Investigates open questions before planning by running parallel research across docs, codebase patterns, and external references, then updates initiative artifacts with evidence-backed findings.
Resolve investigatable unknowns before implementation planning.
Use this skill when answers are discoverable through research. For unknowns that must be built and experienced, use he-spike instead.
he-spec when initiative direction still has open questionshe-plan when requirements or constraints are uncleardocs/specs/<slug>-spec.md with revision notes.he-plancalled_from matches this skill (bash scripts/runbooks/select-runbooks.sh --skill he-research), but never waive/override anything codified here.docs/specs/<slug>-spec.md (preferred) or direct question list from user.docs/plans/completed/, docs/spikes/, docs/generated/.bash scripts/runbooks/select-runbooks.sh --skill he-research and read any returned runbooks. Apply their additions throughout — they must not waive or override gates codified here.Classify each question before investigation:
Category outcomes:
he-spikehigh|medium|low)docs/specs/<slug>-spec.md:
Requirements, Risks, Constraints, or Boundaries as needed.Revision Notes describing what changed and why.When a spec exists, keep it as source of truth:
## Constraints## Requirements## Risks and ## Scope boundaries when findings change direction## Open Questions (Optional) with next actionDo not add implementation-level details that belong in he-plan.
docs/specs/<slug>-spec.md (when spec exists), ordocs/specs/<slug>-spec.md for new initiative intakeRevision Noteshe-spike.| Anti-Pattern | Better Approach |
|---|---|
| Guessing when evidence is thin | Mark as unresolved with confidence level |
| Researching questions that need hands-on exploration | Redirect to he-spike for experience-dependent unknowns |
| Adding implementation details to the spec | Keep implementation in he-plan; spec is intent only |
| Treating inference as fact | Always distinguish and label confidence explicitly |
| Researching in the main thread instead of parallel | Launch one subagent per question for concurrent investigation |
Always use interactive question tool at transitions (AskUserQuestion in Claude Code, request_user_input in Codex Plan mode, or equivalent). Offer:
he-plan (recommended when sufficient clarity)he-spike for experience-dependent unknownsIf running autonomously or no interactive tool is available, continue with the recommended next phase and log an Autonomous transition note in the relevant artifact's Revision Notes.