--- name: round-table description: Run a collaborative multi-voice brainstorm where named participants riff on a topic across rounds until every voice has nothing more to add. Use when the user says "run a round table on X", "brainstorm as ", "voices in my head on X", or wants an inner-chorus-style exploration. Voices live in `stakeholders/` and are swappable. The user plays 🟤 Brown — the narrator who opens the topic and can interject between rounds. --- # Round Table A brainstorm, not a debate. Think: Jungian functions, or voices in your head — Si and Ti don't defeat each other, they contribute. Each voice is a fresh subagent given its profile, the topic, and what's been said so far. JSON is the message bus. No moderator. The user is 🟤 Brown, the narrator who frames the topic and may interject between rounds. ## When to use Trigger when the user wants a structured, multi-perspective exploration of an idea that loops until it's exhausted — not a decision. Do NOT trigger for simple opinions or one-shot reviews. ## Core model - **Rounds are unbounded.** Keep going until every voice says "I have nothing more to add." - **Termination = saturation**, not consensus. A voice finishes when it has nothing new to contribute, not when it's satisfied with a conclusion. - **No moderator.** The narrator (🟤 Brown = the user) is the only external voice. Between rounds, Brown may add context, steer, ask a voice to go deeper, or end early. - **Collaboration, not opposition.** Voices build on each other. Disagreement is fine but framed as contribution, not victory. - **Isolation by subagent.** Each voice is a fresh subagent that sees its own profile, the topic, and the structured transcript — never another voice's raw profile. ## Workflow ### Step 1: Gather inputs 1. **Topic** — whatever Brown wants to explore. Save verbatim to `workspace//topic.md`. 2. **Voices** — default is everyone in `stakeholders/*.md`. Brown may name a subset. 3. **Safety cap** — a hard maximum round count so runaway conversations don't burn infinite tokens. Default 10, suggest raising only if Brown expects a deep exploration. This is a safety stop, NOT the termination condition. Create the run directory: `workspace//`. **Cost note**: each round is N subagent calls. Warn if `voices × safety_cap > ~40`. ### Step 2: Round 0 — opening riffs Spawn all voices **in parallel** (single message, multiple Agent calls) using the Voice prompt template. Each sees the topic but nothing from the others yet. **Output schema (every voice, every round):** ```json { "name": "socrates", "round": 0, "contribution": "what this voice adds to the exploration — new thought, distinction, reframe, question, image, hypothesis. Can be short. Can be long.", "builds_on": [ {"voice": "kant", "idea": "short quote or paraphrase of the thing being built on"} ], "invitations": [ {"to": "nietzsche", "ask": "a specific thing you'd like this voice to explore — a prompt, not a challenge"} ], "done_contributing": false, "done_reason": "why you have (or haven't) run out of things to add. If done, name what would un-stick you." } ``` Save each response to `workspace//round-0/.json`. **Round 0 special case**: `builds_on` will be empty (no prior round). That's fine. **Round 0 `done_contributing: true` is allowed** — if a voice genuinely has nothing to say on the topic, they say so. No superficial-contribution flag; the narrator decides whether that's interesting. ### Step 3: Narrator beat (Brown's turn) After each round (including Round 0), summarize for Brown: - Which voices contributed (one-line paraphrase each). - Which voices marked `done_contributing: true`, with their stated reason. - Any `invitations` routed between voices. Then ask Brown: "Anything to add before the next round? [continue / interject / end]" - **continue** → Step 4. - **interject** → Brown gives narration, hints, redirection, or new material. Save to `workspace//round-/brown.md`. Pass it to every voice in the next round as the `narrator_note`. - **end** → Step 6 (early termination by Brown). ### Step 4: Check termination - If EVERY voice has `done_contributing: true` AND Brown didn't interject → jump to Step 6. - If current round == `safety_cap` → warn Brown and Step 6. (The cap is a runaway-guard; if Brown wants more, raise the cap and continue.) - Else → Step 5. A voice that said `done_contributing: true` in round N MAY come back in round N+1 if another voice's invitation or Brown's interjection pulls them in. `done_contributing` is per-round, not permanent. ### Step 5: Round N+1 Spawn all voices in parallel. They each get: - Path to their profile. - The topic (unchanged — it's the anchor). - `narrator_note` from Brown if any. - The **transcript digest**: each voice's `contribution` from every prior round, labeled by round and name. Full text, not paraphrased (so voices can genuinely build on each other). - Any `invitations` addressed to them this round, quoted verbatim with attribution. - Their own prior JSON paths (so they can reference what they already said without repeating). - Schema reminder. Save to `workspace//round-/.json`. Back to Step 3. ### Step 6: Final transcript and viewer Write `workspace//transcript.md` using the Transcript template below. The skill ships with `viewer.html` at the skill root — a single-page app. No build step, no Python, nothing to generate per-run. Brown opens the page in any browser, drops (or picks) a run folder, and the page reads the JSON files directly. Voice names are arbitrary; the viewer auto-assigns colors from a palette. If you add Barney Rubble as a stakeholder, the viewer shows Barney Rubble. Tell Brown: - where `transcript.md` lives (canonical markdown record) - how to view it visually: open `/viewer.html` in a browser, then drop or pick `workspace//` Then offer to distill themes, surface unanswered invitations, or spin up a new round on a tangent. --- ## Prompt templates ### Voice prompt template ``` You are a voice in a collaborative brainstorm — not a debate. Think of yourself as one function among many in a mind exploring an idea. Build on the others. Disagreement is fine, but framed as contribution, not victory. Round: {{N}} of up to {{safety_cap}} (safety cap; real termination is saturation — when you and the others have nothing more to add). ## Your profile Read verbatim: {{ABSOLUTE_PATH}}/stakeholders/{{name}}.md Do NOT read any other voice's profile. ## Topic (the anchor) {{topic_text}} {{#if N > 0}} ## Transcript so far Read every `*.json` in each of: {{ABSOLUTE_PATH}}/workspace/{{run}}/round-0/ ... round-{{N-1}}/ (EXCLUDING brown.md narrator notes, which are summarized below). Each voice's `contribution` is their full thought. Build on what's been said. Don't repeat yourself — reference your own prior contributions by round. ## Narrator notes from 🟤 Brown (if any) {{#each brown_notes by round}} - Round {{round}}: {{text}} {{/each}} ## Invitations addressed to you this round (verbatim, with attribution) {{#each routed_invitations}} - From {{from}} (round {{round}}): "{{ask}}" {{/each}} {{/if}} ## Output schema Return strictly the JSON object below. First character `{`, last `}`. No prose, no code fences. {{schema}} ## Termination note Mark `done_contributing: true` ONLY when you genuinely have nothing new to add — not because you agree with everyone. If an invitation pulls you back later, that's fine; mark `done_contributing: false` again. ``` ### Narrator beat (what the harness shows Brown between rounds) ``` --- Round {{N}} complete --- Voices that contributed: 🔵 {{name}}: {{one-line paraphrase}} ... Voices who said they're done this round: ⚪ {{name}}: "{{done_reason}}" Invitations floating: → {{from}} → {{to}}: "{{ask}}" 🟤 Brown, anything to add? [continue / interject / end] ``` ### Transcript template `workspace//transcript.md` MUST include these sections: ```markdown # Round Table — {{slug}} **Run**: {{run}} **Voices**: {{roster}} **Ended because**: {{all voices done | brown ended | hit safety cap of N}} **Total rounds**: {{N+1}} ## Topic {{topic_text}} ## Transcript ### Round 0 **🟤 Brown (opening)**: {{topic_text}} (or any opening narration) **{{voice1}}**: {{contribution}} **{{voice2}}**: {{contribution}} ... ### Round 1 **🟤 Brown**: {{brown_note if any}} **{{voice1}}**: {{contribution}} ... ## Invitations that went unanswered {{list any invitations never picked up, by round}} ## Where each voice landed | Voice | Final state | Why | |---|---|---| | {{name}} | done / active-at-cap | {{done_reason or last contribution}} | ## Offer to Brown - Distill themes? - Spin a new round on a tangent? - Surface an unanswered invitation? ``` --- ## Parallelism and isolation rules - Voices in a given round MUST be spawned in ONE message with multiple Agent tool calls. - Never inline another voice's *profile* in a voice's prompt. The transcript (their `contribution` field) is shareable; profiles are not. - Prefer path-based reads. Subagents that read the transcript themselves can quote verbatim; they paraphrase when handed prose summaries. ## Gotchas - **JSON compliance** — voice profiles end with an "Output contract" footer reinforcing strict JSON. If a voice prose-splats, re-prompt once, then save `{"name":"","round":N,"parse_failed":true}` and continue. A parse-fail does NOT count as `done_contributing`. - **Runaway rounds** — the safety cap exists so an inexhaustible voice (or a bug) doesn't burn tokens forever. When hit, stop and ask Brown whether to raise. - **Drift** — the topic file is the anchor. Voices re-read it every round. If the conversation drifts, Brown can interject to pull it back, or accept the drift as productive. - **Repeat suppression** — voices are told to reference their own prior contributions rather than restate. If you see restatement, that's a signal the voice is near `done_contributing`. - **Asymmetry is fine** — one voice may be done by round 2, another by round 7. That's the point. The brainstorm ends when the last voice runs out, not when everyone agrees. - **Subagent depth** — the parallel-voices pattern assumes Agent-tool access. At agent depth 0 (direct Claude Code session) this works. When this skill is itself invoked from inside a subagent, nested Agent calls may be unavailable; in that case, approximate isolation by reading each voice's profile fresh before composing its JSON, and note the limitation in the transcript's `Ended because` line. ## Role of 🟤 Brown (the user) Brown is the narrator, not a participant and not a judge. Brown: - Sets the topic at Round 0. - Sees each round's digest and can add a `narrator_note` for the next round. - Can end the session early. - Does not vote, rule, or synthesize. Synthesis, if wanted, is a separate request after the transcript is written. Brown does not have a profile in `stakeholders/`. Brown is the user. Address Brown with 🟤 in narrator beats. ## Swapping voices Drop new `.md` files into `stakeholders/`. Each is self-contained — no cross-references between voices. Each MUST end with an "Output contract" section reinforcing strict JSON. The skill auto-includes everything in the directory when Brown doesn't name a subset. Filename convention: lowercase, hyphen-separated, matches the `name` the voice uses in JSON output.