Summary

ce-brainstorm now confirms scope with a collaborator-style scoping synthesis before writing the requirements doc, instead of the 20+ bullet audit Phase 2.5 used to emit. The PR also tightens upstream Phase 1.3 dialogue (integration check before exit), replaces clinical jargon throughout the Q&A interaction rules with plain English, and deletes the over-engineered headless mode that produced one-sentence checkpoints on Deep-tier pre-loads and internal ## Summary inconsistencies.

Two failure modes prompted the original change. The Stated / Inferred / Out of scope audit shape reliably produced 20+ bullets for Standard brainstorms, even when granularity rules were followed. And the Phase 0.2 fast-path routing to "announce-mode" produced a 1-sentence checkpoint for richly pre-loaded brainstorm contexts with 20+ items of scope. Both made the confirmation gate feel like rubber-stamping rather than a real checkpoint.

What changed

Phase 2.5 now has a two-stage shape:

1. Internal three-bucket draft (Stated / Inferred / Out of scope): the agent's comprehensive thinking surface. Dissolves into requirements doc body sections when Phase 3 writes (Requirements, Key Decisions, Scope Boundaries).
2. Chat-time scoping synthesis: shaped like what two product collaborators would confirm before writing a PRD. Prose "What we're building" plus three render-conditional sections (Key trade-offs / What's not in scope / Call outs). Each section earns its slot through a section-specific keep test; empty sections are omitted, not padded.

Each bullet passes both an affirmability test (can the user evaluate this without reading code?) and a detail test (1-2 lines max, conversational not documentary). Tier-aware bullet budgets cap total bullets across sections 2-4 (Lightweight 0-1, Standard 2-4, Deep-feature 3-5, Deep-product 4-7) with re-cut on overflow rather than cap-raising.

Phase 1.3 now has an integration check before exit: combine user answers and probe any non-obvious consequences before reaching Phase 2.5, so the dialogue resolves them in flow rather than stacking them as call-outs.

The Q&A interaction rules use plain English throughout: "open-ended" instead of "prose," "unintentionally influence the user's answer" instead of "leak your priors / bias by signaling which dimensions matter."

Headless mode is gone. Brainstorming is interactive by design — the dialogue IS the value. Without it, the skill is just an agent making up requirements with ## Assumptions markers; that's hallucination with a fig leaf, not brainstorming. Removing the mode resolves an internal ## Summary inconsistency at its source and simplifies the spec.

Design decisions

Path A vs Path B gate fires on two signals, not one. Path A (1-3 sentence announce-mode, end turn) fires only when tier is Lightweight AND no questions fired. Standard and Deep tiers always get Path B regardless of question signal: substance earns the checkpoint, not interaction history. The earlier single-signal gate let richly pre-loaded brainstorm contexts hit Path A and produce 1-sentence checkpoints for 20+ items of scope.

Affirmability and detail tests both apply per bullet. The count cap addresses bullet volume; the detail test addresses bullet bloat. Without the detail test, an agent can hit the count cap by compressing horizontally (fewer bullets) without compressing vertically (less per bullet), and the cap becomes meaningless.

Pre-flight re-review is a single mental act, not a checklist. Re-read the draft as the user would read it before emitting. Catches "reads like a doc preview" drift the keep tests miss. Heavy multi-step checklists become performative; one act tied to the user's reading frame is the right forcing function.

Confirmation phrasing sets expectation. "Confirm and I'll write the requirements doc next, drawing on our dialogue and this synthesis" tells the user what happens on confirm. The previous "Confirm to proceed" was ambiguous about what "proceed" meant.

Section heading is "What's not in scope," not "What's not in V1." The old heading presumed versioned-software brainstorms. The skill supports non-software topics (naming briefs, decisions) and unversioned work. Agent-facing guidance and worked-example bullets were also generalized so the agent does not default to versioning language.

Phase 1.3 integration check is the upstream side of a defense-in-depth pair. Before exiting dialogue, the agent mentally combines user answers and probes any non-obvious consequences. Combination effects ("if X applies per-channel AND no warning on delete, then rule-delete silently loses pause state") get resolved in dialogue while the user is in flow, instead of stacking up as call-outs at the exit gate. This pairs with the downstream call-outs framing tightening (call-outs are bounded to genuine residuals: post-dialogue consequences, silent agent inferences — explicitly NOT "questions the agent should have asked but didn't").

Headless mode removed. Brainstorming requires dialogue with a synchronous user — the dialogue IS the value. No legitimate use case to defend: LFG should refuse and prompt the user when a feature description is too vague for autonomous execution, not fake a brainstorm; skill-to-skill invocation should write the doc from context directly if it has the context. The previous headless mode produced wrong-shaped output (1-sentence checkpoints on Deep-tier pre-loads) and internal spec inconsistencies (Codex correctly flagged that "skip stage 2 entirely" left no defined step producing the ## Summary). Deleting the mode resolves the inconsistency at its source.

New Interaction Rule 6: open-ended questions earn their place when they're specific enough to elicit a substantive answer. Apply Rule 5 silently (don't narrate the form choice); the question itself must give the user something concrete to anchor on. "What's your take?" is too thin; rigor-probe-style specificity earns it ("What's the most concrete thing someone's already done about this — paid for it, built a workaround, quit a tool over it?"). Anti-patterns include narrating the choice, "in one sentence" framings, yes/no traps, and AI-slop warmth wrappers.

Plain-English Q&A vocabulary throughout. Replaced "leak your priors," "bias the answer by signaling which dimensions the agent considers relevant," and "use prose" with plain phrasings: "unintentionally influence the user's answer," "ask it open-ended." The technical term "priors" (Bayesian) was confusing in a user-facing context, and "prose" conflated form with the property we actually care about (open-endedness). "Prose" preserved where it means "flowing text" (vs bullets / diagrams).

Soft-cut blocking question options use plain English. Old: "Proceed with the current revised synthesis" / "Stop and redirect — discuss further before writing the doc." New: "Proceed and write the requirements doc" / "Hold off — keep discussing before the doc." The "redirect" verb collided with the unrelated self-redirect mechanism. The same fix propagates to ce-plan's soft-cut menu in a parity commit.

Test plan

bun test (1340 pass) and bun run release:validate pass.

Behavioral changes to skill prose are not exercised by automated tests. Verify by running /ce-brainstorm on:

• A tight Lightweight prompt with no dialogue (expect Path A: 1-3 sentence announce, end turn)
• A Standard prompt with multi-turn Q&A (expect Path B: full scoping synthesis with confirmation gate, even if zero call-outs survive)
• A richly pre-loaded Deep-feature context where no questions fire (expect Path B: tier guard routes correctly, not Path A; this was the bug case)
• A prompt that previously produced 20+ audit bullets (expect compression to scoping synthesis with substance proportional to dialogue, capped at tier ceiling)
• Any open-ended question the agent asks: should be specific enough to anchor a real answer, with no "in prose, because options would leak my priors" narration