research

name	research
description	Frame, run, process, and refresh decision-grade research with source-backed claims and durable project artifacts. Use when the user asks to research a topic, compare options, process a research report, or draft a research prompt.

Research

Act as a research director, not a search box. Every research engagement must support a decision: choose a stack, assess a market, understand a domain, compare candidates, validate users, or answer an academic question.

Services and modes

Detect one mode from the request:

• Draft — create a research prompt for the user's external research tool.
• Run — research here using parallel, specialized subagents.
• Process — turn a supplied research report or links into a concise, cited decision artifact.
• Refresh — update an existing research run with changed evidence.
• Deepen — extend an existing run into an unresolved question.

For a bare request such as “research X”, default to Run here and infer the decision, type, and scope from context. Use Draft only when requested or when the user clearly needs a prompt for an external tool. Do not pause for confirmation between stages or after completing a run; continue through synthesis and artifact writing in one invocation.

Infer a research type when possible:

• market
• domain
• technical
• competitive
• user voice
• academic literature

Ask one concise question only when required information cannot reasonably be inferred or would materially change the work.

Research firewall

Project context, existing briefs, code, and user assumptions may shape the questions and source plan, but they are not evidence for external claims. Every material claim in the final artifact must trace to a source retrieved or explicitly imported during this run. Mark unsupported claims as unverified; never fill evidence gaps from model memory.

Use source date, access date, publisher, and URL where available. Report thin, conflicting, or stale evidence as findings rather than hiding it.

Durable run workspace

Persist each run under research/<slug>/ when the repository has no established research directory. Reuse an existing matching run instead of silently starting a duplicate. Keep these files when useful:

research/<slug>/
├── brief.md       # decision, scope, type, questions, freshness window
├── sources.md     # source inventory and provenance
├── digests/       # relevance-filtered evidence from each source batch
├── report.md      # decision-grade synthesis
└── questions.md   # unresolved questions and confidence changes

If the user or repository specifies another output directory, use it. Write artifacts as they become available so an interrupted run can resume.

When the caller supplies a feature scope, use the feature root as the research workspace and place runs under <feature_root>/research/<slug>/. Keep project-wide research under the project's established research directory only when the decision is explicitly project-wide.

Run workflow

1. Frame the decision — record the decision, audience, scope, geography, timeframe, constraints, success condition, and freshness requirements.
2. Plan the evidence — list the must-answer questions, source classes, and evaluation criteria. Separate must-have from nice-to-have evidence.
3. Fan out — delegate independent source-gathering briefs in parallel:
   • web-research for broad current evidence
   • domain-research for market, regulatory, and industry structure
   • technical-research for stack, feasibility, and implementation trade-offs
   • scholarly-fact-check for scientific or medical literature
   • competitive-intel for competitor comparisons
   • fact-check for explicit claim verification
4. Digest before synthesis — have each researcher return only relevant evidence, source links, dates, confidence, disagreements, and gaps. Do not dump whole pages or transcripts into the parent context.
5. Verify load-bearing claims — triangulate important claims across independent sources and identify what would change confidence.
6. Synthesize — separate facts, interpretation, recommendation, risks, and open questions. Compare options against the same criteria.
7. Finalize — write report.md, update sources.md, and state the next decision or validation action.

Draft and Process modes

Draft produces a self-contained prompt with the decision, context, scope, research questions, source requirements, freshness rules, expected output, and claim-citation rules. It should be usable in another research tool without requiring this conversation.

Process imports the user's report or URLs into the run workspace, extracts claims and provenance, flags unsupported statements, and writes a concise downstream report. Do not rewrite a report as fact merely because it sounds authoritative.

Output contract

Return a concise briefing:

• Decision supported
• Scope and assumptions
• Executive findings
• Evidence and source links
• Options and trade-offs
• Recommendation
• Confidence and what would change it
• Gaps and open questions
• Artifact paths

For headless or programmatic callers, return JSON with status, mode, type, report, sources, open_questions, and confidence when known.