Agent Skill for Research

An agent skill for research is a reusable workflow that tells an agent how to investigate a question, rank sources, preserve citations, separate evidence from inference, and return a brief a human can audit. It is not just a better version of "research this." It is the operating procedure behind a reliable research run.

That distinction matters because modern research agents can already search the web, read uploaded files, query internal systems, and produce long reports. OpenAI describes deep research in ChatGPT as a workflow where the user chooses sources, reviews a proposed plan, follows progress, and receives a structured report with citations or source links. Anthropic's multi-agent research system uses a lead agent that plans the work and dispatches specialized subagents to search different angles in parallel. Those are powerful patterns, but power alone does not decide which source should win, which citation is weak, or when the agent should stop.

The research skill supplies those rules. If you are still defining the category, start with What Is an Agent Skill?. This article focuses on the narrower job: packaging a source-first research method into a skill that can be reused by a founder, analyst, engineer, or support team.

Prompt	Research skill
"Research this topic."	Defines the question, source hierarchy, search plan, citation rules, and output shape
Accepts whatever appears first	Requires primary sources first when they exist
Treats citations as decoration	Treats citations as evidence that must support specific claims
Blends fact, summary, and recommendation	Separates confirmed facts, interpretation, unknowns, and next steps
Produces uneven work across runs	Produces comparable briefs with the same sections and caveat rules
Keeps going when sources are missing	Stops or downgrades confidence when required evidence is unavailable

For low-stakes brainstorming, a prompt may be enough. For competitor research, technical landscape scans, policy-sensitive summaries, investment memos, academic literature reviews, or customer evidence synthesis, a skill is safer because the method matters as much as the answer.

A research skill should tell the agent what kinds of sources count, how they are used, and how much confidence they deserve.

Source tier	Examples	Use
Primary	Official docs, specifications, filings, product pages, APIs, research papers, datasets	Facts, definitions, dates, product behavior, formal claims
Authoritative secondary	Standards bodies, reputable journalism, analyst reports, library guides, peer-reviewed surveys	Context, synthesis, expert interpretation
Community	GitHub issues, forums, Reddit, Discord, Hacker News, Stack Overflow	Live sentiment, edge cases, implementation pain
Internal	Customer notes, call transcripts, product analytics, support tickets, CRM records	Company-specific evidence and customer reality

The skill should require source labels. A claim from an API reference, a Reddit comment, and a sales transcript should not look equally certain. Community sources are useful, but they usually prove that people are saying something, not that the thing is true.

This is the same source-governance problem covered in Agent Skill for Context Engineering. Context engineering decides what the agent should load and trust. A research skill applies that discipline to discovery: search here first, treat this source as authoritative, use this only as anecdote, and report conflicts instead of smoothing them away.

Use this sequence inside the skill:

1. Restate the research question and decision context.
2. Identify what would count as primary evidence.
3. Decide which sources are required before synthesis can begin.
4. Search primary sources first, then authoritative secondary sources.
5. Use community sources only after labeling their role.
6. Collect claim-level notes with source URL, title, date, and source type.
7. Cross-check important claims against at least two sources when possible.
8. Separate confirmed facts, interpretation, recommendation, and unknowns.
9. Return a concise brief with citations, caveats, and confidence.

The point is not to slow every task. The point is to make the agent choose the correct research depth. OpenAI's web search tool documentation distinguishes fast lookup from agentic search, where a reasoning model manages searches, analyzes results, and decides whether to keep searching. A research skill should make that tradeoff explicit: use fast search for a current product date; use a deeper run when the answer needs synthesis across many sources.

Anthropic's research architecture offers a useful named example. Its lead agent breaks a broad query into subtopics and sends subagents to investigate in parallel. A skill for market mapping can use the same shape: one branch for official company pages, one for funding and filings, one for customer sentiment, one for technical docs, and one for risks. The lead agent should then merge the branches into one answer with source labels and conflict notes.

## Research Brief

### Research question
Restate the exact question and the decision it supports.

### Bottom line
One direct paragraph answering the question with confidence level.

### Confirmed facts
- Claim - source link - source type - why it matters

### Interpretation
- What the facts suggest, clearly labeled as interpretation

### Gaps and caveats
- What could not be verified and why
- What may change over time
- Which sources were unavailable, paywalled, stale, or indirect

### Recommendation
- Only include this section when the evidence supports a decision

### Evidence inventory
- Primary sources
- Secondary sources
- Community or anecdotal sources

This template keeps the answer auditable and prevents the agent from hiding uncertainty in smooth prose. It also gives human reviewers a stable shape across different topics.

Citations should be treated as structured data the workflow preserves, not as links the model invents at the end.

OpenAI's Deep Research API cookbook shows a response that includes inline citations, source metadata, and intermediate steps. OpenAI's file search docs also expose annotations and allow developers to include retrieved search results in the response. Anthropic's search results documentation uses source, title, and content blocks so Claude can cite supplied search results accurately. Perplexity's Search API returns structured result objects with fields like title, URL, snippet, date, and last updated, while Sonar returns prose with built-in citations.

Those implementation details matter. If the workflow has raw search results, it should preserve the source map before synthesis begins. Perplexity's citation parsing guide recommends collecting search results before processing streamed text so the URL map is available when citations appear. Its output-control guidance is more direct: do not rely on a model to generate valid links inside structured output; use links returned in citations or search result fields.

A practical research skill can encode the same rule:

Citation rule:
- Never fabricate a citation from memory.
- Cite only URLs, file paths, document IDs, or chunk IDs returned by a tool or inspected directly.
- If a claim was inferred from multiple sources, cite the strongest supporting sources and label the inference.
- If a citation points to a broad page but not to the specific claim, mark it as weak support.

This is where research skills and retrieval systems meet. Agent Skill vs RAG covers the bigger comparison, but the short version is simple: retrieval supplies candidate evidence; the skill decides how evidence may be used.

Research agents often use web search, document stores, databases, browser tools, dashboards, or MCP servers. The skill should define tool rules before the agent starts calling them.

For a technical research skill, the rules might be:

• Use official docs, release notes, and repository files before general web search.
• Use file search or RAG for uploaded specs, policy libraries, or internal knowledge bases.
• Use web search for current public information, company pages, filings, and fresh community sentiment.
• Use internal data only when the user has asked for internal analysis.
• Do not cite private customer data in public-facing output.
• Include source URLs, file paths, document IDs, or query names for decisive claims.
• Mark every source by type and recency.
• Stop if required systems are unavailable.

The tradeoff is control versus convenience. Web search can find fresh material quickly, but it may surface summaries, copied content, spam, or unreviewed posts. RAG over approved documents can be safer, but it may miss new information or fail if chunks are poorly retrieved. A research skill should route the agent to the right tool for the claim being made.

This matters for literature review workflows. A skill for academic research should prefer databases, papers, systematic reviews, and DOI-backed records before general web pages. It should record inclusion and exclusion criteria, avoid turning abstracts into full-paper claims, and flag paywalled sources the agent could not inspect. A 2025 survey of RAG evaluation methods highlights metrics such as factual accuracy, hallucination rate, citation accuracy, and faithfulness. Those are not just system metrics. They are good questions for a research skill's review checklist.

Before returning the research brief, the agent should verify all of the following:

• The question was answered directly.
• Important claims have source-backed citations.
• Primary sources were preferred where available.
• Community sources are labeled as anecdote or sentiment.
• Dates are included when freshness matters.
• Links came from inspected pages, tool citations, file paths, or search result metadata.
• Source limitations are stated.
• Inference is labeled as inference.
• The final answer does not overstate certainty.
• The brief includes a recommendation only when the evidence supports one.
• The agent did not cite a source for a claim the source does not actually support.
• The agent stopped or downgraded confidence when required evidence was missing.

That last item is critical. Nature reported in 2026 that hallucinated citations were appearing in scientific literature, with invalid references possibly affecting many publications from 2025. In community discussions, practitioners describe the same underlying issue in practical terms: one LocalLLaMA thread explains that citation systems need chunk metadata such as document name, URL, and page number; another Perplexity thread warns that asking for full citations can trigger fabricated or unreliable references. Reddit is not the authority here, but it is useful evidence of where builders and users feel the workflow breaking.

The skill should turn those lessons into guardrails. It should require source metadata, not just source-looking text.

A weak prompt might ask: "Compare LangGraph, CrewAI, and OpenAI Agents SDK."

A research skill would turn that into an auditable workflow:

1. Confirm the user's use case: internal automation, production customer workflow, data analysis, or experimentation.
2. Read official docs for each framework before blog posts or tutorials.
3. Capture the date checked, because agent frameworks change quickly.
4. Compare the same dimensions for each option: tool calling, state, memory, human approval, deployment, observability, evaluation, and ecosystem.
5. Use GitHub issues or Reddit threads only for implementation pain and community sentiment.
6. Return a recommendation tied to constraints such as team size, hosted versus self-managed infrastructure, latency, security, and need for long-running workflows.

That produces a decision memo rather than a generic comparison. For example, "OpenAI's API provides hosted web search and file search tools" is a source-backed capability claim. "Use hosted retrieval because your team is small" is a recommendation that depends on budget, security, customization, and maintenance tolerance. A good research skill keeps those claims separate.

A literature review skill needs a stricter contract than a market memo. It should not simply ask the agent to summarize papers.

The skill should define:

• the research question
• databases or search surfaces to use
• inclusion and exclusion criteria
• date range
• whether preprints are allowed
• required fields for each paper
• how to handle unavailable full text
• what counts as a finding versus the agent's interpretation

For example:

For each included paper, extract:
- citation
- publication year
- study type
- dataset or population
- method
- main finding
- stated limitations
- relevance to the research question
- exact source inspected: abstract, full text, appendix, or secondary summary

This makes the output reviewable. A human can see whether the agent over-relied on abstracts, included weak sources, or ignored contradictory studies.

The first mistake is treating citations as proof by themselves. A link proves only that the agent named a source. It does not prove the source supports the claim. The skill should require claim-level support.

The second mistake is treating all retrieval as research. RAG can find relevant chunks, but it does not automatically evaluate authority, bias, freshness, or completeness. A research skill should say when retrieval is enough and when the agent must search outside the provided corpus.

The third mistake is using community sources without a label. Reddit, GitHub issues, and forum posts are excellent for discovering pain points, edge cases, and vocabulary. They are weak sources for settled facts unless the claim is about community sentiment itself.

The fourth mistake is producing a recommendation without assumptions. If the brief recommends a tool, market, pricing model, or technical direction, it should state the assumptions that make the recommendation true.

The fifth mistake is skipping stop rules. Some missing sources should block the result. A skill researching a private competitor cannot produce a trustworthy pricing comparison if the current pricing page is inaccessible. A literature skill should not generalize from abstracts when full methods are unavailable.

Start with the repeatable job, not the tools. "Research" is too broad. "Prepare a source-backed competitor memo for a founder" is a skill. "Summarize literature on RAG evaluation for an engineering team" is a skill. "Find customer evidence for a product positioning decision" is a skill.

Then write the contract:

1. Name the research decision the brief supports.
2. List required inputs.
3. Define source tiers and source labels.
4. Define search order and tool rules.
5. Define citation rules.
6. Define confidence labels.
7. Define stop rules.
8. Provide the output template.
9. Provide a final verification checklist.

Once drafted, test it the same way you would test any reusable workflow. Run it on a clean question, a messy question, and a question with missing evidence. Check whether it starts from primary sources, labels community material correctly, cites claim-level support, and refuses to overstate uncertain conclusions. How to Test an Agent Skill Before You Publish It gives the broader pre-release loop; for research skills, the decisive test is whether a reviewer can trace every important claim back to evidence.

For Skillfully authors, research is one of the strongest skill categories because the result leaves a trail. A good run produces sources, caveats, confidence, and a decision-ready brief. A weak run produces fluent prose that collapses under review. The skill's job is to make the first outcome repeatable.