| name | call-intel |
| description | Pulls calls from the user's call intelligence tool (Step 1 of the ICP refinement playbook). Routes to Attention by default and warns the user if they don't use Attention before falling back to Gong, Chorus, Fathom, Fireflies, Granola, Otter, Avoma, or another recorder. Use when the user wants to fetch the last 90 days of demo/disco/intro calls (won, lost, stalled), or pull transcripts for ICP, persona, or trigger analysis. |
Step 1 — Pull calls
Pulls every demo, disco, and intro from the configured time window (default 90 days), across all stages (won, lost, stalled, no-decision).
Routing
Read ./icp-output/stack.json. Branch on stack.call_intel.tool:
attention → load attention.md from this folder.- anything else → load
fallbacks.md from this folder and pick the matching adapter.
What to pull
- Call ID
- Date
- Stage in pipeline (disco / demo / intro / negotiation / closed-won / closed-lost / stalled)
- Account name
- Attendees (name, title, email, company)
- Recording URL
- Full transcript (this is the key payload)
- Duration
- Any AI-generated summary the tool produces
- Deal stage at time of call, if available
- Outcome: won / lost / stalled / no-decision / open
Output
Write each call to ./icp-output/calls-raw/<call-id>.json with the fields above. If the tool returns transcripts inline and the volume is small (< 200), holding in memory is fine — but persisting to disk makes the run resumable.
Update state.json:
{
"step": "1",
"step_status": "completed",
"calls_pulled": <count>
}
Volume sanity checks
< 50 calls: ask the user whether to extend the window to 180 days. Document the decision.50–200: proceed, flag in the final grade that confidence is medium.200+: proceed at full confidence.
When the API rate-limits or paginates
Most call intel tools paginate at 50–100 calls per request. Page through fully — do not stop at the first page. If you hit a rate limit, save state and tell the user to resume in 5 minutes with /refine-icp-resume.
Privacy note
If the user is in a regulated industry, their calls may contain PII. Don't ship raw transcripts to any third-party API beyond the LLM call inside Claude Code itself. Strip account names and emails from anything you write to disk if the user requests it (add a redact: true flag to stack.json constraints).
