page-one-meeting

Page One

Run a simulated page-one editorial meeting: a chair (you) convenes a small panel of editor personas who debate a pitch with each other — not just report to the chair one at a time — and the meeting ends in a verdict. Default output is a verdict and a shaped pitch. Pass brief as an argument, or have the user ask for it afterward, to extend into a co-authored drafting brief.

Mechanism: the Agent tool with a distinct name per persona and run_in_background: true, then SendMessage to reach agents by name. Agents message the chair unprompted as they finish — no polling. Never use the Workflow tool for this: the debate has to happen between named personalities who can message each other and the pitcher, not a deterministic script.

One thing broke the first time this ran by hand, and this skill exists to prevent it: groupthink. Unguided multi-agent LLM debate has documented failure modes — sycophantic conformity, consensus collapse, lock-in often by the second exchange. The room converged fast and clean, and missed that the whole angle had already been done — a human had to step outside the simulated room to catch it. Phase 4 (mandatory dissent) and the chair's outside check in Phase 5 exist specifically because in-room debate alone didn't catch this.

Phase 0 — Intake (mandatory, abbreviable, no disk writes)

Read whatever raw material the pitcher already gave — if they pasted a page of research and a general direction, that already satisfies most of intake. Silently check it against references/intake-checklist.md, then ask only the 2-3 highest-value missing items at a time, conversationally (ask, don't tell — see managing-editor.md's voice for the model). Use AskUserQuestion only for genuine either/or forks, not the whole checklist. The moment the pitcher says "just start" or otherwise signals impatience, proceed — carry anything unresolved forward as a flagged uncertainty in each persona's briefing, not as a blocker.

If the pitcher names a target outlet, check for pitch-guides/<outlet-slug>.md in this repo (e.g. pitch-guides/vox.md already exists). If it exists, load it, brief the panel with its actual submission format, and shape the final output to match it.

Phase 1 — Cast Assembly (no disk writes)

Pick 3-5 archetypes suited to this pitch's actual tensions, using references/archetype-bench.md as inspiration — not a fixed roster. Invent fresh names and voices every run. Seriously consider a novelty/saturation-checking archetype; its absence is what let an unoriginal angle slip through the first time this ran. Post a one-line "seating the room" announcement naming the cast and why each one is here for this specific pitch.

Phase 2 — Seating / Opening Takes

Issue all N Agent calls in a single parallel message: distinct name per persona, run_in_background: true, prompt built from the spawn template in references/agent-prompts.md. Each persona's prompt includes the raw material, the intake summary plus flagged gaps, and the roster of fellow editors they can message by name.

Wait for opening takes to arrive as agent-message events — this is automatic, not something to poll for. Relay a synthesized digest to the user, never a raw dump of agent messages.

Phase 3 — Open Debate (one round by default)

Send each seated agent the debate-round follow-up from references/agent-prompts.md: message 1-2 named peers directly with a specific, sharp challenge, cap 3 sent messages, then report a revised position to the chair. Digest again. Default to a single round — only repeat if positions are still wildly split after this one.

Phase 4 — Mandatory Dissent Round (structural, not discretionary)

This does not get skipped, and it is not optional even when the room looks unanimous — especially when it looks unanimous.

From the Phase 3 digest, identify which agent(s) moved most toward consensus. Assign dissent duty to one of them, not to whichever persona's archetype is "the skeptic" — asking a built-in skeptic to dissent just produces a rehash of their existing position; forcing someone who just agreed to argue the opposite produces genuinely new reasoning. Use the dissent-round template in references/agent-prompts.md: two required angles every time — (a) steelman the strongest case that the pitch is fundamentally wrong, not a minor note; (b) name the closest existing coverage of this exact angle, or say explicitly if you can't rule out that it's been done. If the agent still believes the consensus holds after genuinely trying, they say so and name the specific evidence that would change their mind.

Record the dissent findings in pitch.md regardless of outcome. The record that the check happened is part of the artifact, not just a gate to pass through.

Phase 5 — Verdict + Shaped Pitch (first disk write)

1. Create pitches/<YYYY-MM-DD>-<slug>/. Derive the slug from the shaped angle the debate landed on, not the raw topic — folder creation waits until this point because the angle isn't knowable before the debate happens.
2. Synthesize the consensus and the Phase 4 dissent findings.
3. Chair's outside check. Before finalizing, briefly step outside the room and reread the shaped pitch as a skeptic would, independent of what the panel just agreed on: is this angle actually saturated — name the closest coverage, or admit you can't rule it out; has the language drifted into sales/scarcity register.
4. For anything under real doubt about novelty or fact-checkability, optionally invoke the managing-editor agent as a second, genuinely separate gate — it wasn't primed by the room's framing and carries live WebSearch/WebFetch. Skippable when the dissent round and outside check already resolved the question.
5. Write pitch.md: verdict using managing-editor.md's vocabulary (Green light / Revise and resubmit / Kill it), the shaped pitch (angle, headline, spine, reader, proof), the cast roster, the debate synthesis, the Phase 4 dissent section (kept even when resolved), the outside check, and any flagged gaps carried from intake.
6. Report a short human summary to the user — not the file pasted in full.

Phase 6 — Drafting Brief (argument-gated, optional)

Triggered by /page-one-meeting brief, a follow-up like "build the drafting brief," or by offering it at the end of Phase 5. Always re-brief the cast fresh from the saved pitch.md rather than relying on agents from earlier phases still being alive — there's no reliable way to confirm that, especially across a new session. Assign one section per persona (mirrors the shape in references/pitch-brief-template.md), each pulling real published analogues via WebSearch/WebFetch. Write pitch-brief.md.

Reference files

Load these only when the phase that needs them is reached:

• references/intake-checklist.md — Phase 0 question menu
• references/archetype-bench.md — Phase 1 exemplar archetypes
• references/agent-prompts.md — Phase 2-4 spawn/debate/dissent templates
• references/pitch-brief-template.md — Phase 6 output shape