Bill Belcamino. Start a conversation

The system  ·  a case study, written 2 October 2026

How a one-person company runs on a team of AI agents.

This is the operating system behind my ten websites: what it is made of, what the agents are allowed to do, what they are not, and what broke on the way. I am writing it down because the skill that matters now is not prompting a model. It is engineering the system around the model so that a business can trust it. Everything here is real, and the numbers are from the day this was written.

The parts What broke

Runs on

ClaudeClaude CodeWordPress REST APIGoogle DriveGA4 & Search ConsoleHostingerPlaywright

The operation in numbers

Counted on 2 October 2026, not estimated.

10live WordPress sites, all built and kept current through the system
41scheduled agent tasks on the master list, most firing up to three times a day so a missed run catches up the same day
8AI roles with written briefs: growth lead, SEO lead, data and citations builder, site builder, freshness verifier, analyst, marketing lead, QA verifier. See the org chart
60+numbered operating rules, each one written after something went wrong or Bill asked for it, with the date
825 vs 441sessions from AI assistants vs Google organic across the portfolio, last 28 days (GA4, 1 Oct 2026, my own visits removed)

Two more that matter: 21 playbooks (repeatable procedures an agent loads for a specific job, like an SEO audit or a site health check) and a 16,000-file shared folder that is the only place any of it lives.

The parts

Six things, in the order a change moves through them.

None of this is exotic. It is a shared folder, some text files and a set of rules, held together by the discipline to never let the agents work anywhere else.

01 · Memory

The Company Brain

One Google Drive folder is the whole company. Nine canonical files say what the business is, what it is trying to do, what is true right now, the hard limits, the house voice, who decides what, the commitments and the decisions with their reasons. Below that: a folder per site, a folder per growth project with its task list and log, the automation runbooks, the playbooks, a journal and an outputs tray.

Every agent, every scheduled run and every chat session opens by confirming it can reach that folder and reading the rules. If it cannot, it stops. Nothing is worked on from memory or from a scratch folder on a laptop, because that is how a company drifts out of sync with itself.

02 · Law

The operating rules

Sixty-odd numbered rules in one file that every agent reads first. They are short, dated and quote the moment they were made. A sample: never delete, archive. Cite everything or mark it unconfirmed. Fix every instance of a problem, not just the one in front of you. Never send, post or publish without the owner’s yes. When a tool stops working, stop and ask rather than work around it. Check whether the same task is already running before you start. Use the real clock, never a guessed date.

The rules are the product. The model changes every few months; the rules are what make the next model safe to hand the keys to.

03 · People

Eight roles, one org chart

An agent takes a role by loading its brief on top of the rules: growth lead, SEO lead specialist, data and citations builder, site builder, freshness verifier, analyst, marketing lead, QA verifier. A role reads only the project it is pointed at. The growth lead names the week’s three items per project on Monday, the lead role does them, the verifier checks the live result, and the analyst reports the numbers.

The roster is a table, and the org chart below is drawn straight from it. The whole team fits on one screen, which is more than most real teams can say.

04 · Time

Scheduled tasks

Forty-one tasks on the master list. The daily growth run at 5:45 in the morning works every growth project without being started. A site health and traffic digest at 6:48. Event verification for the San Francisco and Georgia sites on Monday and Thursday. A weekly vendor pricing check for the software database. A daily race-data check. A monthly fact audit. A weekly backup to Drive. A press desk that scans the morning’s news for stories our data adds to. Each one writes to a named log and knows how to catch up if the Mac was asleep.

05 · Hands

Deploy, with a way back

Agents change the sites over the WordPress REST API with a per-site application password: content, code snippets, media, settings. No browser, no login page. Before every code update the old version is saved to a backups folder; after it, the live page is fetched logged out and checked. A custom purge route clears the cache on every site. Every log entry ends with a rollback line that says exactly how to undo it.

New pages, affiliate links, money, sending and deleting are reserved for me. Everything else moves on its own, which is the point.

06 · Trust

QA, logs and the approval list

A separate verifier role checks every deploy on every site before the task is ticked. Each run ends by writing what changed, the IDs and the rollback into the Brain, then one list of anything that needs my yes, each item with a recommendation so I can answer in a word. A claim board records which task is running where, so two sessions never do the same work twice.

I read the approvals list. I do not read the logs unless something looks wrong. That is what trust looks like once it is earned.

The org chart

Who owns what, and who says yes.

Eight AI roles and one person. Every box has a written brief: what the role does every day, and the one line it never crosses. The chart is drawn from the same roster file the agents load before they start work.

Owner · final approver

Bill Belcamino

AI Operator & Systems Engineer

Owns: strategy, the rules, and every yes. Sending, posting, spending, affiliate links and deletes all come to me.

Daily touchpoint: one approvals list, each item with a recommendation, answered in a word.

Every role loads the same rulebook first: the Company Brain and 60+ numbered operating rules. A role that cannot reach it stops.

Runs the week

Growth Lead

  • Picks each project’s three priorities every Monday
  • Reports what shipped last week against each project’s targets
  • Sends me one numbered approvals list, with a recommendation on every item
  • Runs the week-4 and week-8 checkpoints: what to keep, what to drop
  • Keeps the shared task board in sync with every project
  • Reviews the mistake ledger every Monday: each miss must have a check that catches it next time

Scope · all 7 projects

Never: builds anything. Its job is to keep the plan moving.

Independent check

QA Verifier

  • Fetches every changed page live, logged out, with the cache bypassed
  • Runs a scripted gate on every deploy; nothing is called done until it passes
  • Checks the whole page, not just the change: its address, duplicates, structured data and how it looks on a phone
  • Checks no page lost a link it had and the home page and menus still load
  • Scans for red lines: no invented price, date, testimonial or health claim
  • Sends any failure back to the role that built it

Scope · every deploy, every site

Never: fixes what it checks. It did not build it; it proves it works.

Specialists · each leads its own projects, checked by QA

Search

SEO Lead

  • Owns every address: one page per topic, nothing competing with itself in search
  • Rewrites titles and intros to match what searchers already type
  • Puts the answer in the first two sentences
  • Batches a week of rewrites, before and after, for my yes
  • Crawls internal links, fixes every orphan page, re-crawls to prove zero

SEO lead on all 10 sites · project lead on 2

Never: adds a claim that is not already on the page with a source.

Citations

Data & Citations Builder

  • Builds data hubs, datasets, calculators and guides from primary sources
  • Ships every dataset with its method, sources, a CSV and a cite-this line
  • Lists what we couldn’t verify, in its own section
  • Keeps schema and llms.txt current so AI assistants can read and quote it
  • Drafts cite kits for journalists that I send myself

Lead on 3 projects

Never: publishes a figure that was not computed from the dataset.

Build

Site Builder

  • Builds comparisons, alternatives pages, guides and calculators from one spec per page
  • Stores every figure once, feeding the table, calculator, FAQ and schema
  • Runs a consistency check across old and new pages before publishing
  • Tests every calculator before it ships
  • Leads engineering on every site: snippets, plugins, settings, DNS and email changes follow one checklist, checked before and after

Build lead on 2 projects · engineering lead on all 10 sites

Never: invents a price it cannot verify on the vendor’s page that day.

Facts

Freshness Verifier

  • Checks dates, prices, addresses and hours against the organizer’s or vendor’s own page
  • Records the source URL and the date checked for every fact
  • Fixes every copy of a stale fact, not just the one it found
  • Runs the events desk: searches before it creates, so each event exists once
  • Clears pricing alerts within the week
  • Logs corrections publicly where the site has a corrections page

Scope · all 7 projects

Never: guesses. A fact it cannot verify is marked, not filled in.

Numbers

Analyst

  • Keeps each project’s scorecard: baseline, latest, target, direction
  • Runs a monthly AI-citation panel: 50 fixed prompts per site, cited or not
  • Pulls every number from Google Analytics and Search Console
  • Writes the daily traffic and site-health digest
  • Watches for quiet drops between audits: rich results, sitemaps, mail records

Scope · all 7 projects

Never: estimates. If Google Analytics or Search Console can’t show it, it isn’t reported.

Outreach New

Marketing Lead

  • Plans each week’s hero pages, channels and one experiment
  • Answers journalist and expert-quote requests while they are open
  • Drafts press, organizer, Reddit and LinkedIn outreach that gives the recipient something useful
  • Logs every contact in one CRM, so nobody gets pitched cold twice
  • Reviews what worked each week and picks the next test

Lead on 3 projects

Never: sends anything. It drafts; I send every pitch myself.

Solid line: reports toDashed line: checks or advises

Drawn from the team roster as of 2 October 2026. The hard line above every box: nothing is sent, posted, paid for, linked for money or deleted without my yes.

A normal Tuesday

Arizona time · from the master schedule

5:45Daily growth run picks the top open items on seven growth projects and does them as the owning role.
5:52(Mon and Thu) San Francisco events calendar: dates, prices and venues re-verified against the organizers.
6:48Site health and traffic digest from GA4 and Search Console lands on a private page, with event submissions and anything that needs a decision.
6:58Press desk scans the last 24 hours of SF news for stories our datasets add to and drafts the reporter emails. Drafts only; I send.
7:12Event submissions intake reads the site’s inbox, verifies, publishes what fits, logs the rest.
9:52Vendor pricing monitor re-reads every software vendor’s pricing page and flags the rows that changed.
10:13Big-event guides: anything six weeks out that does not have a guide yet gets one, built from cited sources.
10:38Answer posts: short pages for the questions people actually ask AI assistants, on the sites where that is worth doing.
1:45Growth run, second firing. Reads its own log, sees it already ran today, stops. Or finishes what was marked incomplete.
4:25Organizer replies: fixes organizer-confirmed facts, drafts replies, never sends.
4:50One email to me with the day’s submissions and pitches: accepted, rejected, open.
8:45Third growth-run firing, the catch-up for a Mac that was asleep at dawn.

What it has produced

Sites AI assistants cite, because the facts are verified.

The portfolio’s biggest traffic source in the last 28 days was not Google. It was AI assistants: 825 sessions against 441 from organic search.

That did not come from writing more. It came from building sites as data hubs: a San Francisco closure tracker where every entry is verified by address against the city business register, not by name; a software pricing database where every row carries the date an agent last checked the vendor’s own page; race pages where the data is re-checked daily; a supplement checker refreshed weekly against the FDA’s own list.

Each site also carries the plumbing a model needs to quote it safely: a Person and Organization graph, a machine-readable summary at /llms.txt, dated verification lines, the sources each article rests on listed in its schema, and answer pages for the questions people actually put to an assistant. Once a month, and the same day as any template or schema change, a script scores every site on whether an assistant can reach it, understand it, attribute it and trust it, and whatever slipped is fixed that day. On the research sites, articles also show a “Cite this page” line and the datasets come with CC BY 4.0 citation lines. The agents keep all of it current, which is the part a human never would.

And the smaller things that would never get done otherwise: a shared, generated seasonal game that ships to several sites from one spec with 240 automated tests; font and security fixes rolled out across the sites in one session; a weekly resume refresh; a Reddit scout that finds the threads where the sites could genuinely help.

projectsfClosure trackerVerified by address against the city register; five entries pulled before publishing for being wrong
helpcomparePricing databaseEvery price from the vendor’s own page, dated; weekly agent re-check
race.mxRace dataDaily check in season, weekly results and citation review, monthly fact audit
all 10Citation plumbingA cache purge route on every site; llms.txt, entity schema and security basics rolled out across the sites
sharedSpot & ScoreOne spec, one generator, 240 tests, deployed per site as a code snippet
nextAnswer postsWeekday agent run across the sites, reviewed after six weeks against the numbers

What broke, and the rule it produced

The system is the scar tissue.

Almost every rule exists because something went wrong first. These are the ones I would tell you about on the first call, because you will hit them too.

Work done outside the Brain was lost

Early on, agents kept context in a scratch folder on the laptop. It drifted out of sync with the Brain and nobody could tell which copy was true. Rule now: the Brain is the only workspace. Every run confirms it can reach the folder before doing anything, or stops.

Half-finished ships

A page set went live without its share images and its security block. Rule now: a build is not done until every item on its runbook is live, ticked one by one. Anything skipped is a blocker to raise, never a silent omission.

Working around a refusal

When a permission check or a tool refused a step, the agent’s instinct was to find another route. That is exactly how an agent ends up doing something nobody approved. Rule now: a step that stops working is paused and reported. Carry on with unrelated work; never route around the check.

Missed runs when the Mac sleeps

Tasks bound to my computer fired once, found it asleep, and the day was lost. Rule now: every such task fires up to three times a day and is idempotent. It reads its own log, stops if today is done, or finishes an entry marked incomplete.

Two sessions doing the same job

With fifteen sessions open at once, two of them started the same job. Rule now: a claim board. Before the first write, a session records what it is about to change; others reconcile instead of repeating.

Decisions buried in the middle of a report

I missed questions because they sat mid-paragraph. Rule now: every report ends with one numbered decision block, each item with a recommendation, and nothing after it. If there is nothing to decide, it says so.

The host rejected updates

The CDN in front of the sites returns a 403 for HTTP PUT. Updates now go as POST, reads carry a cache-buster, and the fact is written into the deploy script’s header so no agent rediscovers it.

Spending more than the work was worth

The most capable model is not the right default for a nightly run. Rule now: scheduled tasks and subagents run on the standard models; the expensive one is used only when I ask for it by name.

What this means for you

The same loop fits a company with staff.

Swap my ten sites for your recurring work: reporting, research, content, data upkeep, customer follow-up. The Brain, the rules, the roles and the approval list transfer directly; only the task list changes. If you want it built, send a short note and I will introduce you to the build team I trust. If you want the person who runs it on your team, send that note too.

bill@belcamino.com Connect on LinkedIn

    Or send it from here

    Same inbox, no middleman. I reply to everything I can genuinely help with.