Build log
"AI agent prompts" usually means a one-liner like "act as an SEO expert". That got me nowhere. What worked better was writing each agent a proper brief: one job, the input files, numbered steps, an output file, a NEVER list, and the same approval gates pasted at the bottom. These are the five briefs behind my local SEO agency test. So far: 154 businesses audited, 10 cold emails sent, no replies recorded and $0 revenue.
The shape of every brief
ROLE: You are the [role]. INPUT: [files it reads] TASK: 1. ... 2. ... 3. ... OUTPUT: [file it writes, with source URLs and today's date] NEVER: [things it must not do] [PASTE APPROVAL GATES HERE]
The output file matters most. If an agent only replies in chat, the next agent can't pick up its work, and I can't check it.
1. Market Researcher
Finds and scores markets and prospects. Its brief says to tear down the top organic results, reject combos owned by national franchises, label every keyword estimate as qualitative and never invent search volumes, and always pair the city with the state (Medford OR is not Medford WI).
2. Site Builder
Turns a filled site brief into a static site with build_site.py, records every fact in a sources file, fixes pages the script flags as thin, and zips the site. It is told not to deploy. Its NEVER line: fake reviews, invented addresses, license numbers, guarantees, or prices presented as quotes.
3. Outreach & Compliance
Audits homepages, generates mini-audit PDFs and drafts emails, only to addresses the business publishes. In my run, 92 of the 154 audited businesses had a published email, and the agent generated 78 mini-audits.
4. QA / Verifier
This is the one I'd keep if I could only keep one. Before anything is sent, it re-checks the claims against the live site. When it re-checked 30 audits, 10 PDFs cited an issue the business had already fixed, and 2 were skipped. Without that step, a third of that batch would have opened with a wrong claim.
5. Chief of Staff
Keeps the queue, writes the daily debrief and puts every decision I need to make under "Decisions needed". On a small team it can also do the research and QA.
The approval gates
Every brief ends with the same gates: never spend, never send, never submit, never publish. Draft, queue and ask instead. Every price or statistic needs a source URL and a checked date. The 10 emails I sent all went out by hand after I approved them, on 10 Oct 2026. A further 18 are drafted and not sent.
Where this leaves me
- Businesses audited: 154 (97 US, 57 AU, 21 cities)
- Mini-audits: 78. Emails sent: 10. White-label pitches drafted: 50
- Replies: not checked yet. Clients: 0. Revenue: $0.
Good briefs make agent work checkable. They don't make anyone reply.