althing-poll

Run a quick synthetic poll using the Althing MCP server. One question, a handful of AI personas, one synthesized writeup.

Question: $ARGUMENTS

What to do

1. Resolve the question. If $ARGUMENTS is empty, ask the user for the question before continuing. Do not invent one.

2. Pick personas. Default to the built-in zero-config persona set by omitting the personas argument. Only build a custom list when the user names a specific audience (e.g. "three enterprise SRE personas"). In that case, either:

   • Call mcp__althing__list_persona_packs and get_persona_pack to pull an installed pack, or
   • Hand-author 3–5 dicts with name, age, occupation, background, and personality_traits.

   Keep panels small — run_quick_poll caps at 3 personas when running in sampling mode (no API key).

3. Run the poll. Call mcp__althing__run_quick_poll with:

   • question: the user's question
   • personas: omit for the default set, or pass your custom list
   • synthesis: leave as true (default) — the synthesis pass is the point of this command

   Do not pass model unless the user explicitly asked for a specific one. The MCP server picks a sensible default (haiku in BYOK mode, or the host's own model in sampling mode).

4. Report the result. Present the output in this shape:

   Question: <the question>
   Panel: <n> personas (<mode: BYOK model-name | sampling>)
   
   ## Themes
   <bulleted list from the synthesis>
   
   ## Per-persona responses
   - <persona name>: <one-line summary of their response>
   - ...
   
   ## Notable divergences
   <any disagreements the synthesis flagged>
   
   Cost: <$X.XX from the response's `cost` field, or "n/a" in sampling mode>
   

   Keep each persona summary to one line. The full transcripts are in the raw tool output — don't re-print them.

Guardrails

• Synthetic panels are exploratory, not validating. End the report with a single-line caveat: "Synthetic panel — directional signal only, not a substitute for real user research."
• Don't loop. If the user wants a second question, that's another /althing-poll invocation, not a follow-up inside this one. For multi-question structured studies, suggest the /focus-group skill instead.
• Report errors honestly. If the tool returns an error field (missing API key, persona cap exceeded, etc.), show it verbatim and suggest the fix — don't silently fall back.