staruhub/ClaudeSkills / skills/Geek-skills-deep-research
deep-research
VerifiedUse this skill when the user wants an evidence-based research memo, literature review, market/policy/technical landscape, or a multi-source decision brief with citations, trade-offs, and a clear conclusion. Best for tasks that need synthesis across multiple external sources, iterative follow-up research, or a reusable written artifact. Do not use for quick factual lookups, single-source summaries, simple Q&A, summarizing one document the user already provided, plan-only requests where the user explicitly defers the actual research, or when the user wants a short answer with no need for cited evidence (a short but evidence-backed conclusion is still a brief memo, which this skill covers). Chinese trigger examples: "帮我调研", "深度研究", "综述报告", "技术选型分析", "竞品研究", "政策分析".
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 11.0 KB |
| assets/report_template.md | markdown | 1006 B |
| evals/routing-evals.json | config | 3.8 KB |
| evals/runbook.md | markdown | 967 B |
| references/evaluator-prompt.md | markdown | 4.0 KB |
| references/handoff-format.md | markdown | 3.9 KB |
| references/landscape-scan.md | markdown | 1.2 KB |
| references/methodology.md | markdown | 5.7 KB |
| references/observability.md | markdown | 3.0 KB |
| references/quality-gates.md | markdown | 5.5 KB |
| references/report-assembly.md | markdown | 3.4 KB |
| references/research-notes-format.md | markdown | 2.2 KB |
| references/subagent-prompt.md | markdown | 1.7 KB |
| references/tension-discovery.md | markdown | 1.2 KB |
| scripts/emit_run_summary.py | script | 5.0 KB |
| scripts/source_evaluator.py | script | 11.9 KB |
| scripts/verify_citations.py | script | 10.7 KB |
references/methodology.md
View on GitHub# Deep Research Methodology V8
V8 is built around one idea: **use the lightest workflow that still produces a trustworthy evidence artifact**.
## 1) First decide whether deep research is justified
Deep research is justified when at least one of these is true:
- the user wants a written artifact they can reuse or share
- the task needs synthesis across multiple sources
- the task is decision-oriented and trade-offs matter
- the user wants a second round / deeper follow-up
- the topic is broad, fast-moving, contested, or cross-domain
If none of those are true, do something simpler.
---
## 2) Choose output type before you search
| Output type | Use when | Default target |
|---|---|---|
| Brief memo | concise but evidence-backed answer | 800-1800 words |
| Full report | comprehensive research deliverable | 2500-6000 words |
| Delta update | continue prior work or summarize changes | 600-1800 words |
Choose the artifact first, because it determines how much context, search, and evaluation work is justified.
---
## 3) Single-agent first, fan out only when earned
### Use single-agent when
- there are 1-2 main questions
- evidence threads overlap heavily
- the task is short or medium complexity
- the extra coordination cost would outweigh the gain
### Use subagents when
- there are 3+ separable threads
- different roles or domains truly matter
- parallel work will reduce time or improve coverage
- you have a writable workspace and a place to merge artifacts
### Subagent rules
- prefer 2-4 subagents, rarely more than 5
- each subagent owns one crisp thread
- avoid two agents answering the same sub-question
- merge by notes, not by copying raw search output around
---
## 4) Build a context plan, not just a task plan
Before investigating, define three buckets.
### A. Essential now
Load these immediately:
- the user’s exact request
- output type and audience
- freshness / geography / scope constraints
- core methodology and report assembly rules
### B. Useful later
Defer until needed:
- detailed evaluator prompt
- subagent prompt
- tension discovery rules
- landscape scan playbook
- observability guide
### C. Keep out unless needed
Avoid loading by default:
- old run logs
- every prior report
- entire literature dumps
- every reference file in the skill tree
This is the difference between **context organization** and “shove everything into the prompt”.
---
## 5) Investigation loop
For each research thread:
1. Start with 2-4 broad searches.
2. Identify named entities, standards, datasets, papers, products, policies, laws, or experts.
3. Chase the highest-value entities with targeted searches.
4. Read/fetch the best sources for the claims that actually matter.
5. Write notes with facts, analysis, gaps, conflicts, and support snippets.
6. Stop when the next search is mostly duplicate / low-value, or when confidence is already good enough for the chosen artifact.
### Stop because you know enough, not because you are tired.
A thread is usually “good enough” when:
- the core question is answered from multiple relevant sources
- the main counterpoint or limitation is also covered
- additional searches mostly repeat known facts
---
## 6) Optional modules: use them selectively
### Tension discovery
Use when:
- the topic is hype-heavy
- the task is a recommendation or selection
- the popular framing might hide the real trade-off
Skip when:
- the user only needs a descriptive overview
- the topic is stable and not especially contested
### Landscape scan
Use when:
- the domain is unfamiliar or broad
- literature, standards, repos, filings, or ecosystem mapping matters
- you need to derive roles or clusters from the evidence landscape
Skip when:
- the task is narrow and already well scoped
- the domain can be covered directly without a separate scan
### Reverse search
Run when:
- the draft has benefits but almost no costs
- the notes have no failure modes, constraints, or counter-evidence
- the recommendation feels too clean
---
## 7) Evidence hierarchy is claim-specific
Do not blindly prefer one source class everywhere.
Ask: **what kind of evidence can actually support this claim?**
Examples:
- product feature claim → official docs / release notes / source code
- law or policy claim → official text / regulator / court or agency guidance
- scientific effectiveness claim → peer-reviewed studies / meta-analyses / trial reports
- market trend claim → filings / reputable research / financial disclosures / high-quality industry data
- operational failure mode claim → incident reports / postmortems / engineering docs / credible case studies
A source can be famous and still be the wrong support for the claim.
---
## 8) Merge by notes, re-open raw sources only when needed
By default, the lead agent should synthesize from notes and the registry.
That keeps the active context smaller and more stable.
Re-open raw/fetched sources for:
- high-stakes claims
- unresolved conflicts
- suspicious or overly compressed notes
- final decision-driving numbers or legal wording
This is the sweet spot: **notes-first, raw-source-on-demand**.
---
## 9) Final output rules
Every final artifact should:
- answer the user’s actual question
- distinguish evidence from interpretation
- include limitations / trade-offs
- calibrate certainty honestly
- avoid fake precision
Only force a decision framework when the user is choosing.
Only force a contrarian move when a real dominant narrative exists.
Otherwise prefer a **non-obvious insight**.
---
## 10) Learn from runs
After the run, record:
- which workflow shape was used
- where time or tools were wasted
- where evidence was thin
- what made the final result better
- what should become a new test case
A skill gets better when failures become future routing tests or output checks.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |