Deep Research

Deep research workflow for /deepresearch with sources, claims, synthesis, and resumable state.

Use this for /deepresearch <question> and for any request that needs more than a quick sourced answer: literature mapping, market/technical due diligence, state-of-the-art reports, factual investigations, or research that should survive across several agent turns.

Core rule

Do not answer from memory. Build an evidence ledger first, then synthesize. Prefer local/project sources when the question is about the user's workspace; otherwise use web and academic sources.

When to delegate

Use existing skills as lanes when relevant:

• 00-web-grounded-answers for factual web grounding and citations.
• academic-research-hub for PubMed, arXiv, Semantic Scholar, Google Scholar style work.
• literature-search for systematic-review methodology, PRISMA-like screening, inclusion/exclusion criteria.
• researchclaw only for full autonomous research pipelines with experiments/paper generation.
• OMOC/team when the research needs parallel lanes: scout, skeptic, verifier, synthesizer.

Default workflow

1. Create or resume a ledger under .deepresearch/<slug>/ using scripts/deepresearch.py init.
2. Restate the research question, expected output, known constraints, and stop condition.
3. Split the question into 3-7 research lanes. Typical lanes: background, strongest evidence, contrary evidence, current/latest status, implementation details, risks, open questions.
4. Search each lane. For unstable or recent facts, browse. For scientific/medical claims, prioritize papers/reviews and authoritative databases.
5. Record every useful source with scripts/deepresearch.py source add.
6. Record atomic claims with scripts/deepresearch.py claim add, linking each claim to at least one source id when possible.
7. Mark claims as supported, weak, conflicting, or unverified after cross-checking.
8. Write notes and interim summaries frequently with scripts/deepresearch.py note add so a crash or context reset does not lose the state.
9. Run scripts/deepresearch.py brief before final synthesis to inspect coverage and gaps.
10. Produce the answer with: direct conclusion, evidence map, important caveats, sources, and next actions if useful.

Quality bar

• Important factual claims need citations.
• Separate evidence from inference.
• Include uncertainty when sources conflict or evidence is thin.
• Prefer primary sources for technical/scientific claims.
• For recommendations or decisions, include tradeoffs and what would change the conclusion.
• Keep raw source snippets short; summarize instead of copying.

Output shapes

For a normal answer:

• Short answer.
• What I found.
• Evidence and caveats.
• Sources.

For a report:

• Executive summary.
• Method.
• Findings by lane.
• Evidence table or bullet ledger.
• Gaps and confidence.
• Sources.

For a long running investigation:

• State path: .deepresearch/<slug>/
• Current decision.
• Completed lanes.
• Open lanes.
• Next command or next research pass.

Commands

Resolve scripts/deepresearch.py relative to this skill directory.

python3 scripts/deepresearch.py init --question "..." --slug optional-slug
python3 scripts/deepresearch.py lane add --slug optional-slug --name "contrary evidence" --question "What would falsify this?"
python3 scripts/deepresearch.py source add --slug optional-slug --title "..." --url "..." --kind paper --reliability high
python3 scripts/deepresearch.py claim add --slug optional-slug --text "..." --source S001 --status supported
python3 scripts/deepresearch.py note add --slug optional-slug --text "..."
python3 scripts/deepresearch.py brief --slug optional-slug

Integration with OMOC

When OMOC is active, /deepresearch can be one lane inside a goal/team/ralph loop. Use the deepresearch ledger as durable evidence, then have OMOC verifier tasks consume the ledger before checkpointing the goal.

Suggested composition:

1. /goal defines the decision or report to ship.
2. /deepresearch builds evidence under .deepresearch/<slug>/.
3. /team splits lanes across workers/verifiers.
4. /ralph repeats bounded cycles until the ledger has enough evidence or the goal is blocked.

Never let a deepresearch loop run forever without a stop condition. Use explicit coverage criteria: minimum source count, minimum verifier pass, unresolved contradictions, or deadline.