Agent Brief

An agent brief is the document you write once and paste as a system prompt to turn an AI into a goal-directed worker instead of a turn-based chatter. It answers four questions in strict order. What is the single sentence goal. What tools does the agent get access to. What rules constrain every decision. And when does it stop and hand you the finished deliverable. This brief is what separates designers who ship real work with agents from those still pasting chatbot output into Figma at 2 a.m. in 2026. It scales the tight prompt engineering discipline from 2024 into the agent loop of plan, act, observe, iterate so the thing actually finishes instead of asking for clarification every third step.

It is not a rambling paragraph typed into Claude Desktop. It is not a system instruction that only tells the model to act like a senior designer without nailing deliverables. It is not code, not a n8n workflow, and not the vague wish list you give a copilot hoping it reads your mind. It is not a one-shot prompt. Miss any of the four parts and the agent either loops forever burning tokens, improvises with stock slop, or stops early with half-finished garbage. A weak brief produces the average of everything the model has seen. A tight one produces work that looks like it came from your own desk.

The research-to-moodboard agent every Brainy studio now deploys on day one of a client project shows what a strong brief looks like in practice. Goal reads: from a discovery call transcript produce a moodboard in Figma plus a creative brief in Google Drive using our house template. Tools list is explicit: Google Drive MCP to read the transcript and write the final brief, web search, image fetch limited to editorial sources, and Figma MCP to populate a dedicated moodboard frame. Rules section gets ruthless. Pull images only from Are.na, It's Nice That, Brand New, museum archives, or studio portfolios. Never touch Shutterstock, Getty, or Unsplash. Caption every image with its source URL directly on the Figma frame. Write the brief in Brainy house voice: opinionated on craft, zero corporate filler, neutral on facts. If the transcript leaves a brand adjective unclear, flag it as needs confirmation instead of inventing something safe. Stop condition is binary: Figma frame must contain at least twelve captioned references, brief saved as PDF in the shared Drive folder, and both URLs posted back in the chat. Paste this into Claude Desktop with MCP connections wired, point it at a fresh transcript, and walk away. First run might pull one generic image or slip into LinkedIn voice. You review it like any junior deliverable, add one new rule, and rerun. By version four the agent ships moodboards a strategist can hand to a client with only minor notes.

The spec-to-handoff agent closes the design-to-dev black hole that still eats most teams. Goal: given a Figma file URL, produce a developer handoff doc in Notion containing component inventory, token audit, layout specs, and open questions. Tools: Figma MCP plus Notion write access. Rules: never invent a component name, map everything against the live components directory, flag every non-autolayout frame as a risk, include direct Figma node IDs, list any missing states without assuming interactions. Stop when the Notion page is published and its URL is returned. Run this after design sign-off and devs receive a 90-second snapshot that lists every token mismatch and risky layout instead of discovering them during implementation. The design QA agent takes it further. It hits a staging URL at 1440px, 768px, and 375px using Playwright, diffs the screenshots against the Figma file of record via Claude vision, then outputs a categorized Markdown report in the qa/reports folder. Blocking issues get annotated screenshots. Non-blocking gets tolerance thresholds of 2px. The brief forces the agent to ignore tiny pixel noise but scream when a button variant is wrong. These three briefs now run weekly inside Brainy studios and cut entire job functions from tedious execution.

Write an agent brief when the workflow repeats, crosses multiple tools, and has clear success criteria. Competitor audits that scrape Brand New and Figma files, weekly token hygiene checks against a design system repo, automated QA on every staging deploy, turning Notion meeting notes into formatted pitch decks. These are perfect because the loop can run unsupervised once the brief is tight. Never write one for taste decisions like whether a brand direction feels emotionally flat. Skip it for novel strategy, first-principles positioning, or anything requiring client trust. Ambiguous goals such as make the site better produce either infinite loops or generic output that wastes money. If you cannot state the stop condition in one concrete measurable sentence the agent has no idea when to quit and you will watch your token bill climb while it keeps iterating.

Version every brief like code. Store it in a text file, check it into git, and treat every failed run as a bug report that adds one sharper rule. After ten reviewed cycles the agent becomes reliable enough to run on schedule through n8n or Zapier while you sleep. The first three runs always expose where your thinking was sloppy. That exposure is the point.

A tight agent brief turns AI from a faster intern into a tireless junior who ships work worth reviewing instead of redoing.