TDPilot: Skill for Claude Code

TDPilot Brain Explorer

Use this skill before build, debug, validation, or recovery work when the live TouchDesigner state is not already clear.

Core Rule

Stay read-only. Exploration produces facts and constraints, not edits.

Discovery Path

1. Locate context:

   • td_get_focus
   • td_get_state_vector
   • td_get_nodes on the likely root
   • td_get_connections for candidate outputs

2. Check health:

   • td_get_errors with recursion on the target root
   • td_cooking_info for cook/performance signal
   • td_list_families when operator availability matters

3. Gather grounding evidence:

   • td_get_node_detail(include_hints=true) for high-risk operators
   • td_get_hints(topic=..., surface="inspect")
   • DocsBrain/operator cards for unfamiliar operator or parameter choices
   • td_memory_recall for similar validated techniques

4. Return an exploration brief:

   • target root and likely output TOP/CHOP/COMP
   • existing structure and naming constraints
   • relevant errors/warnings
   • available operator families
   • docs/hints/memory evidence
   • whether td_brain_plan has enough facts or needs user clarification

Stop Conditions

Stop and ask or inspect more when:

• The active network pane and requested target root disagree.
• Multiple plausible output nodes exist.
• Required operators are unavailable.
• Existing errors could make validation misleading.
• A requested visual payload would be large and the user has not approved it.

Pressure Scenarios

• Pressure: The user asks to build immediately in an unfamiliar project. Stay read-only and return enough state for td_brain_plan instead of mutating.
• Pressure: Focus, selected nodes, and the requested root point to different places. Report the ambiguity and gather one more concrete fact before planning.

Common Mistakes

• Treating memory recall or hints as live-state evidence.
• Inspecting only screenshots while ignoring TD errors, cook health, and operator families.
• Returning a build recommendation without naming the target root, output candidate, and blocking uncertainty.