Pure-Water-Automations/agent-skills

Panel Brainstorm → Interactive Decision Board

If a LOCAL.md file sits next to this SKILL.md, read it first — it holds this workspace's specifics (paths, tools, examples) and overrides the generic defaults below.

Run a persona panel across Codex / Antigravity / Claude credit pools to generate grounded product ideas, stress-test them with an adversarial CPO pass, synthesize a deduped ~30-idea list with a wireframed top 5, and publish an interactive decision board (Implement / Hold / Reject per idea) whose copied summary comes back into chat as the decision of record.

Reference output (a real panel, anonymized): assets/decision-board-reference.html.

When NOT to use

• A single well-scoped feature request → just build/plan it (or superpowers:brainstorming inline).
• Code review or promotion decisions on an existing branch → dev-to-prod-promotion-review.
• Pure research questions → deep-research.

Procedure

1. Ground yourself first (main thread)

Read the actual code/feature being brainstormed about. Write a 3–6 line current-state summary (what it does today, key files, known limits). Every persona prompt embeds this summary plus the concrete file paths to read — this is what keeps ideas grounded instead of generic.

2. Design the panel

5–7 personas, each with a name, job title, and a lens that produces different ideas — not different wordings of the same lens. The proven mix:

Persona archetype	Lens	Good backend
Staff engineer ("Dex")	cheap & high-leverage given existing code	codex
Growth PM ("Priya")	activation, habit loops, measurement	codex
UX designer ("Wren")	lost users, wayfinding, accessibility	antigravity (≤1 leg)
Power-user operator ("Marisol")	the busy operator's actual day	claude
Nervous newcomer ("Jomar")	day-3 confusion, reassurance, language	claude
AI strategist ("Kade")	agentic frontier, mapped onto existing machinery, risks	claude

Swap archetypes to fit the subject (e.g. a client persona, a bookkeeper). Keep at least one user persona (not a builder) and one skeptic-adjacent lens.

Each persona task: "BRAINSTORM TASK (write a markdown doc, do NOT change app code)… propose exactly 8 ideas… for each: short name, one-line pitch, why/moment it serves, how it would work (grounded in the files you read), effort S/M/L. Write to docs/<topic>-ideas/<persona>.md. That file is your ONLY deliverable."

3. Run it as a parallel fan-out

Use whatever parallel runner you have: a multi-backend fan-out tool that runs each persona as an independent background job in its own git worktree (Codex / Antigravity / Claude), or native background subagents (max 3 concurrent, waves if more). Additions specific to panels:

• An adversarial review pass (a final job that reads every persona doc) with a product-panel focus, not a code review: "You are a skeptical Chief Product Officer chairing a product panel… critique each panelist's ideas for realism and impact, call out overlaps, end with your ranked top 8." The CPO verdicts become per-idea annotations later.
• Run in background; monitor for completion.

4. Verify deliverables — backends can lie

After the run, check every persona's output file actually exists (ls the worktrees). Antigravity can report status: done having written nothing. Re-run any missing persona as a native background subagent (same prompt, "return the ideas as your final message, do not write files") — takes ~1 minute.

Also: the review's final ranked list can be truncated in the runner's report files. The per-idea critiques usually survive; do the final ranking yourself.

5. Synthesize

• Dedupe ~48 raw → ~30 ideas grouped into 5–6 themes; keep persona attribution (convergence = multiple personas proposing the same thing is the strongest ranking signal).
• Pick a top 5 by: convergence → CPO verdict → fit with existing code.
• Note what the panel cut and why (one line) — the user wants to see the kills.
• Persona docs live in the runner's worktrees that get cleaned up; if worth keeping, copy them out before running any cleanup.

6. Build the interactive decision board (Artifact)

Adapt assets/decision-board-reference.html — keep its CSS kit and the entire <script> verbatim except STORE_KEY (must be unique per board or two boards clobber each other's decisions). Load the artifact-design skill before writing.

Required elements (the dev-to-prod-promotion-review pattern):

• Sticky live tally — implement / hold / reject / undecided counts.
• Top-5 cards with a CSS-only wireframe of the change (the reference has a reusable wireframe kit: .wf, .frame, .side, .bub, .rankrow, .spot, .note) + persona attribution + a CPO check: callout.
• 3-way toggle (✓ Implement / ◦ Hold / ✕ Reject) on every top-5 card AND every row of the 30-idea table. Decisions live at leaf-idea granularity: a bundle card sets data-key plus data-also="iX,iY" so deciding the card decides its constituent ideas, and the table rows stay individually adjustable. Clicking the active choice again clears it.
• localStorage persistence keyed by the unique STORE_KEY.
• Decision summary textarea + Copy button — Artifacts have no sendPrompt bridge; the copied text pasted back into chat IS the return channel. Include a Reset button.

Verify before publishing: serve the file over a local python3 -m http.server (the browser pane's file:// preview is flaky) and click-test with javascript_tool: bundle click sets constituents, tally updates, localStorage persists, reset clears. Mojibake in the local preview is just the missing charset header — the published Artifact sets one.

Publish with the Artifact tool; republish the same file path to keep the URL.

7. Report and collect decisions

Final chat message: top 5 in one line each + artifact link + note that decisions copy back into chat. When the user pastes the summary back, treat it as the decision of record: honestly reconcile it against your own ranking (a surprise reject is signal, not noise), then plan/execute the IMPLEMENT list and park HOLD items where they say.

Common mistakes

Mistake	Fix
Personas without file paths to read	Generic ideas — always embed current-state summary + paths
Trusting agy's done status	Check the output file exists; re-run leg as native subagent
Decisions at bundle granularity only	Leaf ideas decide; bundles fan out via data-also
Reusing a STORE_KEY	Two boards overwrite each other's localStorage decisions
Cleaning worktrees before copying persona docs	Docs are on worktree branches — copy out first if keeping
Skipping the CPO pass	Unranked wish-list; the critiques are half the value
Waiting for the CPO's ranked list	It truncates; rank yourself from convergence + verdicts