SuperAPI AI Directory

gtm-lead-research

Writing & research

View repository

Updated 10-15 21:37

About this skill

Research Leads

Recipe

  1. Resolve the GTM context repo and target organization node.
  2. State Using GTM context: <display name> — <N> ICPs and <M> personas visible.
  3. Report a missing prerequisite and stop when that node has no personas.
  4. Inspect the org chain, target-node ICPs and personas, safe packets, and web evidence when available.
  5. Preserve supplied persona labels without re-segmenting.
  6. Separate inspected findings, unverified claims, tentative hypotheses, conflicts, and provenance without silent joins.
  7. Interpret persona relevance, timing, risks, and outreach angles.
  8. Assign one high, medium, or research-needed priority.
  9. Render the lean person-brief spine and bulk priority distribution when applicable.
  10. Answer save requests with the org-only principle and complete copyable brief.
  11. Close No files, git history, or external systems changed.

Details

  • Treat the supplied persona label as immutable input; flag one absent from the visible set instead of assigning or substituting another.
  • Inspect only the selected node's own ICPs and personas; use the root-to-target org.md chain only for organization context.
  • Use web access when available and relevant; state its absence instead of pretending evidence was inspected.
  • Never open, reproduce, or describe an unsafe or tokenized link; list it only as Private source withheld.
  • Put actually inspected role facts under Findings and non-inspected statements under Unverified claims.
  • Keep role, influence, pain, and stakeholder interpretations explicitly tentative hypotheses.
  • Preserve source, publisher, and date beside material findings; never silently join separate sources or people.
  • Keep conflicts visible, including dated role conflicts, rather than selecting the convenient claim.
  • Use high only for inspected persona relevance plus a clear active or dated signal without material conflict.
  • Use medium when useful relevance or timing evidence has a material conflict or unverified dependency.
  • Use research-needed when inspectable evidence is absent or the persona label is not visible.
  • Set Confidence and Needs review from evidence completeness and conflict.
  • Render every brief in the exact output-reference order; use None evidenced for an empty section.
  • Open bulk output with Research-priority distribution, then render one independent complete brief per person.
  • A lead is the research subject, never the user, operator, or active identity.
  • Research is response-only: never write, commit, update a CRM, or change an external system.
  • A save request is not an approval flow; explain the org-only storage boundary and provide copyable Markdown.

Calls

  • Read references/context.md before research to resolve repo and node, enforce node-local visibility, state counts, and handle ambiguity or zero personas. If unavailable, report that the contract could not load and stop.
  • Read references/evidence.md before inspecting sources to apply link safety, evidence boundaries, provenance, joining, and calibration. If unavailable, use safe supplied packets only, mark claims unverified, and use research-needed.
  • Read references/output.md before rendering to apply one-off, bulk, save-request, and closing shapes. If unavailable, render the nine Recipe spine sections in order and use the exact close.