Run AI agents
as a system.
I help small businesses put Claude Code and AI agents to work as a repeatable system, not a one-off trick. Everything is delivered async — every deliverable is a document you keep.
One workflow, written down (A$300) See services
Name the three tasks that eat the most time. You get back a single page. No charge, no data required.
Where this usually helps
AI drafts fast, and nothing checks at the same speed
Output volume went up. So did the time you spend reviewing it. The bottleneck simply moved.
Quality depends on who is paying attention that day
The same instruction gives different results. There is no written standard, so there is nothing to check against.
Repetitive work is eating the week
Templated documents, first replies, fixed-format reports. High volume, zero tolerance for error.
How I work — 4 phases + a quality gate
Not "let the AI do everything," but a design with an explicit human approval boundary. That is what keeps quality stable, and what keeps responsibility clear when something goes wrong.
Scoring is three-axis: process 30 + output quality 40 + factual grounding 30. Anything below the bar is sent back automatically and never reaches delivery.
Services
Claude Code Starter
A$2,000 / 2 weeks / 50% deposit
Pick one target workflow. You get instruction templates, a SKILL.md, a quality rubric, a setup guide, and a 10–15 min recorded screencast (screen only) — plus 14 days of Q&A after delivery.
AI Operations Advisory
from A$500/mo / cancel anytime
A monthly written report and unlimited chat Q&A. The standard tier adds one improvement build per month. Still no video calls.
Prices are indicative in AUD; invoices can be issued in AUD or JPY. Work starts once the 50% deposit clears.
Case studies
680 of 684 pages were worthless — auditing and rebuilding an AI-generated site
A machine audit of 684 AI-generated pages found more than 680 were thin auto-generated stubs. Reclaiming crawl budget with noindex, rebuilding the pages worth keeping, and turning the whole thing into a scoring gate that runs on every future piece.
Running four departments as one person — building a video production team out of AI agents
Video production split into planning, production, QA and analysis, with a pass gate between each department. The human approves and publishes, nothing else. Quality scores of 96/100, with nothing dependent on one person's attention.
Making AI Agents Model-Proof — a Handoff System That Survives Model Upgrades
Codifying an entire operating procedure into a portable SKILL.md — Hard Rules, per-task Definition of Done, a human-only scope, and an adversarial scoring rubric — so any model or agent resumes work with zero re-briefing.
Notes
Written for owners rather than engineers — how this works in practice, and where it goes wrong.
A checking standard for one AI task (free template)
A blank template for deciding, in writing, what AI may draft in a specific business task, what a person must keep, what counts as good enough to send, and what happens when it goes wrong. Copy it, fill it in, use it. No attribution required.
The completeness check for a first reply to a travel enquiry
A checklist you can use today, without buying anything. The expensive part of a first reply isn't writing it — it's discovering three days later that you can't quote because something wasn't asked.
Where AI can and cannot sit in a bookkeeping workflow
A boundary you can draw today, without buying anything. The rule is simple: nothing that produces a figure a client will act on. Everything around that figure is fair game.
Hirotoshi Yamaguchi
I'm a sole trader on the Gold Coast, Queensland (ABN 64 998 187 645), working with small businesses on AI operations. I've built and run a news publication, an iOS learning app and a content automation pipeline on my own — from concept through daily operation.
What I care about most is finishing what I said I'd finish, in the form I said it would take. I put that ahead of technical skill, because the thing that usually went wrong on someone's last project was that it got abandoned halfway, or it wasn't what they were told they'd get.
AI is not infallible; it fails sometimes. So I don't tell people it can be left to run on its own. I define who notices the failure and who fixes it, and I hand that boundary over along with the work. Everything happens asynchronously, and every decision stays on the record.
Why fully async
Skipping video calls isn't laziness, it's the design. Time zones stop mattering, every exchange stays on the record, and I don't take hours out of your week. Everything you need in order to decide arrives as a document.
- 1Get in touch — the A$300 written workflow is the easiest way in, but email works too.
- 2Scoping sheet (20–30 minutes to fill in) plus two or three rounds of Q&A by email. I reply within 48 hours (AEST).
- 3Quote — a fixed price and an explicit list of deliverables (definition of done).
- 450% deposit clears, then work starts.
- 5Delivery — instructions, SKILL.md, setup guide, recorded screencast. 14 days of Q&A follow.
Contact
The quickest way in is the free repetitive work review. Otherwise email contact@guardrail-ops.com — I reply within 48 hours (AEST, business days).