26 atom kinds
From vision to deploy_target. A closed vocabulary covering the full surface of a typical PRD.
AtomPRD
Product requirements as atoms — small, identifiable, self-contained units that LLM agents can read, diff, and translate into running code.
Why AtomPRD
Section titled “Why AtomPRD”A 40-page Word PRD does not fit in a context window. AtomPRD shreds requirements into atoms — 500–2000 token units with stable IDs and typed relationships — so an LLM agent only pulls the slice it needs and so a diff can land on a single feature instead of a chapter.
2 layers
Intent is prose — what and why. Behavior is structured — how. Five kinds carry both.
4 provenance patterns
Every Layer 2 field is tagged: A hand-authored, B convention default, C inferred, D LLM-suggested.
MCP-native
Hosted MCP gateway. Cursor, Claude Code and Cline pull atoms as context — no copy-paste prompting.
Quick links
Section titled “Quick links”- Install + first project — ~30 seconds.
- The 26 atom kinds — reference table.
- Two-layer model — Intent + Behavior.
- Four patterns — A / B / C / D provenance.
- User flow → Overview — end-to-end systematic journey.
- Codegen → SenLang — the reference target.
- MCP server — IDE / agent integration.
What you can do
Section titled “What you can do”- Ship a PRD an LLM can act on. Pull atoms, diff atoms, generate code from atoms.
- Layer 2 lowers to runnable.
feature.behavior.steps→ SenLang Action chain.criterion.behavior.before_after→ scenario fixture. Deterministic. - Mix automation with judgement. Pattern C (free, deterministic) + Pattern D (LLM, reviewed via diff). Lock fields you trust.
- Stay portable. Spec is MIT. Reference codegen is OSS. Hosted authoring is the commercial layer.