Content · build guide
AI Brief Machine
From one objective, keyword, or product to a full creative brief in under five minutes
What it does
Give the AI Brief Machine a single input — a campaign objective, a keyword, a product — and its multi-step prompt chain returns a structured creative brief in under five minutes. The chain runs in stages: audience definition, angle, format, messaging hierarchy, and distribution notes, with each stage's output feeding the next so the brief stays internally consistent instead of reading like five disconnected paragraphs.
The chain itself is the compounding asset. Every run surfaces a weak prompt or a missing brand rule; you fix it once and every future brief improves. Over months you build a brief library your writers trust and a prompt chain tuned to your brand — not a blank chat window someone re-improvises each time.
This version runs entirely on a saved prompt chain, a Google Docs template, and an intake sheet — no API connections and no build tooling, matching the automation's core dependency: one input per run. It suits a team that wants consistent briefs this week and is fine running the chain by hand for each request.
This version wraps the same prompt chain in an agent: intake triggers a run automatically, the agent pulls read-scoped brand and performance context over MCP, drafts the brief, and parks it at a human gate before anything reaches production. It suits teams producing enough briefs per week that hand-running the chain has become the bottleneck.
Before you start
Prerequisites.
Path A
Build it with manual data extracts.
This version runs entirely on a saved prompt chain, a Google Docs template, and an intake sheet — no API connections and no build tooling, matching the automation's core dependency: one input per run. It suits a team that wants consistent briefs this week and is fine running the chain by hand for each request.
Lock the brief template first
Create a Google Docs template with the five sections in fixed order: audience definition, angle, format, messaging hierarchy, distribution notes. Add a header block for the input, requester, date, and the prompt-chain revision that produced it. The chain writes into this shape, so agree on it with your writers before you build anything.
Build the prompt chain in Claude
Write one saved prompt per stage, where each stage takes the previous stage's output as its input — audience feeds angle, angle feeds format, and so on. Store the chain in a shared Claude Project with your brand context (tone rules, banned claims, segments) loaded as project knowledge. Record the chain in a change log from day one so every later revision is traceable.
Set up an intake sheet
Make a Google Sheet with columns for the input (objective, keyword, or product), audience notes, deadline, requester, and status. Optionally paste supporting rows from a search term or GA4 landing-page export into a context column — it sharpens the audience and angle stages, but the chain runs fine without it.
Run the chain per request
Take the next intake row, run stage one, paste its output into stage two, and continue through all five. Paste the final output into the template. A run takes under five minutes; resist editing mid-chain — note weaknesses instead so they get fixed in the prompts.
Review, file, and log
The owner checks the draft against brand rules, makes corrections in the doc, saves it to the brief library, and marks the intake row done. Log any correction that reveals a prompt gap — those notes are the raw material for the next chain revision.
Iterate the chain monthly
Once a month, review the logged corrections, update the weakest prompts, and note the revision in the change log. Keep prior prompt wording so you can roll back if a revision makes briefs worse.
Cadence: On demand per campaign request — a run takes under five minutes plus a short review; batch the intake queue weekly if requests pile up, and revise the chain monthly.
Path B
Integrate it with MCP connections.
This version wraps the same prompt chain in an agent: intake triggers a run automatically, the agent pulls read-scoped brand and performance context over MCP, drafts the brief, and parks it at a human gate before anything reaches production. It suits teams producing enough briefs per week that hand-running the chain has become the bottleneck.
Encode the chain as an agent workflow
Rebuild each stage as a step with structured output — named fields for audience, angle, format, messaging hierarchy, and distribution notes — in n8n, Make, or a Claude agent workflow. Store the prompts in version control or a versioned doc so every change is reviewable and reversible.
Wire intake as the trigger
A new row in the intake sheet or a new ticket in your project tool starts a run — no one pastes anything by hand. Reject malformed intakes back to the requester with a note rather than letting the agent guess at missing inputs.
Connect read-scoped context
Give the agent MCP connections to your brand guidelines folder and, optionally, performance data such as GA4 or Google Ads exports in BigQuery. All connections are read-only and scoped to specific folders and datasets; write access exists only for the briefs output folder.
Let the agent draft and self-check
The agent runs the five stages, then validates the draft against your brand exclusion list — banned claims, restricted audiences, off-limits comparisons. Sections it can't ground in the provided context get flagged for the reviewer instead of filled with plausible-sounding filler.
Hold every brief at the human gate
The finished draft lands in Google Docs and a Slack notification goes to the chain owner, showing the input and which sources the agent used. Only approved briefs move to the writers' queue. The agent recommends; it never publishes.
Keep the audit trail per run
Each run writes input, prompt revision, sources read, output, approver, and timestamp to a log sheet or BigQuery table. When a brief goes wrong, the log tells you exactly which revision of the chain produced it — and rolling back is one change.
Tools & guardrails
Run it safely.
- Every brief passes a named human approver before it reaches writers or production — the agent drafts, it never publishes
- Agent access is read-only and scoped: brand-docs folder and performance datasets in, one briefs folder out, nothing else
- The prompt chain is versioned and every run logs input, prompt revision, output, and approver — any bad brief traces to a revision and rolls back in one change
- A brand exclusion list is checked on every run; flagged briefs route to the owner instead of the queue
Want new build guides as they ship?
Optional — every guide on this site is free either way. Leave your details and we'll email you when new automations join the library, nothing else.
Prefer it built for you?
We stand this automation up as a governed sprint on the stack you already run — scoped, built, and handed over with the guardrails configured. No new platform, no migration.