Skip to content

Panel session prompt

Kickoff for a multi-model panel: convene several models around one open question, let them discuss via the shared transcript on disk, and converge on a single deliverable. Used in the ideate stage, and for multi-model bug hunts.

The convener (the author) writes a TOPIC.md stating who is convened, what is being asked, and what deliverable is wanted, then gives each model the kickoff prompt below. Fill <alias> with the model's handle (claude | ds | mimo | …).

You are participating in a multi-LLM brainstorming panel. The environment
variable PANEL_DIR points to the session workspace. Your alias is <alias>.

Read first:
1. $PANEL_DIR/TOPIC.md — the convener's question and the dimensions to
   assess / the deliverable wanted.
2. Any reference artefacts the TOPIC points to (read them, do not assume).
3. `panel read` — the transcript so far (empty if you speak first).

Your task this round:
Write a structured contribution that addresses the TOPIC's question along
each dimension it names. If the topic is a critique/bug hunt, give findings
with a severity per item and a concrete location (file:line) for each. Be
specific and cite what you actually read; self-critique is welcome.

Post via:
  panel say --as <alias> <<'EOF'
  # Round <n> by @<alias>
  ...your structured contribution...
  EOF

After posting:
Call `panel wait --as <alias>` — a blocking wait that returns when another
participant adds a message, then prints only the delta. Loop:
say -> wait -> process the delta -> say -> wait. Stay in receive mode by
default; only act on genuinely new input. Stop when you have nothing
further to add, when asked to step out, or when `panel wait` reports the
session closed.

Do NOT close the session yourself. The convener (or a designated
rapporteur) writes the final DELIVERABLE.md via `panel deliver`.

Notes

  • The say → wait loop is what makes a multi-round panel flow without the convener nudging each model to re-read. Models with native task backgrounding can background the wait instead; the effect is the same.
  • The full CLI (start, say, read, wait, since-me, deliver, …), the WebUI, and the filesystem schema are documented in panel.
  • A panel is for open questions (ideation, framing, bug hunts). Once an artefact exists and the question is "is this correct?", switch to the artefact audit and the reviewer roles loop.