# Screen Desk

> First-look deal screening for private equity teams. Paste an inbound teaser, CIM excerpt or broker email; the browser extracts the deal facts, checks the numbers against each other and screens them against the fund's criteria for free. A paid screen writes the one-page screening memo (pass, further diligence or hard pass), and its open questions hand off to a diligence meeting prep of 15-20 questions.

URL: https://screen-desk.skillsafe.ai/
API tutorial: https://screen-desk.skillsafe.ai/api.html
Derived from the agent skills @anthropics/deal-screening (primary) and @anthropics/dd-meeting-prep, from anthropics/financial-services-plugins (Apache-2.0).
Model: gpt-terra. Runs are metered in SkillSafe credits and need a signed-in account; the prescan, the criteria table and the starter question list are free.

## What the free prescan does (in the browser, no AI)

- Extracts revenue, EBITDA (adjusted and reported), EBITDA margin, revenue growth, asking price or valuation, EV/EBITDA multiple, largest-customer and top-N customer concentration, recurring revenue share, headquarters, places named, deal type (platform, add-on, recap, minority, majority, carve-out, growth), management continuity (staying, leaving, mixed) and seller motivation - each with its line and period. Markdown or tab-separated tables are read row by row, and a "($M)" header scales bare numbers.
- Checks the material against itself: the same year's revenue or EBITDA stated two ways; EBITDA above revenue; a stated EBITDA margin that is not EBITDA / revenue (more than 1 point apart); a stated growth rate or CAGR that no span of the revenue series produces (1.5 points tolerance); a stated multiple the price and EBITDA do not imply; a forecast growing more than 10 points a year faster than history; add-backs above 20% of adjusted EBITDA; negative EBITDA; the largest customer at 20% of revenue or more; management leaving; missing revenue, EBITDA, price or reason for sale.
- Screens the deal against the fund's box, the deal-screening skill's ten rows: revenue range, EBITDA range, EBITDA margin, growth, sector fit, geography, deal size / EV (stated, or implied by multiple x EBITDA), valuation multiple, customer concentration, management continuity. Each row is meets, misses, unknown (the material is silent) or no target. The box is saved to the user's account.
- Writes a starter question list for five meeting types (management presentation, expert call, customer reference, advisor check-in, site visit) from dd-meeting-prep's question bank, with neutral probes for what the prescan flagged, ending with "What haven't we asked about that we should?"

## Lanes (the `task` field)

- `screen` - the screening memo: deal facts, one row per criterion with the model's call and comment, bull case (2-3), bear case (2-3), red flags by severity, 3-8 first-call questions, every prescan flag confirmed or dismissed, a bottom line, and a verdict: `further_diligence` (take the first call), `pass` (decline for now), `hard_pass` (outside the box; must rest on a missed criterion or a critical red flag).
- `prep` - the meeting prep: logistics, exactly three objectives, 15-20 prioritised questions grouped by topic with 5-8 must-asks, every concern addressed by a question that cites it, no leading questions, the closer last; benchmarks from the material (quoted) or to pull before the meeting (no figures); red flags with the neutral probe; follow-up requests.

The screen hands off to the prep with one button: its red flags and bear case become the concerns (C1..Cn) and its first-call questions travel as K1..Kn.

## Input (every value a string)

task, material (required), facts (required, JSON-encoded: figures, qualitative facts, criteria rows with the browser's calls, box verdict, prescan flags), question (optional), meeting (prep: type, attendees, duration_minutes, focus, concerns), screen (prep, optional: the carried screen), retry_note (reformat retry only).

## How the reply is checked

Every quote must be an exact span of the material. Criteria calls on numbers must agree with the browser's arithmetic; sector, geography and management are keyword reads the model may overrule with a reason. Every prescan flag must be answered once. Dollar figures in the deal facts must be figures in the material (arithmetic is noted, not failed). In the prep, every concern must get a question, every reference must exist, questions may not lead the witness, and benchmarks from the material must carry figures the material has. Disagreements are shown next to the result.

## Limits

Reads text only - no PDF parsing (paste the text), no market data, no comparable companies, no background checks. First-look work, not investment advice. Long material loses its middle on paragraph boundaries, with a marker; the prescan still covers every line.