Plan, retrieve, cross-check and write in Claude Code
Deep Research
Trace findings back to original evidence
Organize research that needs several sources: define the scope, preserve sources and evidence, and produce a cited report with counterevidence and limitations that can be revisited.
Host documented by the author
- Claude Code
- GitHub stars
- 1,202
- Checked
- Categories
- Writing & ResearchContent & Knowledge
Decide whether you need a map or a comparison
Use Quick to map available evidence and open questions. For a choice between two options, use Standard with shared conditions. After delivery, trace a conclusion that could change the choice through its claim, excerpt and original passage, then decide what to investigate or test next.
Pick a task and get started
Install into Claude Code’s skills directory and confirm that deep-research can be read. Then choose a task below and replace its scenario and paths with your material. A first run can stop at Markdown and evidence records.
Supply the intended use, audience, time window and exclusions. Quick covers scope, retrieval and packaging to produce an initial map.
Install the deep-research entry
Clone into the author’s Claude Code path, then select the commit reviewed here. The destination should not already exist; inspect its version and local changes if it does. Explicitly invoke deep-research in Claude Code and confirm search and file tools are available.
Read the author’s setup guidegit clone https://github.com/199-biotechnologies/claude-deep-research-skill.git ~/.claude/skills/deep-research git -C ~/.claude/skills/deep-research checkout f2f2c0fa4e7617ca84c86b63f4bb40f77a746933After installation, confirm retrieval and persistence
Run the git clone command above in a terminal, then return to Claude Code. Ask it to confirm that it can read ~/.claude/skills/deep-research/SKILL.md and identify the available search, file and command tools. The optional search setup is brew tap 199-biotechnologies/tap && brew install search-cli, followed by search config set keys.brave YOUR_KEY for at least one provider. Without search-cli, an available WebSearch can serve as the fallback. Choose a new directory such as ~/Documents/[Topic]_Research_[YYYYMMDD]/ and record the actual provider, report path and material assumptions in run_manifest.json. Relative script paths are based on the Skill’s root directory.
Read installation and search setupGive your AI this first task
These are editorial task prompts. Their expected outputs were not generated in a site test. Only material actually retrieved and inspected should be recorded as evidence from the run.
Use deep research in quick mode to help two web engineers decide whether browser-side local speech recognition is worth prototyping. Use the execution date for recency; focus on publicly accessible official documentation, project code and limitations from the past year. Exclude cloud transcription services. Do not run models or present reported comparisons as tests performed in this run. State the scope and material assumptions, then identify the available search tools. Record queries actually executed, access dates and sources seen only as abstracts or inaccessible pages. Separate documented findings, contested points and questions that require a prototype. Do not pad the bibliography to reach a source count. Create a dated topic directory under ~/Documents/ and first deliver Markdown, sources.jsonl, evidence.jsonl, claims.jsonl and run_manifest.json. Use the template’s section headings. Suggest the next three questions worth investigating. Check each factual statement in the summary against a saved excerpt and its location in the original source, leaving gaps explicit.
Check what you receive
- An initial report with the search date, scope and assumptions, separating documented findings from prototype work still needed.
- A source registry, excerpts with source locations, claim ledger and run manifest that identify the search tools actually used.
- Spot-check three factual statements in the summary against real pages and passages; abstract-only access should be marked.
More about this skill
Deep Research is a research workflow for Claude Code. It suits technology choices, field surveys and contested claims: turn a broad question into a scope and search plan, inspect different sources, account for conflicting evidence and give conditional conclusions. A single-fact lookup or ordinary debugging rarely needs this process.
Quick provides an initial map. Standard, the default, adds planning, triangulation, outline refinement and synthesis. Deep adds critique and revision; UltraDeep extends the investigation through more iterations of the same stages. Standard and deeper modes can loop section by section through retrieval, evidence capture, drafting and targeted follow-up. Evidence that challenges the initial framing should change the report. The author’s timing and length figures are workflow targets, not delivery guarantees.
Its practical value is the material kept alongside the report: stable source identities, excerpts with source locations, claims linked to evidence, and a manifest of the question, mode and assumptions. Citation, structure and claim-checking scripts help with review, while retrieval, interpretation and file assembly still depend on the assistant. A polished long report is useful only if those records are authentic, complete and support its conclusions.
Sources and verification
Sources checked
Checked the entry, methodology, setup, scripts, schemas, templates and author tests at fixed commit f2f2c0f. We did not execute its research tasks, search services or export workflow, or validate the quality of generated reports.