From ensembled-research
This skill should be used when the user asks to "research a topic", "create a library doc", "write a reference document", "ensemble research", "research with multiple agents", or wants comprehensive, unbiased research on any topic. Spawns multiple independent researchers on the same topic with no pre-assigned angles, then ensembles results into a single document using independent agreement as a quality signal.
How this skill is triggered — by the user, by Claude, or both
Slash command
/ensembled-research:ensembled-researchThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Produce high-quality reference documents by spawning multiple independent researchers on the same topic, then synthesizing their outputs into a single document using ensemble principles. The key insight: independent agreement across researchers is a strong quality signal, while unique findings from a single researcher are preserved but weighted accordingly.
Produce high-quality reference documents by spawning multiple independent researchers on the same topic, then synthesizing their outputs into a single document using ensemble principles. The key insight: independent agreement across researchers is a strong quality signal, while unique findings from a single researcher are preserved but weighted accordingly.
Spawn N researchers (default: 5) in parallel using the Task tool. All receive the identical prompt with no assigned angles or aspects.
Researcher prompt template:
Research [TOPIC] — [brief description]. Use web search, Context7, and any sources
available to build a comprehensive reference document in markdown.
IMPORTANT: When retrieving web content, use `curl -sL <url> | sed 's/<[^>]*>//g'`
via the Bash tool instead of WebFetch. WebFetch summarizes content through a small
model, likely cached — all researchers would receive the same lossy summary,
destroying independence at the retrieval layer. curl returns raw content that you
interpret independently, which is what makes ensemble agreement meaningful.
Cover whatever is most important and useful for a developer who will be using [TOPIC]
in production with [RELEVANT STACK]. Don't limit yourself to any particular aspect —
cover architecture, APIs, patterns, deployment, client libraries, configuration,
persistence, security, clustering, whatever is relevant. Include code examples, version
numbers, and practical pitfalls.
Return findings as a single markdown document.
Spawn configuration:
subagent_type: general-purposemodel: sonnet (cost-efficient for research volume)run_in_background: true (all run concurrently)researcher-1 through researcher-NAs researchers complete, save each output to a named file:
[topic]-researcher-1.md through [topic]-researcher-N.md~/workspace/library/)Monitor progress via output file sizes. Researchers typically produce 100-200KB of raw output including tool call logs.
Before synthesis, cross-reference key facts across all researcher outputs to identify divergences. Build a correlation table for the facts with the largest variation across researchers.
What to correlate:
How to present: Create a table with researchers as columns and facts as rows. Flag discrepancies:
Unique finds: Note any content that only a single researcher found. These are either:
Present the correlation analysis to the user or use it to inform the next phase.
For each divergence identified in correlation, verify against authoritative sources using web search. Focus on:
Highest priority (always check):
Methodology:
WebSearch to find URLs, then curl -sL <url> | sed 's/<[^>]*>//g' to retrieve raw content — do not use WebFetch, which summarizes through a cached small model and introduces systematic bias shared across all researchersCommon hallucination patterns:
Output: Produce a corrections table:
| Fact | Researchers Said | Reality | Source | Action |
|---|---|---|---|---|
| Library version | 0.35-0.37 | 0.46.0 | crates.io | Correct all |
| Binary size | 10-20MB | <20MB binary, <10MB Docker | Official FAQ | Distinguish both |
With correlation and fact-checking complete, synthesize into a single document. Apply these principles:
Convergence weighting:
Fact corrections:
Structure selection:
Code examples:
Pitfalls section:
Write the ensembled result as the final document (e.g., nats.md). Clean up the individual researcher files unless the user wants to keep them.
Update any indexes (e.g., CLAUDE.md library table) with the new document entry.
Every ensembled document gets a compact fingerprint encoding its provenance:
5
The fingerprint is just the researcher topology. Everything else is fixed by the pattern: researchers are always lo (cost-efficient), synthesizer is always hi (high-capability), correlation and fact-checking always happen.
Examples:
5 — standard ensemble, 5 identical-prompt researchers5×3 — one augmentation round (5 base, 3 augmented)5×3×3 — two augmentation rounds (maximum, 11 researchers)7 — larger base cohort, no augmentationEach ×N after the first is an augmentation round of 3 researchers. The fingerprint is stored with the document for provenance tracking.
The base pattern uses identical prompts. An augmented ensemble adds cohorts with additional framing:
The augmented cohort may surface content the identical cohort missed (e.g., fleet deployment angle, security implications, integration patterns). In correlation, augmented researchers are treated identically — independent agreement is the signal regardless of prompt variant.
Limits: Maximum 5×3×3 (11 researchers). Beyond that, the document needs to be split or the base cohort re-run with a refined scope.
Composability: An existing ensembled document can be augmented later. Launch 3 new researchers with an augmented prompt, correlate against the existing doc, re-synthesize. The fingerprint grows from 5 to 5×3.
| Parameter | Default | Notes |
|---|---|---|
| Researcher count | 5 | More researchers = more diversity but higher cost. 5 is a good balance. |
| Model | sonnet | Cost-efficient for research. Use opus for highly technical topics. |
| Prompt variation | None | All researchers get identical prompts. Variation defeats the purpose. |
npx claudepluginhub viablesys/claude-plugins --plugin ensembled-researchCollects and ranks multi-source technical information (docs, GitHub, Stack Overflow, papers, blogs) before writing PRDs, ADRs, docs, or articles. Outputs a ranked source dossier to memory/research/.
Deep research on any topic by conducting web searches and fetching content. Can also add files or URLs to a research index. Useful for building knowledge bases or investigating unfamiliar subjects.
Researches topics comprehensively with web search, scraping, and AI agents, producing timestamped Markdown documentation in .claude/research/.