Deep Research Skill | Agent Skills

Deep Research

Plan the search, run it in passes, and write a cited report.

Decision Tree

User request
  |
  ├─ Answerable in 1-2 searches? ──> STOP. Use WebSearch directly.
  ├─ Debugging or code question? ──> STOP. Use Explore agent or Read tool.
  ├─ "Compare X vs Y" detected? ──> COMPARISON mode (see methodology.md)
  └─ Research question ──> Select mode:
       ├─ User says "quick" or simple topic ──> QUICK
       ├─ Default ──> STANDARD
       └─ User says "deep" or "thorough" ──> DEEP

Modes

| Phase | Quick | Standard | Deep | |-------|-------|----------|------| | 1. PLAN | 2 orientation searches, 3 sub-queries | 2 orientation searches, 5 sub-queries | 3 orientation searches, 7 sub-queries | | 2. RESEARCH | Parallel sub-queries only | + 2 sub-agents + 1 gap-fill round | + 3 sub-agents + 2 gap-fill rounds + source curation | | 3. SYNTHESIZE | Direct synthesis, no critique | Outline refinement + critique | Outline refinement + three-angle critique + delta-queries | | Target sources | 8+ | 15+ | 25+ |

Length follows evidence: Quick answers the question; Standard and Deep cover every sub-question with cited evidence.

Execution

Step 0: Run date +%Y-%m-%d for the real current date. Use that date in every search.

Phase 1 -- PLAN: Read reference/methodology.md and follow the PLAN phase.

Phase 2 -- RESEARCH: Keep following reference/methodology.md for the RESEARCH phase.

Phase 3 -- SYNTHESIZE: Read reference/report-assembly.md for how to write the report. Read reference/quality-gates.md for what has to pass. Follow the SYNTHESIZE phase in reference/methodology.md.

Validation: Resolve the bundled validator, then run it up to three times until the report passes:

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
RESOLVER=""
for candidate in \
  "$REPO_ROOT/plugins/v1tamins/scripts/resolve-skill-root.sh" \
  "${CLAUDE_PLUGIN_ROOT:-}/scripts/resolve-skill-root.sh" \
  "$HOME/.claude/skills/v1-deep-research/../../scripts/resolve-skill-root.sh" \
  "$HOME/.codex/skills/v1-deep-research/../../scripts/resolve-skill-root.sh" \
  "$HOME/.claude/plugins/cache"/*/v1tamins/*/scripts/resolve-skill-root.sh \
  "$HOME/.codex/plugins/cache/v1tamins/v1tamins"/*/scripts/resolve-skill-root.sh; do
  [ -x "$candidate" ] && RESOLVER="$candidate" && break
done
if [ -z "${RESOLVER:-}" ]; then
  echo "ERROR: Could not find resolve-skill-root.sh" >&2
  exit 1
fi
SKILL_ROOT="$("$RESOLVER" v1-deep-research scripts/validate_report.py)"

python3 "$SKILL_ROOT/scripts/validate_report.py" --report [path]

Output Contract

Every report includes:

Executive Summary (200-400 words, standalone, no citations)
Structured body sections with inline citations [N]
Methodology note (what was searched, how many sources, date range)
Limitations / gaps acknowledged
Bibliography with numbered entries matching inline citations
All sources hyperlinked

Output location: ~/Documents/Research/[Topic]_[YYYYMMDD]/report.md

Requested video summary

When the user requests a narrated video of the research, hand the central finding, supporting sources, material uncertainty, and a concrete example to v1-explainer-video. Reuse existing research; do not repeat the investigation merely to change formats. Keep the cited report as detailed evidence and select one lesson for the video instead of narrating the whole report. Preserve the requested report or --context output; video generation is an additional requested format, not a default research step.

Flags

--quick / --deep -- override mode selection
--context -- emit compact summary for other skills (no full report)
--compare -- force comparison mode for "X vs Y" queries

Agent Skills: Deep Research

Install this agent skill to your local