container build, and docs
AGENTS.md	instructions for agents working in this repo

The two halves are independent and verify separately:

cd skills && pytest    # 108 tests, standard library only
cd webapp && pytest    # 215 tests, offline, with a scoped coverage gate

License

MIT — see LICENSE.