Everyone says they use AI now. Here is what I actually built, what it runs on, what it changed, and what breaks.
ExtendScript (JSX) · Illustrator · Roland VersaWorks · Graphtec
Print-production prep
Problem Every wrap, decal, and sign job needed three hand-built files — print, plotter, master — plus a fiddly 20× scale-up and RGB→CMYK conversion before it could hit the Roland XP-640. Slow, with a standing chance of sending the wrong file to the plotter.
Built DecalPrep_Automate.jsx turns a finished design into _Print.ai, _Plotter.ai, and _Master.ai in one run. AutoSpa_ScaleUp_CMYK_Preflight.jsx scales art and artboards 20×, converts color, runs a preflight report, and saves the full-scale file. Design → script → VersaWorks RIP → print → laminate → Graphtec contour-cut on registration marks.
Result Hand-prep between design and print is gone, and so is the misfiled-plotter-file failure mode.
What breaks Illustrator's canvas limit forces the design-small-scale-up workaround; and the preflight report is there because color conversion fails quietly.
AppleScript · Illustrator · Python · ReportLab
Batch export & document generation
Problem Six in-shop TV display boards needed re-exporting as a batch; tiered proposals were rebuilt by hand per prospect.
Built AppleScript that drives Illustrator externally for batch saves and exports. A Python/ReportLab generator that produces branded, tiered proposal PDFs on demand — the fire-department fleet ceramic proposal shipped this way. Plus browser-based pricing calculators and a referral-code generator with QR output that feeds GoHighLevel workflows.
Result Board exports are a script run. Proposals are generated, not assembled.
What breaks Scripting Illustrator from outside is flaky: you have to bring the app to the front first, and you poll the filesystem for the saved file instead of trusting the save call to return on time.
ExtendScript · Brand design system
Creative variants from the brand rules
Problem Producing every size and format of one creative is mechanical work that eats a designer's day.
Built Scripted batch generation against the master brand system — palette (black / #EF5323 / white / grey), the three-font stack, and layout specs applied programmatically, so variants stay on-spec by construction. The pattern is in the code below.
Result Creative iteration stopped being rate-limited by production time.
GoHighLevel · Zapier · Webhooks
Lifecycle & CRM automation
Problem Leads arrived through five channels and got followed up inconsistently — or not at all.
Built A routing and follow-up layer: inbound from web, calls, and social normalized into one pipeline, with triggered sequences through booking and post-service follow-up, tied to the same brand system as the ads.
Result Follow-up happens whether or not anyone remembers to do it. The transactional message and the ad now read like the same company.
Claude · Scheduled task · Browser automation
The agent that writes and posts
Problem Consistent publishing fails for the ordinary reason — it competes with billable work and loses.
Built A Claude agent on a Tuesday/Thursday schedule. It checks my recent LinkedIn activity so it does not repeat a topic, drafts against a fixed brief and voice guardrails, posts to LinkedIn through browser automation, and logs what it covered for the next run. The essays on this site are AI-drafted and labeled the same way.
Result Runs Tuesdays and Thursdays without me; first scheduled run September 15, 2026. Every AI-written piece is labeled as such.
What breaks No human review before it posts — a deliberate trade, and why the brief is narrow and the label is mandatory. And the site is not in the loop: the agent posts to LinkedIn on its own, but essays reach this site by hand. Closing that gap is next.
Claude · Scheduled task · Gmail · Calendar · Drive
Opportunity intake pipeline
Problem Inbound opportunities need research and a fast, specific reply, and the research is the slow part.
Built A scheduled Claude job that reads inbound mail, filters real opportunities from job-board noise, researches the company and role, writes a tailored cover letter and call script into a per-company folder, labels the thread so it is never processed twice, and stages a reply as a Gmail draft — never auto-sent, because that is where a human belongs.
Result Research-to-reply collapsed from hours to minutes, with the judgment call still mine.
What breaks Filtering is the weak point: a false positive wastes a research run, a false negative loses a lead. It is also paused whenever the desktop it depends on is offline — a dependency I would remove in the next version.
The pattern, in code
ExtendScript
// Batch-render one creative across the spec'd size set.
// Layout rules live in the brand system, not in my head.
function renderVariants(doc, specs, outDir) {
for (var i = 0; i < specs.length; i++) {
var s = specs[i];
doc.documentPreferences.pageWidth = s.w;
doc.documentPreferences.pageHeight = s.h;
applyGrid(doc, s.margin, s.columns);
fitType(doc, s.scale); // type ramp, not free-hand resizing
enforcePalette(doc, BRAND.palette);
var out = new File(outDir + "/" + s.name + ".pdf");
doc.exportFile(ExportFormat.PDF_TYPE, out, false);
}
}
Simplified from production — the pattern, not the whole file. The point is that the brand system is executable: specs are data, and the layout rules are enforced by the script rather than by someone remembering them at 6pm.
