TDPilot: Skill for Claude Code

TDPilot Brain Builder

Use this skill for non-trivial TouchDesigner construction or transformation. Load tdpilot-production separately only for production/show-safe work.

Core Rule

Classify the whole request before choosing a brain route. Exact validated patterns may use td_brain_plan; artistic, multi-domain, spatial, camera/depth, fog, or architecture-shaped requests go directly through ground → author → propose. Execute only a valid plan with complete intent coverage and fully lowered semantic edges.

One Build Loop, Two Brain Routes

1. Inspect only what changes the plan

• Start with td_get_focus, the target root, direct connections, and relevant parameters/errors.
• Use td_get_state_vector only when broad project health or performance is relevant.
• Put independent reads in td_tool_batch; never batch mutations as a fake transaction.
• Establish the exact target root, protected paths, desired stable output, and any external device/file dependency.

2. Extract the requirements

List every material capability, input, output, constraint, spatial feature, behavior, quality, binding, and validation need. Each required facet must have evidence from a concept/module, parameter, connection, binding, or assertion. Do not let one recognized keyword hide the rest of a multi-domain request.

3A. Pattern-shaped route

Use this only when TDPilot has an exact validated topology or technique composition for the complete request:

1. Call td_brain_plan(..., detail_level="summary").
2. Inspect route, intent_coverage, required operators, output, risks, and validation contract.
3. Execute only when required coverage is complete, operations are nonempty, and unresolved_semantic_edges is empty.
4. If it returns host_authored, clarify, incomplete coverage, or an unsupported combination, continue with route 3B. Never execute the partial match.

3B. Concept-shaped route

1. Before taste-critical authoring, recall compatible evidence with td_memory_recall and td_knowledge_recall, or call grounding with include_memory=true.
2. Call td_brain_ground with the full intent, target root, constraints, desired output, mode, and trace_level="summary".
3. Read the returned grounding_id, task facets, live state, corpus evidence, ranked candidate operators, parameter semantics, availability, recall bundle, exemplars, and exact authoring_contract.
4. Author the smallest draft that covers every required facet. Candidate operators are ranked suggestions, not permission to invent unavailable or undocumented operators: additions still require an official card, live availability, and parameter semantics.
5. Call td_brain_propose with the same grounding_id, draft_schema_version="2", and detail_level="summary".
6. On rejection, fix the machine-readable cause and re-propose. On acceptance, check server-recomputed coverage and any stripped parameters. A host claim that coverage is complete is never authoritative.