
Lead Researcher
- 4 installs
- 14 repo stars
- Updated July 28, 2026
- samhvw8/dotfiles
Helps with ai & agent building tasks.
About
lead-researcher is a Claude Code skill for ai & agent building. It helps solo builders move faster with AI-assisted development.
- lead-researcher
- AI & Agent Building
- AI-coding skill
Lead Researcher by the numbers
- 4 all-time installs (skills.sh)
- Ranked #13,349 of 16,556 AI & Agent Building skills by installs in the Skillselion catalog
- Data as of Jul 29, 2026 (Skillselion catalog sync)
npx skills add https://github.com/samhvw8/dotfiles --skill lead-researcherAdd your badge
Show developers this skill is listed on Skillselion. Paste this into your README.
| Installs | 4 |
|---|---|
| repo stars | ★ 14 |
| Last updated | July 28, 2026 |
| Repository | samhvw8/dotfiles ↗ |
What it does
Helps with ai & agent building tasks.
Files
Lead Researcher
Orchestrator-worker pattern for complex research. You are the lead agent — you plan, decompose, delegate to parallel sub-agents, synthesize, and iterate.
Based on: https://www.anthropic.com/engineering/multi-agent-research-system
When This Skill Applies (Complexity Test)
Use lead-researcher when 2+ of these are true:
| Signal | Example |
|---|---|
| Multi-domain | "Compare auth approaches across mobile, web, and IoT" |
| Multi-perspective | "Evaluate X from security, performance, and DX angles" |
| Contradictory landscape | Topic where communities disagree |
| Breadth + depth needed | 5+ independent sub-questions |
| Multi-language value | Chinese/Russian communities have distinct signal |
| Synthesis-heavy | Answer requires weighing tradeoffs across findings |
If only 0-1 signals match → use research skill directly (single-agent research).
Workflow
Phase 1: Plan (You, the Lead)
Before spawning agents, think through:
1. Decompose the question into 3-5 independent sub-questions 2. Assess complexity per sub-question (simple=1 agent, comparison=2, deep=3) 3. Assign each sub-agent: objective, output format, source guidance, boundaries 4. Budget total effort: max 3 parallel agents per wave
Write a brief plan (3-5 bullets) before delegating. Save it — context will grow large.
Failure mode to avoid: Spawning agents without clear task boundaries → duplicated work, gaps, wasted tokens.
Phase 2: Delegate (Parallel Sub-Agents)
Spawn sub-agents with the researcher agent type. Each gets:
[Specific objective — what to find]
[Output format — what to return]
[Source guidance — where to look, what languages]
[Boundaries — what NOT to investigate, what's handled by other agents]
RECOMMENDED SKILLS: research - use for search methodology and gh CLI patternsScaling rules:
| Complexity | Agents | Tool calls each |
|---|---|---|
| Fact-finding | 1 | 3-10 |
| Comparison | 2-3 | 10-15 |
| Deep multi-field | 3 (max per wave) | 15-30 |
Fire all independent agents in a single message (parallel execution).
Phase 3: Synthesize (You, the Lead)
When agents return:
1. Evaluate — Are there gaps? Contradictions? Missing angles? 2. Cross-reference — Do findings from different agents agree? Where do they conflict? 3. Decide — Is another wave needed? (max 2 waves total, 6 agents total) 4. Merge — Combine into coherent findings with source attribution
Iteration trigger: Only spawn wave 2 if wave 1 revealed a critical gap you couldn't have predicted upfront.
Phase 4: Report
Structure the final synthesis:
- Lead with the answer/recommendation
- Present tensions between sources honestly
- Every claim MUST cite its source — inline
[text](url)next to each claim - List unresolved questions at end
- End with
## Sourcessection listing all URLs used - No claim without a source — if you can't cite it, don't include it
Anti-Patterns
| Trap | Fix |
|---|---|
| 5+ agents for simple query | Check complexity test — probably just research skill |
| All agents searching same thing | Give explicit non-overlapping boundaries |
| Endless iteration waves | Max 2 waves. Ship what you have. |
| Over-investing in one angle | Balance effort across sub-questions |
| No plan before spawning | MUST write decomposition first |
Quality Check
Before reporting to user:
- [ ] All claims have a source
- [ ] Conflicting findings are presented as tensions, not hidden
- [ ] Each sub-question got adequate coverage
- [ ] Report is concise — sacrifice grammar for brevity
- [ ] Report saved to
./research/YYMMDD-<topic>.md