Custom AI roles — the catalogue this role comes from.

Researcher — Assemblable today

Answers a research question from the sources you give it, and renders the answer as a document you can send.

Build this one

Start from this role and tell the designer what changes — no need to describe it again.

Start from this role

What it does

The Researcher takes a question and the material you hold on it — a reading list, a document set, a folder of PDFs, decks and spreadsheets — reads each source into its own text before quoting it, and writes the answer with every claim tied to the source it came from. The brief separates what the sources say from what the reading of them supports, builds a synthesis table that shows which sources carry which claim, and renders to PDF, DOCX or HTML. A second model can be attached to argue the other side of a contested point before the brief commits to it.

What you get

ArtefactProduced by
Research brief — the question, the answer, the evidence under it and the gaps in itdocuments/report v0.2.0, rendered by documents/office v0.3.0
Sourced report — each claim carrying its source and the passage it rests ondocuments/report v0.2.0
Synthesis table — source, claim, what it supports, and how stronglydocuments/data_table v0.2.0
Source register — every document read, its format, its date and what it was used fordocuments/data_table v0.2.0
Cross-check on a contested claim by a second model, in parallel with the firstQwen qwen3.8-flash, optional supplier
Reading PDF, DOCX, PPTX and XLSX, and reading back any document you supplydocuments/extract v0.2.0, documents/office v0.3.0
Reaching the workspace from a chat instead of a browserconnections/telegram, optional

Nothing in this assembly fetches the open web: the parts read, write and render, and the two suppliers are models. So the evidence base is what you hand over — which is the honest constraint worth knowing before the design conversation, and the reason the source register exists. If live retrieval from the web is part of the job, that is a part to be added, not behaviour to be assumed.

The parts it is made of

What it costs

A month of this role is 88 paid calls and $0.64. That is the whole bill.

The calls are two models: DeepSeek at $0.00700 for the reading and the drafting, Qwen at $0.00875 for a cross-check, across eight briefs and eight tables in twenty working days. The document parts carry no published per-call price and sit outside the total below.

The cross-check is a choice: without it the month is 72 calls and $0.50400.

CallUnit priceCallsCost
DeepSeek deepseek-v4-flash — reading the source set, six calls a brief$0.0070048$0.33600
DeepSeek deepseek-v4-flash — drafting the brief, the report and the table, three calls a brief$0.0070024$0.16800
Qwen qwen3.8-flash — cross-check on a contested claim, two calls a brief (optional)$0.0087516$0.14000
Documents: extract, report, office, data_tableno published per-call price8 briefs, 8 tablesnot costed
connections/telegram (optional)no published per-call priceone channelnot costed
Total88 paid calls$0.64400

An illustrative month, twenty working days: eight briefs, each over a source set of ten to fifteen documents — six reading calls to work through them, three drafting calls for the brief, the report and the table — so 72 DeepSeek calls, plus two Qwen cross-checks a brief, so 16. A wider source set costs more per brief, roughly linearly. The counts are assumptions, not observations.

The arithmetic is short. Seventy-two DeepSeek calls at $0.00700 is $0.50400. Sixteen Qwen calls at $0.00875 is $0.14000. The two add to $0.64400 for 88 paid calls: $0.08050 a brief, $0.03220 a working day and $7.73 over twelve months at the same pace. Without the cross-checks the month is $0.50400 — the second model is a choice, not a requirement, and a brief that settles nothing contested wastes it. The document parts carry no published per-call price, so eight briefs and eight tables sit outside the total: the figure understates the work. At the metered ceiling of $0.00955 the DeepSeek calls cost $0.68760 and the month $0.82760, 29 per cent more. The 0.50 USDC opening balance is seventy-one calls at $0.00700 — six briefs at these counts, or about sixteen working days.

Services it buys, per call

What it refuses

Three boundaries, each one a limit the workspace enforces rather than a preference of the agent's.

It will not carry an unsourced claim. Every assertion in the brief points at a document, a passage and a date. Where a statement rests on the model's own reading rather than on a source, it is labelled as such in the text — and where it would change the answer, it is left out instead of being softened into agreement with what the sources do say.

It will not invent a citation. No plausible-looking title, author, year, page or link to fill the end of a sentence. A missing source is reported as missing, and the sentence it was meant to carry is marked as resting on nothing until the document is supplied.

It will not pass a summary off as a source. A secondary account it was handed — an abstract, a review, a press summary — is cited as second-hand and never becomes the authority for a figure. Where the primary document is not in the set, the brief says the claim is corroborated only at one remove, and at what strength.

Questions

What does it cost to start?

0.50 USDC on Base, in your own wallet and credited to the workspace as working capital rather than a fee. At $0.00700 a call that is seventy-one calls, and a brief at the counts above is eleven — nine DeepSeek calls and two cross-checks — so the balance is six briefs, or about sixteen working days at this pace. If you skip the cross-checks it stretches to eight briefs. The three ceilings stop one long reading session from spending it.

What access does it need?

The source set, and that is the main input: ten to fifteen documents a brief in a format documents/extract can read, with their dates. It needs no mailbox and no account, and an optional Telegram channel lets you reach the workspace from a chat. The documents and the register live in the workspace volume, which is yours, and one client's corpus does not enter another's context.

What will it not do?

Retrieve from the web, invent a citation, or present a summary as a primary source. It also will not let a strong question force a weak answer: where the sources do not settle the point, the brief says so, gives what they do support, and names the document or the kind of evidence that would close the gap. The conclusion is yours to draw from a stated evidence base rather than inferred from a confident paragraph.

Start

The first useful output is one brief on one question you have already tried to answer by hand, built from the documents you would have read: the answer, the evidence under it and the gaps it leaves. Start at /design?start=researcher: about ten exchanges settle the source set, whether a second model is attached and the ceilings, and a person approves the proposal before anything is provisioned.

Build this one

Start from this role and tell the designer what changes — no need to describe it again.

Start from this role

Other roles