td-feature-design — turn an idea into an implementable spec

A spec is good when a builder can implement it in one file + one test without re-deciding anything. Aim for that bar.

Procedure

1. Read the context first. docs/ROADMAP.md (the idea is usually already scoped there), CLAUDE.md (the conventions), and 1–2 neighbour tools in the target src/tools/layer*/ for the exact file pattern.
2. Pick the layer / altitude:
   • Layer 1 (src/tools/layer1/) — an artist tool that builds a whole wired+arranged network (goes through orchestration.ts).
   • Layer 2 (src/tools/layer2/) — a building block (connect, control panel, animate, external IO).
   • Layer 3 (src/tools/layer3/) — atomic node CRUD / inspection / raw-Python escape hatch.
   • src/tools/vault/ — Obsidian-vault bridge. A src/prompts/ entry — when the value is guidance to the model, not a deterministic build (multimodal / natural-language / critique ship as prompts).
   • Decide the file path now: src/tools/layer<N>/<camelCaseName>.ts.
3. Design the Zod input schema — a param table: name, type, default, enum values, notes. Defaults matter: device-sourced inputs (camera/audio) default to a synthetic/file source (live device is opt-in) because device capture can hang TD on a macOS permission modal.
4. Lay out the TD network topology — the operators it creates, how they wire, and the live controls it exposes. Reactive features must end on a Null CHOP ready for bind_to_channel, reusing the shipped binding path instead of inventing one. Verify every operator type against the KB (tdmcp://operators/… or search_operators) — never invent a type.
5. Plan the bridge approach — almost always a Python payload via buildPayloadScript (__PAYLOAD_B64__) executed through the client and parsed with parsePythonReport. Propose a new REST endpoint only when streaming or performance genuinely demands it.
6. Sketch a UI wireframe when the feature has a surface (control panel / control surface / phone remote / web dashboard / chat UI): an ASCII layout or component list naming each control and what parameter it drives.
7. List probe-first risks — anything to validate live before locking the API: platform-specific operators (Syphon/Spout/NDI/Video-Stream-Out exist only on some OS/licenses), KB-missing operators (the KB lags ~14 recent ops; ~22 dir(td) names aren't createable), device permissions, and time-dependent chains that read 0 on a paused timeline.
8. Write the test plan — what the offline msw unit test should assert (operators created, params set, wiring, the returned shape).
9. Write integration notes — exactly which shared files the integrator must edit (layer*/index.ts, src/cli/agent.ts command name + flags, docs regenerate automatically).

Output

One file per feature at _workspace/01_design_<feature>.md with these sections in order: Summary · Layer + target path · Input schema (param table) · Network topology · Bridge/Python approach · UI wireframe (if any) · Probe-first risks · Test plan · Integration notes.

Anti-patterns

• Don't design a feature that needs the builder to edit shared files — that's the integrator's job; note it instead.
• Don't specify operators you haven't confirmed exist; mark them UNVERIFIED — probe live.
• Don't rebuild a shipped primitive (binding, tempo clock, preset/cue engine) — wire into it.
• Don't over-specify defensive params for states that can't occur; match the feature's real surface.