Custom AI roles — the catalogue this role comes from.
Finance desk — Assemblable today
A three-analyst desk — macro, financial and trade — that works one brief in parallel and hands over a synthesised view with its disagreements named.
Build this one
Start from this role and tell the designer what changes — no need to describe it again.
What it does
The Finance desk takes one brief and works it three ways at once. The brief is a question, the material that binds it — statements, curve snapshot, price series, filings, each with an as-of date — and the reader who has to act on the answer. The desk splits it into three workstreams: macro computes the curve, breakevens, growth and FX; fundamental values the company from its own statements; trade measures the price series into indicators, a signal and a backtest with costs. All three run against the same brief and the same as-of date.
They run independently — the design, not a detail of scheduling. None reads another's conclusion while it computes. A view that has seen its neighbour's answer is not a second opinion; it is the same opinion with better manners. The disagreement the desk exists to find only exists while the three are separate.
What comes out is a synthesis, not three notes stapled together. Where the three agree, one line says so and names the inputs behind it. Where they diverge, the divergence is a row of the disagreement table: the three values, the input and method behind each, the size of the gap, and whether it is a gap of inputs, method or timing. The note leads with the view the evidence supports, lists the disagreements in the order of what they would change, and hands over to a named reader. A desk that returns three notes and no view has done the work and skipped the job.
What you get
| Artefact | Produced by |
|---|---|
| Macro view — curve, breakevens, growth, FX; every figure computed and labelled | compute/macro_analysis v0.1.0, documents/data_table v0.2.0 |
| Fundamental view — multiples, a DCF bridge, a ratio set and the risk that moves them | compute/fundamental_analysis v0.1.0, documents/data_table v0.2.0 |
| Trade view — indicators, a composite signal, a backtest with its costs stated | compute/technical_analysis v0.1.0, documents/data_table v0.2.0 |
| Disagreement table — one row a contested figure: three values, the input and method behind each, the gap and its kind | documents/data_table v0.2.0 |
| Synthesis note — the supported view first, then each disagreement and the figure that carries it | documents/report v0.2.0 |
| Coverage block — source, as-of and corroboration status per figure, and what could not be obtained | documents/report v0.2.0 |
| PDF or DOCX of the note and its tables | documents/office v0.3.0 |
| Statements, snapshots and series read before anything is computed | documents/extract v0.2.0 |
The disagreement table is a first-class artefact: it is produced whether or not the three views agree. The desk declares seven parts and no connection — it renders to PDF or DOCX inside the workspace and stops. Add the mailbox or a Telegram channel at design time if the reader is elsewhere; this page does not pretend the desk has one today.
The parts it is made of
- documents/reportv0.2.0Report
the synthesised note the desk hands over - documents/data_tablev0.2.0Data table
the comparison table across the three views - compute/macro_analysisv0.1.0Macro analysis engine
the macro computations, not mental arithmetic - documents/extractv0.2.0Reading floor (pure Python)
reads the releases and documents the client supplies - documents/officev0.3.0Document engine (pure Python)
reads and writes working documents - compute/fundamental_analysisv0.1.0Fundamental analysis engine
multiples, DCF and cross-checks, computed - compute/technical_analysisv0.1.0Technical analysis engine
indicators and the backtest, computed
What it costs
The gallery entry prices three compute operations — compute/macro_analysis, compute/fundamental_analysis, compute/technical_analysis — at $0.006125 a call on Base, and DeepSeek at $0.00700. The engines ship in the image and run in-process, so $0.006125 is the over-the-wire price and the month is an upper bound. The counts are assumptions, not observations.
| Call | Unit price | Calls | Cost |
|---|---|---|---|
compute/macro_analysis — yield_curve, breakeven_inflation, growth, fx, four a brief | $0.006125 | 32 | $0.19600 |
compute/fundamental_analysis — statements, multiples, dcf, risk, four a brief | $0.006125 | 32 | $0.19600 |
compute/technical_analysis — indicator, signal, backtest, three a brief | $0.006125 | 24 | $0.14700 |
DeepSeek deepseek-v4-flash — six a brief: three view drafts, two synthesis passes, one table write-up | $0.00700 | 48 | $0.33600 |
Documents: extract, report, data_table, office | no published per-call price | 8 notes, 32 tables | not costed |
| Total | 136 paid calls | $0.87500 |
Review: DeepSeek: per-call chat completions over x402 (2026-10-31)
An illustrative month, twenty working days: eight briefs, two a week, each running eleven engine operations — four macro, four fundamental, three technical — so 88 engine calls, and six model calls a brief, so 48.
The arithmetic is short. Eighty-eight engine calls at $0.006125 is $0.53900; forty-eight model calls at $0.00700 is $0.33600; the two add to $0.87500 for 136 paid calls, which is $0.10938 a brief, $0.04375 a working day and $10.50 over twelve months. The document parts carry no published per-call price, so eight notes and thirty-two tables sit outside that total. At the metered ceiling of $0.00955 the 48 model calls cost $0.45840 and the month $0.99740, 14 per cent more; in-process, where the engines are not bought, the engine lines cost nothing and the month is $0.33600. The 0.50 USDC opening balance is seventy-one model calls at $0.00700: at seventeen paid calls a brief it covers four briefs, with $0.06250 in hand.
What it refuses
Three boundaries, each one a limit the workspace enforces rather than a preference of the desk's.
It will not average the three views into a consensus. No sentence of the form "the desk's view is broadly neutral" where the macro read says one thing and the fundamental read says another. Averaging is not synthesis; it is a fourth view nobody computed, and it is the failure this desk prevents. Where the three disagree, the disagreement is measured and left standing for the reader.
It will not present a forecast as a computed figure. The engines compute over data they are handed and never fetch. Where the next print matters, the note carries a view labelled opinion, with a confidence level and the condition that would change it — and if that view cannot be kept apart from the arithmetic, it is dropped rather than dressed as a result.
It will not give advice. No recommendation, no position, no hedge, no allocation; nothing here is investment, legal or tax advice. The desk produces the evidence and the argument between its own parts. The decision, and the liability with it, stay with the person who defends them.
Questions
Why a desk rather than one agent?
One agent asked for a view picks a lane: the analyst who reads the statements also decides, quietly, that the curve does not matter. The second and third views are the ones that would have changed the answer, and they are worth having only if computed independently. One analyst's note is eight paid calls; a desk brief is seventeen.
What does the disagreement cost me?
Nothing in calls. The three views are computed from the same inputs on the same brief, and the table is one data table inside the same seventeen. What it costs is a decision: where the table shows a gap, the reader chooses between the figures or asks for the input that would settle it. A desk that averaged the gap away would cost the same and be worth less.
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 — four briefs at $0.10938, and part of a fifth. The three ceilings — calls per message, spend per message, spend per day — stop one broad instruction from spending it.
Start
The first useful output is one brief worked three ways: a question you would otherwise answer from a single document, with the statements, a curve snapshot and a price series you already hold, and the disagreement table even where it has one row. Start at /design?start=finance_desk: about ten exchanges settle the brief, the inputs, the reader 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.
Other roles
- Writerassemblable
- Researcherassemblable
- Financial Analystassemblable
- Macro Analystassemblable
- Trade Analystassemblable
- ISO Auditorassemblable
- Purchase managerassemblable
- Customer serviceassemblable
- Salesassemblable
- Social Mediaassemblable
- Curator / Editorassemblable
- Real Estate Assistantassemblable
- Personal assistantassemblable
- Publishing deskassemblable