Skip to content

docs(skills): clarify who runs the multi-scene dispatch pipeline#448

Open
vanceingalls wants to merge 2 commits intofeat/visual-style-promptfrom
vance/skill-multi-scene-dispatch
Open

docs(skills): clarify who runs the multi-scene dispatch pipeline#448
vanceingalls wants to merge 2 commits intofeat/visual-style-promptfrom
vance/skill-multi-scene-dispatch

Conversation

@vanceingalls
Copy link
Copy Markdown
Collaborator

What

Brief description of the change.

Why

Why is this change needed?

How

How was this implemented? Any notable design decisions?

Test plan

How was this tested?

  • Unit tests added/updated
  • Manual testing performed
  • Documentation updated (if applicable)

vanceingalls and others added 2 commits April 23, 2026 11:04
The parallel fan-out in Phase 1 and Phase 2b requires the Agent/Task tool,
which in Claude Code is only available to the top-level conversation agent.
Dispatched subagents cannot spawn further subagents, so they can't drive the
parallel pipeline.

Add guidance in SKILL.md and references/multi-scene.md:
- Top-level agent: run the full parallel pipeline
- Nested subagent: author fragments sequentially, still run through the
  assembler and lint gates, note the constraint in the final report

This surfaced in eval runs where the skill was invoked from a nested
general-purpose subagent and the multi-scene dispatch silently fell back to
serial authoring with no indication to the caller.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Single-pass authoring under-decorates when a composition has more than one
scene. The eval run showed p1 branch (3 scenes, single-pass with expansion)
regressing vs main on background-layer density — the authoring agent
followed the expansion's scene-element list literally and omitted the
house-style atmosphere stack (ghost type, radial glow, framing rules).

Giving each scene its own subagent via the multi-scene pipeline keeps
per-scene density and decoration consistent, because each fragment agent
reads house-style.md and applies "2–5 decoratives per scene" to its one
scene without competing for context.

The threshold is now "2 or more scenes". Single-pass is reserved for true
one-scene compositions (title cards, standalone overlays).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Copy link
Copy Markdown
Collaborator Author

vanceingalls commented Apr 23, 2026

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant