Deep Research

Deep Research V8.1

This skill is for evidence-rich research outputs, not for every question that happens to mention “analysis”.

The V8 shift is simple:

• Single-agent first. Start with one lead agent and only fan out when parallel work will clearly help.
• Thin harness, fat skill. Put reusable judgment and workflow here; keep deterministic checks in scripts.
• Context organization over prompt stuffing. Load the minimum active context bundle, then pull in references only when needed.
• Eval and observability built in. A good report is not enough; the run must also be diagnosable and improvable.

What this skill should produce

Choose the lightest artifact that satisfies the task.

If the user did not ask for a long report, default to Brief memo.

When NOT to use this skill

Do not activate for:

• quick fact lookups or simple definitions
• summarizing a single provided article/PDF/page
• short comparisons the model can answer directly from 1-2 sources
• brainstorming without evidence requirements
• tasks where the user explicitly wants a short answer, not a report

If in doubt, ask yourself: Does this task need a reusable evidence artifact and multi-source synthesis? If not, do something simpler.

Org-policy boundary

This skill does not replace system policies, enterprise guardrails, or repo-level instructions. Put these outside the skill:

• data handling / PII / compliance rules
• approval requirements for external access or irreversible actions
• org-wide style and review policy
• environment-specific permissions

Keep those in system prompts, AGENTS/CLAUDE/OpenAI config, or the harness. This skill owns the workflow, not the company’s permanent red lines.

Active context bundle

At activation time, keep the active bundle small.

Always load first

1. This SKILL.md
2. references/methodology.md
3. references/report-assembly.md
4. references/research-notes-format.md

Load on demand

• references/subagent-prompt.md only if you actually dispatch subagents
• references/handoff-format.md only when a delta update continues a prior round
• references/evaluator-prompt.md only if you run the evaluator
• references/quality-gates.md before finalization
• references/observability.md when emitting metrics or diagnosing regressions
• references/tension-discovery.md only for contested / decision-heavy topics
• references/landscape-scan.md only when literature or ecosystem mapping matters

After compaction or context reset Reload only:

• research-plan.md
• active task notes
• registry.md
• unresolved issues list
• the one reference file for the current phase

Do not reload the whole skill tree unless the run drifted badly.

Workflow

P0 — Scope, route, and choose the lightest mode

Create workspace/research-plan.md with:

• research question
• intended audience
• freshness requirement
• geography / market / jurisdiction
• output type (brief / full / delta)
• stakes: low / medium / high
• why this skill is justified

Then choose the orchestration mode:

Do not fan out just because subagents exist.

P0.5 — Optional modules (not mandatory by default)

Use optional modules only when they earn their keep:

• Tension discovery (references/tension-discovery.md): use for contested, hype-heavy, or decision topics where mainstream framing may be wrong.
• Landscape scan (references/landscape-scan.md): use when the domain is unfamiliar, broad, or literature-heavy. For non-academic topics, this can be an ecosystem/standards/vendor scan rather than arXiv.
• Reverse search: use when costs, failure modes, counter-evidence, or operational constraints are missing.

P1 — Plan the evidence work

Break the task into 1-5 research threads. Each thread needs:

• one crisp objective
• starting queries
• what “done” looks like
• what evidence would change the conclusion

If using subagents, each subagent gets one focused thread. Avoid overlapping ownership.

P2 — Investigate, extract, and write notes

Follow references/research-notes-format.md.

Rules:

• search broadly first, then chase named entities, standards, datasets, products, trials, laws, or papers
• fetch and read the best supporting sources for the highest-value claims
• write notes that separate facts, analysis, gaps, and unresolved conflicts
• capture support snippets/paraphrases for the top claims so later verification is easier

The lead agent should work from notes by default, but may inspect raw/fetched sources again when:

• two sources materially conflict
• a claim is high-stakes or decision-critical
• a note looks suspiciously weak or over-compressed

P3 — Build registry and verify evidence

Create workspace/registry.md from approved sources only.

Use scripts/source_evaluator.py as a helper, not an oracle. Authority scores are heuristics. Final acceptance depends on claim fit, evidence type, and whether the source can actually bear the weight of the claim.

Use scripts/verify_citations.py before finalization.

Evidence rules:

• core claims should lean on the strongest available evidence for that claim type
• anecdotes illustrate; they do not anchor the conclusion
• conflicting evidence must be surfaced, not silently averaged away
• if the topic is high-stakes, spot-check raw support for top claims before shipping