Vibe Coding Meetup · Denver

Vibe coding an entire department

Not one tool. An AI agent that runs our Supplements department end to end: estimates, adjuster communication, approvals, completion certificates, invoicing, and follow ups.

Miguel Santos · Precision Exteriors Restoration

The department

Supplements: a whole job family, not a task

Insurance adjusters are not building code experts, so code required items are missing from most roof claims. This department finds what the scope missed, documents it, and gets it paid. Skilled, repetitive, high volume work.

Xactimate estimates Adjuster communication Job approval Completion certificates Invoicing Financial worksheets Public adjuster comms

Every one of these now has an AI playbook.

Architecture

One router, one playbook per function

The department is a thin router: identify which function the job is, open exactly one playbook file.
The agent never loads the whole department. Only what this job needs. Context stays small, behavior stays sharp.
Each playbook holds the hard won rules for its function, written once, corrected forever.
New functions become new playbooks. The department grows without retraining anything.

The work finds the agent

A self maintaining queue

A dashboard watches our job management system and detects which jobs need a supplement, no human triage.
That dashboard is the agent's work queue. It picks up the next job the same way a new hire would.
An hourly follow up engine watches every open conversation with an adjuster and drafts the next nudge when one goes quiet.

The hardest piece

The core tool has no way in

Estimates must be built in Xactimate, the industry's locked desktop app. No API.
Carrier scopes arrive as PDF only, never data.
DevTools protocol and Windows accessibility automation: both dead ends. The app renders raw pixels.

Rule I live by now: when a tactic fails twice, stop repeating it. Research the root cause instead.

So we gave it eyes and hands

Vision plus keyboard, verified

The agent OCR locates every field on screen, types the entry, then reads the value back to confirm it landed. A write you did not read back did not happen.
The full price catalog turned out to live on disk in the app's own backups. Extracted once, so any plain English line resolves to an exact code offline.
The thinking (parse the PDF, find the missing code items, justify them) lives outside the fragile UI layer and was shipping value before the hands existed.

Agents checking agents

Nothing ships on one opinion

Every finished batch gets three parallel reviewer agents with different lenses: process fidelity, correctness, tooling root cause.
They have caught real money errors a human skim would have missed, including work that silently came in below the carrier's own number.
Every catch becomes a new hard rule in the playbook. The department gets stricter with every batch.

Teaching it

I record my screen, it learns the SOP

One click starts a screen and mic recording while I work a job the manual way.
The agent transcribes it and diffs my real workflow against its assumptions, then corrects its own playbook.
This is how every function joined the department: show it once, correct it in writing, never teach it again.

Beyond estimates

The rest of the job, handled

Adjuster follow ups: an autonomous engine runs hourly and drafts the next email in any thread that stalls.
Completion certificates and invoicing: playbooks carry each job from approved work to money in the door.
Financial worksheets: recalculated when new paperwork lands.
Everything is draft only. A human approves before anything leaves the building.

Where this goes

Autonomy is a ramp, not a switch

Stage one: the agent drafts, a human reviews everything. That is today.
Stage two: draft for approval, review by exception.
Stage three: routine jobs run themselves, edge cases escalate to a person.
Each stage is earned by a track record, never assumed.

If you build one of these

Five things I would tell you

Automate the department, not the task. A router plus playbooks scales; a script does not.
Separate brain from hands. Ship the thinking before the automation.
Verify by read back. A write you did not read back did not happen.
Make tools honest. Success by default hides every failure that matters.
Let agents audit agents. Different lenses catch what you skim past.

Miguel Santos · Precision Exteriors Restoration