> ## Documentation Index > Fetch the complete documentation index at: https://docs.wpos.ai/llms.txt > Use this file to discover all available pages before exploring further. # Recipe: agent briefs > Writing build briefs that let an AI agent one-shot a site: the API-card-first rule, one-call verification, and fix-round budgets. A build brief is the document you hand an agent along with a target site. The difference between a brief that one-shots a four-page site and one that burns an hour is not the design direction; it is the operating rules. These are the ones that survived real batch runs. `eu-studio brief` generates a ready-to-run brief for the current environment; this page is the doctrine behind it. ## The API-card-first rule Make the agent's first act `exjsx api` (or reading the printed [API card](/ultra/api-card)), and explicitly forbid grepping framework source, conventions files, examples, or memory files. The card is the complete authoring API: intrinsics, sx keys, kit helpers, gotchas. Agents that explore source instead of reading the card spend their budget on archaeology and still author against internals that were never public surface. ## One-call verification Verification is where agent builds bleed time: each hand-written Playwright probe is a tool call plus a model round-trip. The brief pins the whole gate to single calls: ```text theme={null} eu-studio check --pages /,/about/,/pricing/,/contact/ \ --form /contact/ --accordion /pricing/ --burger / --nav /=about ``` One call, one browser session, the full structural matrix plus every interaction test, in about a minute. After a fix that touches one page, the brief says to re-check only that page. All measurement needs (screenshots, section rhythm, geometry, text ink, font paint, diff sheets, copy extraction) are pinned to `eu-studio measure `; the brief bans hand-written Playwright and PIL outright. ## Fix-round budgets Give a hard budget and an exit: **max two fix rounds, then report residuals instead of chasing.** Convergence loops without a budget do not converge; they oscillate around anti-aliasing noise. The brief's report format enforces the exit: page IDs and URLs, the final check JSON verbatim, the screenshot directory, and accepted deviations as a terse list. No prose narrative; the JSON is the record. ## Structural diversity guidance Left to defaults, models converge on the centered-column SaaS formula (hero, three feature cards, stats band, testimonial, CTA). If you want design, ban the default in the brief and make diversity testable: * At most one section per page may be a centered single-column block. * At least three distinct layout archetypes across the site: split-screen, asymmetric offset grid, sticky side rail, full-bleed color blocks with hard seams, oversized display type as layout, horizontal timeline, overlapping stacked cards, editorial multi-column text, diagonal section breaks, a table as a design object. * Vary section rhythm: at least one very tall and one very short section per page; backgrounds must alternate meaningfully. * Pages must not share one master section layout with swapped copy. * The type scale must take a position: huge display contrast somewhere, or a strict editorial scale, not the safe middle everywhere. ## Trust the gates Tell the agent explicitly: when `lint` prints a `verified:` fact or `check` passes a probe, that fact is **proven**. Re-verifying proven facts with hand-rolled scripts is the most common form of wasted budget. The positive-assertion design of the gates exists so the agent can build on them. ## Spend thinking where it pays The brief should direct effort: deep design thought at theme and authoring time (one-shot the pages well), quick mechanical execution everywhere else. Screenshots, crops, measurements, and redeploys are mechanical; deliberating over them is waste. Concretely: write the whole site before the first deploy, deploy with `--fast`, run one check call, judge shots at 1920 and 390, fix, targeted re-check. ## Environment rules that prevent derailments Small print that saves whole sessions, worth copying into any brief: * The target site is already running; do **not** provision anything or kill any process. * The server self-heals on crash; on a sudden request failure, wait about 15 seconds and retry before diagnosing. * Do not open pages in the Elementor editor on the playground. * A deploy 422 names the offending envelope; the server validator is authoritative, fix the named thing. * If CSS breaks: `eu-studio doctor` once; if the file store keeps flaking, `carry-css` once at the end. ## Related What the one-call gate actually asserts. The measured rebuild flow briefs point at.