# Research > Probe project and technical questions. Findings are positions to be challenged with explicit sources, not conclusions. Phase 6 advisor pass argues against the synthesis before it lands. Do NOT use for design planning or plan generation (use /think instead). - Skill: `thkt/research` (Agent Skill, multi-file: 9 files) - Install (CLI): `npx skillmds@latest add thkt/research` - Raw SKILL.md: https://api.skillmd.com/api/skills/thkt/research/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Research & Search - Author: thkt (https://skillmd.com/u/thkt) - Updated: 2026-09-10 - Page: https://skillmd.com/skills/thkt/research --- # /research - Project & Technical Investigation Investigate the codebase and record findings with sources, without implementation. ## Input The research subject is taken from `$ARGUMENTS`, a free-text topic or question. If empty, prompt via AskUserQuestion. ## Source notation Facts are `file:line` or command output, inferences `inferred from X`, unverified `unknown, requires X`. This is the source notation the Phases and the report template refer to; no other form is accepted. ## Phase 1: Outcome Anchor Read `.claude/OUTCOME.md`. If absent, generate the stub via /outcome. If the investigation steps into Non-goals, confirm with the user before proceeding. ## Phase 2: Prior Research Scan Derive the lowercase hyphenated subject slug from `$ARGUMENTS` and run `${CLAUDE_SKILL_DIR}/scripts/find-prior-research.ts .claude/workspace/research`. Parse the JSON `{ candidates: [{file, shared}, ...], slug_words: int }` (shared descending) from stdout. - No candidates: set the report's Prior research to `none found` and move on - A candidate with shared >= 2, or with shared equal to `slug_words`: carry forward per the table below - A candidate with shared == 1 left after that: the slug runs two words or more, so the filename overlap is all it rests on. It is excluded from the carry-over table below and lands only in the report's References, with its path and shared count | Extract | Carry to | Handling | | ----------------------- | -------- | ----------------------------------------------------------------------------------------------- | | Key Findings table | Phase 7 | Re-verify or supersede as baseline | | Constraints table | Phase 4 | Use as input only while its Domain matches the current one, otherwise return it to re-discovery | | Disconfirmation results | Phase 7 | Reference | ## Phase 3: Intent and Domain Clarification Skip if `$ARGUMENTS` clearly indicates both. Otherwise ask via AskUserQuestion. Intent is chosen from Feature planning / Bug investigation / Understanding; Domain from Data model / API / Infrastructure / General. ## Phase 4: Domain-Scoped Parallel Investigation Two inputs arrive here. The intent and domain Phase 3 settled pick the row of the domain table below, and the Constraints table Phase 2's handoff table passes is taken as input only when the handoff's Domain matches the current one. Launch Explore / ugrep / bfs / Read in parallel. Append each command and its raw output verbatim to the scratch. This is the audit trail; Phase 7 Disconfirmation quotes it directly and does not reconstruct. State the source for each finding in place. For Feature planning or Bug investigation intent, also invoke `Agent(subagent_type: explorer-feature)`. The spawn runs in the background, so keep the other searches going while its completion notification is pending. Take the result as a single JSON object `{ findings: [{ statement: string, source: string }] }`, and do not move to the next Phase before it arrives. When that trigger fires, or when a `.codegraph/` index exists, read ${CLAUDE_SKILL_DIR}/references/tactics.md and apply the tactics whose trigger matches. With a Domain other than General, read ${CLAUDE_SKILL_DIR}/references/domain-scope.md and scope the searches by its roots and words. Once the findings are in, read ${CLAUDE_SKILL_DIR}/references/verification.md and apply the verification matching the finding's kind. This read happens on every run, whatever the intent and the domain. ## Phase 5: Strong Inference (Bug investigation only) Apply ${CLAUDE_SKILL_DIR}/../../rules/core/OPERATION.md § Debug Investigation Protocol to eliminate the bug, then once the root cause is confirmed, run ${CLAUDE_SKILL_DIR}/references/verification.md § Same-origin sweep. ## Phase 6: Advisor Pre-Synthesis Check Invoke `advisor()` with no parameters. Advisor sees the full conversation history. If it flags a missed area or weak inference, return to Phase 4 to narrow the scoping. Skip the invocation only when all conditions hold, and record the skip reason in the output. - Phase 2 hit prior research, and no Phase 4 finding falls outside its Key Findings - Intent is Understanding and Domain is General - No claim crosses a repository boundary or drives PR scope ## Phase 7: Synthesis 1. If Phase 2 found prior research, integrate the inherited findings / constraints into Key Findings, marking each re-verified or superseded 2. Confirm each finding carries a source in the source notation. Mark gaps `unknown, requires X` 3. Triage each finding. A Next Action goes only to a finding tied to a direct answer to the `$ARGUMENTS` question, to advancing or protecting an OUTCOME.md Behavior / Constraint, or to handling a real incident (issue / bug report). Such a finding states its linkage in the action cell. Every other finding gets Next Action `record only`. Either way, all findings stay listed 4. Record Disconfirmation. If Phase 5 ran, write `Covered by Phase 5 elimination`; if skipped, quote the command and raw output from the scratch verbatim. Treat 0 hits as possible tool misuse before absence 5. Confirm every Phase 3 question is answered or recorded as `unknown, requires X` ## Phase 8: Output Generate the report following the skeleton in ${CLAUDE_SKILL_DIR}/templates/research.md, fill in `${CLAUDE_SESSION_ID}`, and save to `.claude/workspace/research/YYYY-MM-DD-.md`. After saving, propose the destination from the template's Next Steps table in the conversation, and run none of them. Attach the saved report's path and the words the slug came from, verbatim. `/think` builds its slug from those words and pulls this report with the same script, so a word that drifts leaves the report unreachable. ## Completion Criteria Not done until all are satisfied. An item whose Condition carries "(...)" is required only when applicable. | Item | Phase | Condition | | ----------------- | ------- | ---------------------------------------------------------------------------------------------------- | | OUTCOME | Phase 1 | `.claude/OUTCOME.md` present | | Prior research | Phase 2 | `Prior research` field filled with the slug or `none found` | | Audit trail | Phase 4 | Scratch captured with commands and raw output verbatim | | Cross-method | Phase 4 | Cross-method verification performed for exhaustiveness claims (when such a claim exists) | | Primary source | Phase 4 | Primary-source verification run on load-bearing external claims, or marked unverified (when present) | | Same-origin sweep | Phase 5 | Sweep performed when Bug intent confirmed a root cause (when applicable) | | advisor | Phase 6 | advisor invoked, or skip reason recorded | | Source | Phase 7 | Every finding has an explicit source or an `unknown, requires X` note | | Triage | Phase 7 | Every Next Action states its linkage (question / OUTCOME / incident) or reads `record only` | | Save | Phase 8 | Output saved to `.claude/workspace/research/` | | Handoff | Phase 8 | Destination proposed from the Next Steps table, with the report path and the slug's words attached |