Dzhuneyt Research Fan-out
Turn one fuzzy research goal into parallel, cost-tiered investigation. Interview to scope → get an approved plan tree → fan out subagents → roll up findings as each completes → deliver a report.
Task: $ARGUMENTS
Distinct from deep-research (non-interactive web-search harness): this skill front-loads an interactive scoping interview and an approvable plan tree, tiers models per branch for cost, and streams a per-fork rollup the user can steer mid-flight.
Stage 1 — Orient
- Read
$ARGUMENTS. If clear, go to Stage 2. - If terms are ambiguous, run 1–3 quick web searches to ground yourself before interviewing. Goal: understand the domain well enough to count research forks correctly — not to answer the question yet.
Stage 2 — Interview (max 3 questions)
Narrow the scope. Hard cap: 3 questions, one at a time.
- Multiple-choice preferred; open-ended when needed.
- After every user answer, reply with a 1–2 sentence glanceable restatement of the understood goal (no more).
- Stop early once scope is clear — do not burn questions for their own sake.
Stage 3 — Plan tree (approval gate)
Present a nested bullet tree — the "big picture" of the research path:
- ~3 top-level branches (adaptive 2–5), each phrased as a question the user should be asking (the unknown unknowns), not a statement.
- Sub-bullets decompose each branch into leaf questions. Leaves are the parallel work units.
- Cap total leaves at ~8 for cost. If trimming, say so and name what was cut.
Then: user approves, or refines free-form. Loop — re-present after each refinement. Do not launch until a clear go-ahead ("go", "run it", "approved").
Stage 4 — Tier and launch
Tag each leaf by intuitive complexity, then launch all leaves as background subagents at max parallelism (single message, multiple dispatches). Use general-purpose agents — web-primary, may also read local files if the topic calls for it.
| Tier | Leaf looks like | Claude Code model |
|---|---|---|
| Cheap | lookup, fact-gathering, mechanical | claude-haiku-4-5 |
| Standard | typical synthesis / comparison | claude-sonnet-5 |
| Deep | ambiguous, high-judgment, cross-cutting | claude-opus-4-8 |
On non-Claude hosts, map tiers to that host's cheap / mid / flagship models.
Give each subagent a tight brief and a fixed return shape:
## <leaf question>
- Findings: <concise, evidence-backed>
- Sources: <urls / refs>
- Confidence: high | medium | low
- Gaps: <what it couldn't resolve>
Stage 5 — Live rollup
Monitor the running forks. As each fork completes, post its condensed findings to chat immediately (don't wait for all).
- User may interject with course-corrections → adjust, spawn, or kill remaining forks accordingly.
- User may also stay silent and keep watching — keep rolling up.
- If a fork fails or returns null: note it in one line, continue with the rest.
Stage 6 — Synthesize
When all forks are done:
- Dedup overlapping findings; flag contradictions between forks explicitly.
- Write the full report to
research/<topic-slug>-YYYY-MM-DD.md— branches as sections, findings, sources, open gaps. - Show the executive summary in chat (answer to the original goal + key findings + confidence + what remains unknown).
Anti-patterns
- Don't skip the interview and jump to launching subagents.
- Don't launch before the plan tree is explicitly approved.
- Don't give every leaf the same model — tiering is the cost mechanism.
- Don't batch the rollup into one dump at the end; roll up per-fork completion.
- Don't phrase top-level branches as answers — they are the questions to investigate.
- Don't let leaves sprawl past the cap without telling the user what was cut.