fix(ce-brainstorm): reduce token cost by extracting late-sequence content

[tmchow]

Summary

• Extract Phase 3 (requirements capture, 130 lines) and Phase 4 (handoff, 90 lines) from SKILL.md into references/ files loaded on demand via backtick path stubs
• Extract visual communication guidance (26 lines) into its own conditional reference, following ce:plan's existing pattern
• Deduplicate interaction rules that were restated verbatim in Phase 1.3
• Update contract tests to verify stubs point to correct reference files

SKILL.md: 387 -> 173 lines (55% reduction, 24.2KB -> 11.8KB)

How the savings work

Phase 3 + Phase 4 make up 53% of the skill but are only needed after the interactive dialogue (Phases 0-2) completes. In a typical brainstorm, 8-17 turns happen before Phase 3 is relevant — each carrying that content in the system prompt for nothing.

The system prompt (where skill content lives) is carried in full on every API call and is never compressed. Extracting late-sequence content to reference files means it's only loaded via Read when actually needed, reducing the per-turn carrying cost during the interactive exploration phases.

Estimated savings per session

Scenario	Pre-Phase-3 Turns	Estimated Savings
Lightweight (9 turns)	~4	~13K tokens (30%)
Standard (19 turns)	~12	~35K tokens (39%)
Deep (26 turns)	~17	~49K tokens (35%)

After both references are loaded in the final few turns, the per-turn cost is roughly neutral — the savings are concentrated in the interactive phases where the most turns occur.

Benchmarking note

We ran eval comparisons but the evals methodology reads SKILL.md via the Read tool into conversation history. This is architecturally different from real skill invocation where SKILL.md is injected as system prompt. The evals confirmed quality parity (both versions produce equivalent brainstorm output) but cannot measure system prompt carrying cost reduction, which is where the savings come from. The theoretical model follows the same pattern validated in #489 (ce:plan) and #509 (document-review).

Test plan

• bun test — 586 pass, 0 fail
• bun run release:validate — metadata in sync
• Contract tests verify stubs point to correct reference files
• Contract tests verify behavioral guarantees in extracted files
• Eval runs confirm quality parity between original and optimized versions

🤖 Generated with Claude Code