Research Agents

PRIMA Scholar

PRIMA Scholar includes three agents for research jobs that take many steps. An agent is a helper that Claude starts in a separate context. It does the searching and reading on its own, writes its full output to a file and returns only a short summary. Your conversation therefore stays short, even when the agent reads dozens of papers.

The Three Agents

Agent	What it does	Model	Output
`research-agent`	Researches a question in depth and writes a themed synthesis with APA7 citations and confidence tags	`sonnet`	A report file and a three to five line summary
`form-completion-agent`	Fills in a structured form, such as a grant application or a literature review matrix, from your library and the databases	`sonnet`	The completed form and a count of the fields filled and the fields left for you
`literature-watch-agent`	Runs one scan for the Literature Watch	`opus`	A digest, a JSON file of DOIs and counts

The model is the one named in each agent’s definition file.

Asking Claude to Use an Agent

Ask for the agent by name and say what you want, including where to save the result.

Use the research-agent to review the evidence on psychological safety in
hospital teams. Save the report to Research/psychological-safety.md.

Claude Code can also hand a request to an agent on its own when the request matches the agent’s description, such as a request for a systematic review. The literature-watch-agent is the exception, because the literature-watch skill starts it and you do not call it directly.

research-agent

The research agent handles questions that need more than one search, such as a literature review or a question with several sides.

It breaks your question into three to five sub-questions that cover different sides of the topic.
It searches the databases for each sub-question, narrowing the databases when the discipline suggests it. For example, it uses ERIC for education and PubMed with Europe PMC for health research.
It picks out the key papers, both highly cited foundational work and recent developments.
It follows citation chains from the most relevant papers, forwards to work that cites them and backwards to the work they cite.
It checks your research library for documents you already hold, and prefers them where they fit.
It writes a report organised by theme, with an APA7 citation for every factual claim and a reference list of only the papers it cites.
It imports to your library the key papers you are likely to cite again.

Each major claim in the report carries a confidence tag.

Tag	Meaning in the research agent
HIGH	Several replicated studies, meta-analyses or systematic reviews
MEDIUM	A single well-designed study with limited replication
LOW	Theory, anecdote, conflicting evidence or preliminary findings

The report opens with the date, the sub-questions, the databases searched and the number of papers reviewed. An introduction follows, then three to five themed sections, a section on gaps and future directions, and the references.

If you do not give a location, the agent saves the report to ~/Documents/Agent Outputs/research/ with a name made of the date, time and topic, such as 2026-10-14-1530-psychological-safety.md. It creates the folder if it does not exist.

form-completion-agent

The form completion agent takes a template with questions or fields and answers each one from evidence. It suits grant applications, compliance forms and literature review matrices. Give it the path to the form and say where to save the result.

It reads the form and numbers every question or field.
It groups related questions, so one search can serve several fields.
For each field it searches your research library first and reads the relevant documents. It searches the databases only when the library does not hold enough.
It drafts each answer with APA7 citations and a confidence tag.
Its instructions are to present completed sections in groups of three to five for your review and to revise answers you say are weak.
It writes the completed form once every section is approved.

Tag	Meaning in the form completion agent
HIGH	Several sources that agree, with a strong evidence base
MEDIUM	One reliable source, or an answer extrapolated from related evidence
LOW	Limited evidence, or theory only

The agent keeps the form’s structure exactly as it is, without moving sections or renaming fields. When it finds no reliable source for a field, it writes [Insufficient evidence — manual input needed] in place of an answer and never fills the gap with guesswork. A reference list closes the form.

If you do not give a location, the agent saves the form to ~/Documents/Agent Outputs/forms/ with the date, time and form name in the file name. It returns the file path, the number of fields in the form, the number completed with citations and the number marked for manual input.

literature-watch-agent

This agent runs a single literature watch scan. From the literature-watch skill it receives the settings file, the start date, the DOIs to leave out and the two output paths. It then searches, screens, writes the digest and imports the selected papers, and returns only the paths and counts. The Literature Watch page describes the whole process.

APA7 citations. All three cite their sources in APA7. The research agent and the literature watch agent use the citation strings that the search tools return.
UK English. All three agents write in UK English.
Lean returns. Each agent writes its full output to disk and returns only the path with a short summary or counts. Open the file to read the work.
Your library grows. The research agent and the literature watch agent add papers to your library. Check Collections and Tags to keep it organised.

Changing an Agent

Each agent is a markdown file in the agents/ folder of your PRIMA Scholar installation. If you installed with the installer, updates overwrite the three agent files, so copy an agent to a new name, such as my-research-agent.md, before you change it. Installing in Claude Code covers both installation routes.