| name | websets-cohort-bios |
| description | Produce uniform, verified background reviews / biographies / profiles for a SHORTLIST of entities (people or companies) by dispatching one research subagent per entity in parallel, then synthesizing and persisting the results. Use when the user asks for "bios", "background reviews", "profiles", "tell me about each of these", or a comparison across several candidates — especially after a [[websets-secondary-review]] produced a shortlist. For a single entity use [[deep-research-item]] instead; this is the parallel, many-entities harness.
|
| argument-hint | ["list of entities OR websetId + which items"] |
| user-invocable | true |
| allowed-tools | Task, Read, mcp__websets-codemode-local__execute, mcp__websets-codemode-local__search, mcp__exa__web_search_exa |
Websets Cohort Bios
Fan out parallel research subagents to build comparable background reviews across a shortlist,
then synthesize and persist. The value is uniformity (every bio answers the same questions) and
parallelism (one agent per entity, dispatched together).
Observe — assemble the shortlist
- Take entities from the conversation, or pull from a Webset (
items.getAll) and the survivor set
of a [[websets-secondary-review]]. For each, gather the starting facts you already have
(name, practice/company, location, contact, any prior verification) so agents enrich rather than
rediscover.
Act — dispatch one subagent per entity, in parallel
Send all Task calls in a single message so they run concurrently (general-purpose, or a
research agent like scout). Give every agent the same strict contract:
- Inputs: the subject + known starting facts ("verify, don't just repeat").
- Verify these dimensions, each with a source URL: education & training; board
certification / credentials; years & career history; clinical/scope focus and fit for the stated
use-case; affiliations; languages; accepted insurance / plans; accepting-new-clients status;
reputation signals (ratings + review themes); and public red flags (licensing-board
disciplinary actions, sanctions, malpractice) — report "none found" explicitly.
- Rules: no fabrication; mark uncertain items "unverified" with the reason; disambiguate
from same-name individuals; resolve and call out conflicting data (stale directories, NPI
mislabels, aggregator errors).
- Output contract (identical for all): ~250-word prose review + a "Quick facts" bullet block +
a "Sources" URL list.
Orient — synthesize
Collect the returns and produce: per-entity bios in a consistent shape, a comparison table, and a
revised ranking with tiers (e.g. top-pick / solid / caution). Surface cross-cutting caveats
once (e.g. "directory accepting/insurance flags go stale — confirm by phone"; "board portals
couldn't be queried directly — self-verify at the licensing board"). Honestly flag where your
synthesized ranking differs from anything previously stated, and name which record is canonical.
Persist (optional, on request)
Write results back so they survive: for each item store.syncItem(...) then
store.annotate({ itemId, type: 'background_review' | 'judgment' | 'contact_verified', value, source }).
Read back with store.query to confirm counts. (If store.* errors on the SQLite binding, see
[[pnpm10-blocks-native-builds]].) A durable local markdown export is a fine fallback if the store
is unavailable.
Notes
- Scale agent count to the shortlist; for >8 consider batching.
- Keep the per-agent prompt asking for sources — uncited claims are the main failure mode.
