Ultimate Deep Research

A production-grade research skill for coding agents. It runs an 8-phase pipeline — scope, parallel retrieval, evidence registry, synthesis, counter-review, mechanical verification — and produces citation-backed Markdown reports instead of confident guesses.

Merged from the strongest open-source research skills (see references/lineage.md). Principle: breadth-first parallel retrieval, evidence on disk, synthesis only from notes, adversarial review, mechanical verification.

Why it is different

• Search, don't recall. Every factual claim traces to a retrieved source, not model memory.
• Evidence survives compaction. Sources and task notes live in research/<slug>_<date>/ on disk, so long runs never lose citations.
• Drafts come from notes only. No invented URLs, numbers, quotes, or DOIs; unsupported claims are cut or marked [unverified].
• Counter-review is mandatory. The red team must surface at least three real issues, or it runs again.
• Mechanical gate. scripts/validate_report.py checks citations, sections, placeholders, and uncited numbers before delivery.
• Fetched content is data, never instructions. Prompt injection inside pages is ignored.

Modes

Mode	Use for	Sources	Subagents	Output
Quick	single fact / sanity check	3–6	0	answer + 3 cites
Standard (default)	most questions	12+	3–5	1.5–4k words
Deep	decisions, comparisons, contested topics	20+	5–8	4–8k words
UltraDeep	due diligence, literature reviews	30+	8–12 + red team	8k+ words

Lenses combine with any depth: Recency (last 30 days + community signal), Enterprise (company deep dive + SWOT), Compare (X vs Y matrix), Discovery (trend scan), Structured-list (outline-first, human-in-the-loop), Academic (papers, DOIs, strength of evidence).

Pipeline

Phase	Output
P0 Setup	Real date, workdir, citation registry, capability check
P1 Scope	3–7 sub-questions, task board with non-overlapping scopes
P2 Retrieve	Parallel subagents, full-page reads for load-bearing sources, evidence objects
P3 Triangulate	Source registry, authority and independence gates
P4 Outline	Topic-first, evidence-mapped sections with gaps and counter-claims
P5 Draft	Prose-first report, per-section confidence, stated AS_OF
P6 Red team	Five adversarial personas, ≥3 real issues, loop back if critical
P7 Verify	Validator, spot-checks, URL re-fetch, ≤3 fix cycles
P8 Deliver	report.md on disk, brief chat summary with confidence and caveat

Install

# OpenCode
unzip ultimate-deep-research.skill -d ~/.config/opencode/skills/

# Claude Code
unzip ultimate-deep-research.skill -d ~/.claude/skills/

# Any agent that reads a skills directory
mkdir -p <skills-dir>/ultimate-deep-research
cp -r ultimate-deep-research/* <skills-dir>/ultimate-deep-research/

The repo root has the flat ultimate-deep-research-SKILL.md for platforms that take a single file; the .skill archive is the full skill with references, scripts, templates, and tests.

Requires Python 3 (standard library only) for the registry and validator. Web search and page fetch are required; subagents are optional — the skill degrades to sequential mode.

Usage

Once installed, ask normally. It triggers on "research", "investigate", "compare", "due diligence", "what's the current state of X", "last 30 days", and similar phrasing.

Typical run:

# 1. Create the workdir and register sources
python scripts/sources.py init --dir research/topic_2026-10-27 --mode standard --as-of 2026-10-27
python scripts/sources.py add --dir research/topic_2026-10-27 \
  --url https://example.com/report --title "Report" --type official \
  --access public --date 2026-10-20

# 2. Write the report, then validate it
python scripts/validate_report.py research/topic_2026-10-27/report.md \
  --sources research/topic_2026-10-27/sources.json --mode standard

sources.py commands: init, add, drop, list, stats, export (numbered bibliography text).

Files

Path	Contents
ultimate-deep-research-SKILL.md	Flat skill definition (root copy)
ultimate-deep-research.skill	Full skill archive (zip)
SKILL.md	Skill definition with pipeline and non-negotiables
references/methodology.md	Subagent briefs, query craft, evidence objects
references/source-governance.md	Accessibility/type labels, authority rubric, gates
references/quality-gates.md	Thresholds, red-team personas, completeness checklist
references/modes.md	Recency, Enterprise, Compare, Discovery, Structured-list, Academic
references/environments.md	Per-platform adaptations and continuation protocol
references/lineage.md	Upstream skills and what was taken from each
scripts/sources.py	Disk-backed citation registry
scripts/validate_report.py	Report validator
templates/report_template.md	Report skeleton with required sections
templates/task_note.md	Per-task evidence note
tests/test_skill.py	15 unit tests (python3 -m unittest tests.test_skill)

Verify the install

python3 -m unittest tests.test_skill -v   # 15 tests, ~20s
python3 scripts/validate_report.py --help

License

MIT © 2026 TonyD2F — see LICENSE.