An eight-phase deep research pipeline with an evidence ledger

It turns "deep research" into a verifiable pipeline

Claude Deep Research Skill (199-biotechnologies, 1,051 stars / 116 forks) is not "let the model search a few more times". It is a research pipeline with an evidence ledger: 8 phases, 4 intensity modes, every step required to be checkable.

Licence first: this repo has no LICENSE file (GitHub returns license = None). Under default copyright that means "all rights reserved" — you can read it and run it locally, but you must get the author's permission before redistributing or using it commercially. We flag it because it is the one thing you must know before adopting it.

Four modes and the phase matrix

You pick a mode from a decision tree; SKILL.md explicitly says do NOT use it for simple lookups, debugging, or questions answerable in one or two searches.

Mode	Phases	Time	Use
Quick	3	2–5 min	Initial exploration
Standard (default)	6	5–10 min	Standard research
Deep	8	10–20 min	Critical decisions
UltraDeep	8+	20–45 min	Comprehensive review

The 8 phases are SCOPE → PLAN → RETRIEVE → TRIANGULATE → OUTLINE REFINEMENT (4.5) → SYNTHESIZE → CRITIQUE → REFINE → PACKAGE. Quick runs only SCOPE/RETRIEVE/PACKAGE; CRITIQUE and REFINE arrive at Deep and above.

The critical note: phases 3–5 operate as an evidence loop per section, not as strict sequential gates — retrieve → evidence store → refine outline → draft → verify claims → delta-retrieve if needed. That prevents the mismatch you get from "search everything once, then write".

The evidence ledger is the soul of it

Besides the report, the output directory ~/Documents/[Topic]_Research_[YYYYMMDD]/ carries four JSONL ledgers, all append-only with canonical IDs:

• sources.jsonl — a stable source registry;
• evidence.jsonl — an evidence store with quotes and locators;
• claims.jsonl — an atomic claim ledger with support status;
• run_manifest.json — this run's query, mode, assumptions, provider config.

The quality bar is stated hard: 10+ sources, 3+ independent sources per major claim (explicitly cluster-independent, not just a count); every factual claim cited inline as [N] and backed in evidence.jsonl; claim-support verification is mandatory — no unsupported factual claim passes delivery; no placeholders, no fabricated citations; prose-first (≥80%), bullets sparingly.

Output contract and self-check scripts

The report structure is fixed: Executive Summary (200–400 words), Introduction (scope, methodology, assumptions), Main Analysis (4–8 findings, 600–2,000 words each, all cited), Synthesis & Insights, Limitations & Caveats, Recommendations, a COMPLETE bibliography (every citation, no placeholders), and a Methodology Appendix. It emits Markdown (the source of truth), a McKinsey-style HTML report (auto-opened), and a PDF (WeasyPrint, auto-opened).

python scripts/validate_report.py --report [path]   # 9 structural checks
python scripts/verify_citations.py --report [path]  # citation verification, max 3 loops
python scripts/md_to_html.py [markdown_path]        # render HTML

Long reports (>18K words) have a separate continuation.md. The autonomy principle: infer assumptions and stop only on critical errors — the same philosophy as Superpowers' "Rulings, not stalls".

Install

git clone https://github.com/199-biotechnologies/claude-deep-research-skill \
  ~/.claude/skills/deep-research
# Optional: wire a search-cli (Brave / Serper / Exa / Jina / Firecrawl) for better retrieval

Runtime is Claude Code. Licence reminder: fine for personal use; secure permission before embedding it in a commercial product.