01
An AI cross-check silently dead for 23 records, failing identically every run for weeks, while the daily digest reported a green tick — because the success flag only counted write failures, not source failures.
AO / FIELD MANUAL 01PDF · 21 pages · 9 parts
THE AGENT OPERATOR'S MANUAL · v1.0 · 15 AUG 2026
Same model. Same prompt. Different result every time.
21 pages on making agent output repeatable: a memory layer that survives the context window, a skill library you reuse instead of re-explaining, verification gates that catch a confident “done” that isn’t, and a 30-day path to install all of it. Instant PDF download. $9 USD once.
Get the manual · $9 USDOne payment · keep it forever · free updates for every v1.x revision
THE DEFECT CLASS
An agent reporting success is not evidence of success.
The manual calls this false green, and it is the most valuable thing the system catches. Four real examples from the author's own production systems, all documented in Part 04:
02
An LLM response truncating mid-JSON. The token cap was too low, finish_reason was never checked, and three identical retries at temperature zero produced three identical failures.
03
A committed file that was the pre-fix build while the running copy was correct. A routine checkout would have silently re-enabled the bug.
04
A staleness watchdog with a 24-hour threshold, checked once daily — a guaranteed blind window. The thing it watched sat 23 hours stale and slipped under it, forever.
If you have accepted a diff because the agent said the tests passed, you have already paid for this manual once.
THE OPERATING LAYERS
The mechanism is not the secret. The assembly is the work.
01layer
MEMORY
A markdown vault as source of truth, plus a global contract file loaded into every session. Plain .md in a folder. No proprietary format, no lock-in.
02layer
SKILLS
Executable procedures with the failure modes already baked in, so the second time you do something you do it correctly by default.
03layer
EXECUTION
Claude Code and Codex CLI running the same contract side by side, so each can check the other’s work.
04layer
VERIFICATION
Deterministic gates plus a fresh-context critic. Never self-approve in the context that authored.
Built on Claude Code, Codex CLI, Obsidian, Playwright and git worktrees. All named, all public, most of them already installed on your machine. What you are buying is the order you assemble them in and the rules that keep them from lying to you.
INSIDE THE MANUAL
Nine parts that turn agent usage into a repeatable practice.
01 / The stack
A clear map of the operating layers, their jobs, and the real tradeoffs behind them.
02 / Operating contract
A compact contract and laziness ladder that keeps the work moving without pretending the system is automatic.
03 / Obsidian memory
A durable memory pattern for keeping decisions, context, and useful history available across work.
04 / Reusable skills
Turn repeated work into procedures you can invoke, inspect, and improve instead of rewriting from scratch.
05 / Verification gates
Catch false-green outputs with explicit checks that make done observable.
06 / Cheap infrastructure
Reason about inference and infrastructure with enough clarity to avoid paying for avoidable loops.
07 / Worked pipeline
Follow one task from brief to verified result, including the points where most agents quietly fail.
08 / 30-day path
A month-long implementation path that turns the ideas into operating habits.
09 / Templates
Copy-paste templates for contracts, skills, memory, briefs, and checks.
What this is not.
Not a prompt pack. There are no prompts to copy.
Not vendor-specific. The contract runs on any agent CLI.
Not a course, a community, or a subscription. One PDF, one payment.
Not for someone who has never run an agent CLI. It assumes you already have and are getting inconsistent results.
HONEST PROOF STATUS
What I can and cannot prove to you.
This launched on 15 August 2026 and has no customers yet, so there are no testimonials on this page and I am not going to manufacture any. What I can tell you is what the system is: 144 skills in the live library, 816 notes in the vault it reads from, and 227 client profiles enriched through it. Every figure in the manual is dated at the point it is stated, because inference prices and hosting prices move and undated numbers become quiet lies. If the manual is not worth $9 USD to you, ask and I will refund it.
Written by Rayyan Mohsin. This is the operating doc I actually work from, cleaned up for someone who doesn't have my context.
The complete manual
The Agent Operator's Manual
21 pages. 9 parts. A practical agent-first operating system in a PDF you can keep beside your work.
What happens when you click.
- You go to Stripe's secure checkout — card, Apple Pay or Google Pay.
- You come straight back to a download page. The PDF is there immediately.
- Your receipt lands in your inbox.
Nothing recurring. Nothing to cancel.
One-time access$9 USD
- ✓ The complete 21-page PDF
- ✓ All 9 parts including the copy-paste appendix
- ✓ The 30-day implementation path
- ✓ Every future v1.x revision, free
Payment is handled by Stripe. Your card details never touch this site.
The charge on your statement will read BUILD & RELEASE AI.
If it's not useful, I'll refund you.
Email rayyan@clientacquisition.io within 30 days with your order number and I'll send the $9 back. No form, no questions, no "what did you try first." I reply within one business day.
I'd rather refund you than have you dispute a $9 charge with your bank. So please ask me first — you'll get your money either way, and it's faster.
FAQ
Questions worth answering.
Isn't all of this on a blog somewhere for free?
Parts of it, scattered, written by people describing a system rather than running one. What you are paying for is the sequence and the failure modes — the things that only show up after the approach has broken in production.
Which tools does it assume?
An agent CLI (Claude Code or Codex), a markdown vault (Obsidian or any folder of .md), git, and tmux. Everything else is optional and named where it appears.
Will it be obsolete in three months?
The tools will move; the contract and the gates will not. Every number in the manual is dated at the point it is stated so you can tell what has aged. You get every v1.x revision free.
21 pages seems short.
It is short on purpose. There are no filler chapters and nothing is padded to justify a price. If it could be a paragraph, it is a paragraph.
Do I need to pay for Claude or Codex on top?
Yes, you need your own agent subscription. The manual is about how to run it, not a replacement for it.
Is this for beginners?
No. It assumes you already use an agent CLI and are frustrated by inconsistent output.
What exactly do I get?
One PDF, 21 pages, 9 parts, downloaded immediately after payment from a page only you can reach. Nothing is emailed to you.
How do I get a refund?
Email rayyan@clientacquisition.io within 30 days with your order number. No form and no questions — I would genuinely rather refund you than have you dispute a $9 charge with your bank.