Skip to main content
Systematically explores a topic using parallel agents across applicable orthogonal dimensions (WHO / WHAT / HOW / WHERE / WHEN / WHY / LIMITS). Unlike a quick research brief, this skill provides structured multi-dimensional coverage with source quality tiers, round-by-round cost gates, and risk-stratified fact verification. Coverage is bounded by a user-controlled round budget; the final report honestly characterises what was covered and what wasn’t.

​
Invocation

/deep-research [seed topic or question]
Optional flag: 
/deep-research [topic] --auto
The --auto flag skips all interactive round gates and runs to max_rounds. Use it for unattended runs. There is no cost circuit-breaker in --auto mode — set an appropriate max_rounds before starting.

​
Model tier strategy

Three tiers balance cost and quality. The coordinator (main session) always handles synthesis and gap detection — these are never delegated.
TierModelUsed forEst. cost / agent
ScouthaikuDepth ≥ 2 directions, low priority, low-stakes verification~$0.05
ResearchersonnetDepth 0–1 high/medium, all seed directions~$0.30–0.60
Deep DiveopusRe-exploration only when exhaustion_score ≤ 2~$3–5
Tier selection logic (applied at spawn time): 
if direction.depth == 0:                              # → Researcher (sonnet)
elif direction.depth == 1 and priority == "high":     # → Researcher (sonnet)
elif direction.depth == 1 and priority == "medium":   # → Scout (haiku)
elif direction.depth >= 2:                            # → Scout (haiku)
elif re_exploration and exhaustion_score <= 2:        # → Deep Dive (opus)
else:                                                 # → Scout (haiku)
Expected cost for a full run: ~15–25(vs 15–25 (vs ~170 with all-Opus).

​
Pre-run scope declaration

Before any agents are spawned the skill shows you a scope declaration and waits for your confirmation: 
Deep research: "{seed}"
Interpretation: [one-sentence interpretation]
Applicable dimensions (N): [list]
Initial directions: {count}
Estimated rounds needed: {low}–{high}
Suggested max_rounds: {recommendation with rationale}
Wall-clock estimate: {time range}

Set max_rounds [default {recommendation}]: _
Continue? [y/N]
max_rounds is a soft gate — when reached with a non-empty frontier the skill prompts you to extend. You can always add rounds. Only --auto converts it to a hard stop. The absolute ceiling is max_rounds × 3. Recommendation formula: 
min_rounds = ceil(initial_directions / 6)  # 6 agents per round
recommended = ceil(min_rounds * 1.5)       # 50% expansion for sub-directions
recommended = max(recommended, 8)          # floor of 8 rounds

​
Workflow

1

Phase 0: Seed validation

Before any directions are generated, three checks fire in sequence:
  • Safety check — if the seed requests harmful or illegal research, refuse immediately.
  • Ambiguity check — if the seed has multiple plausible interpretations, confirm which one to use before proceeding.
  • Input validation — if the seed is too thin (a single proper noun without context), ask for more scope.
2

Phase 1: Seed expansion

Assess which dimensions from WHO / WHAT / HOW / WHERE / WHEN / WHY / LIMITS are applicable using the multi-context table:
DimensionHistorical/socialTechnical/scientificPolicy
WHOKey people, institutionsResearch groups, standards bodiesAgencies, legislators
WHATEvents, phenomenaTechniques, architecturesPolicies, regulations
HOWMechanisms, causationAlgorithms, protocolsEnforcement, incentives
WHEREGeography, settingsDeployment environmentsJurisdictions
WHENChronology, sequenceMaturity level, adoption windowsLegislative calendar
WHYMotivations, driversTradeoffs, design constraintsPolitical economy
LIMITSConstraints, boundariesTheoretical bounds, known failuresEnforcement gaps
Generates 2–4 directions per applicable dimension plus cross-dimensional intersections. Maximum 25 initial directions.
  • 0 applicable dimensions → error; ask user to clarify
  • 1–2 applicable dimensions → warn user; ask to confirm before proceeding
  • 3+ applicable dimensions → proceed
Shows the pre-run scope declaration (see above) and waits for your max_rounds input.
3

Phase 2: Initialize state

Creates deep-research-state.json and deep-research-findings/ in the current working directory. Writes a lock file deep-research-{run_id}.lock before spawning any agents.
4

Phase 3: Research rounds

Each round fires a prospective gate before any agents are spawned:
About to run Round N: {frontier_size} directions queued
Estimated tokens this round: ~{estimate} ({cost_estimate})
Total spent so far: ~{running_total}
Continue? [y/N/redirect:<focus>]
Per round:
  1. Pop up to 6 highest-priority directions from the frontier
  2. Select model tier for each direction
  3. Spawn agents in parallel with an 8-minute timeout
  4. Collect all new directions from completed agents before deduplication
  5. Apply dedup against the stable pre-round frontier snapshot
  6. Update the coordinator summary
  7. Run round-level dimension re-assessment (corrects cold-start errors)
  8. Increment round counter
Timed-out directions are marked timed_out and are not re-queued.
5

Phase 4: Fact verification

After the final research round, before synthesis:
  • Claim extraction — identify the top N significant factual claims (N = min(20, total)). Risk-stratified sampling prioritises: single-source primary → numerical/statistical → contested → corroboration candidates.
  • Citation spot-check — fetch each sampled URL; confirm the attributed claim appears in the source text. For numerical claims: compare exact numbers — semantic similarity is not accepted.
  • Corroboration independence check — for claims cited by 3+ agents, verify the sources are from different organisations, dates, and methodologies.
Paywalled sources are classified as “unverifiable — full text inaccessible.” Accessible sources where the claim cannot be found are flagged as “citation mismatch — manual verification required.”
6

Phase 5: Synthesize

Three-pass synthesis:
  1. Mini-syntheses — each agent writes a mini-synthesis in its findings file.
  2. Theme extraction — coordinator reads mini-syntheses only (not raw findings). A theme is valid only if it requires findings from 2+ distinct dimensions.
  3. Final report — writes deep-research-report.md.
After the report is written, the skill offers an optional deep-qa pass:
QA pass available. Run deep-qa on this report? [y/N]
If you accept, /deep-qa audits the report for citation accuracy, logical consistency, coverage gaps, and counter-evidence gaps.
7

Phase 6: Termination check

The run terminates when any of these is true (first condition wins):
  1. User stops — you chose N at a round gate
  2. Coverage plateau — no new dimensions for 3 consecutive rounds AND all frontier items have exhaustion ≥ 4
  3. Budget soft gate — max_rounds reached with non-empty frontier → you are prompted to extend
  4. Frontier empties — all directions explored (possible since direction reporting is optional)
The final report includes a termination label: User-stopped / Coverage plateau / Budget limit / Convergence.

​
Self-review checklist

Before delivering output, verify all of the following:
  • State file is valid JSON after every round
  • No direction has status in_progress after a round completes
  • Every findings file has: Findings, Source Table, Mini-Synthesis, New Directions (or “terminal node”), and Exhaustion Assessment
  • No direction explored more than 2 times
  • Prospective gate was shown before each round (or --auto was set)
  • Coordinator summary updated each round in structured format, not freeform
  • Fact verification ran before final synthesis
  • Final report includes Spot-Check Sample Results section with explicit limitations
  • Final report uses a termination label from the defined vocabulary
  • Two separate confidence scores in the report: Coverage % and Evidence Quality
  • Model tier correctly selected for each agent

​
Golden rules

Never spawn an agent without reading the state file first. Deduplicate every direction before adding it to the frontier.
A terminal node is valid output. Do NOT force agents to invent new directions to fill the slot.
Always pop the highest-priority direction first. Agent-discovered (child) directions receive a +2 depth bonus over same-tier siblings at the same depth level.
Each direction can be explored at most twice. A third attempt is skipped without re-queuing.
Never spawn agents without showing the user a cost estimate first — unless --auto is set.
Never accumulate raw findings in the coordinator. Use the structured coordinator summary to keep context size predictable.
Web search URLs required for every claim. Training-data-only findings are not accepted.
Never let agents default to a tier. Unintentional Opus usage is the primary source of cost spirals.
Flag all numerical claims in the spot-check. LLM number verification is unreliable — exact comparison is required.

​
Reference files

FileContents
DFS.mdDimension discovery, cross-product expansion, exhaustion map, frontier priority ordering, termination logic
STATE.mdState file schema, direction schema, deduplication contract
SYNTHESIS.mdFact verification protocol, coordinator summary format, final report structure
FORMAT.mdFinal report format, coverage report, spot-check results section