Privacy rules when agents build your personal site

In this note
  1. Building a stage set, not doxxing my living room
  2. Nodding to 2027 work without leaking a syllable
  3. Four build scripts that don't rely on polite prompts
  4. Locked down before build number one

To build Home Port, I handed cloud coding sessions my resumes, career notes, and a few reference photos from home.

That's the weird tension at the center of building a personal site with agents. A personal site only has a pulse if it feels specific and lived-in rather than sanded down into LinkedIn cardboard. At the same time, an agentic build means multiple autonomous sessions reading and writing code, transcripts, image pipelines, and static files destined for the public web.

If your privacy plan is "I'll just read every diff carefully before we launch", you don't actually have a privacy plan. Parallel sessions generate more changes in an afternoon than you can eyeball on a phone in a week, and at 1:00 AM tired eyes miss things.

So we treated privacy the exact same way we treated byte budgets and contrast ratios: write down bright-line rules up front, and enforce the ones that really matter with scripts that break the build.

Building a stage set, not doxxing my living room

Early on, I shared a handful of private reference photos with my lead session so it could calibrate how warm the afternoon light, wood, and rug textures should feel. The lazy path would have been tossing those JPEGs into a repo folder where every worker session could grab them.

Instead, those files stayed strictly on that single container's ephemeral disk: never committed to Git, never sampled as textures, and never placed in a render. That rule carried a real operational cost. Because spinning up a fresh cloud machine starts with a blank disk, any visual check against real-world warmth had to stay inside that original lead session, which meant when we switched models later to save quota, we swapped the model in place rather than migrating containers.

The room itself follows the exact same boundary. On the questions page I tapped the rule that governed every prop: "Kindred pieces only, nothing that could identify the house". The loft is an invented Miami stage set, not my floor plan. Every wood, plaster, and fabric texture comes from procedural Blender shaders or open CC0 libraries fetched by script. Out the window, my employer's building across PortMiami is pure architectural massing with zero corporate signage, every berthed cruise ship is clean unmarked white, and nothing anywhere in the repo narrows my location tighter than Miami.

Nodding to 2027 work without leaking a syllable

Anyone leading multi-year work runs into the same portfolio headache: the work stretching you most right now is usually the work you can't talk about publicly yet.

In the loft, that 2027 work sits across the channel as a ship wrapped in warm shipyard sailcloth stenciled COMING 2027. When the lead asked early on whether internal names or feature details could appear anywhere, my answer was blunt: "We don't need to use the name just yet, only that it is coming in 2027."

My favorite guardrail from that conversation was a tiny prop rule inside the design doc: "If a prop mentions it, the prop is wrong." No sticky note on the whiteboard, no terminal tab on the desk monitor, and no clever Easter egg on a coffee-table book jacket was allowed to hint at internal names or unannounced mechanics. Left to their own creative instincts, builders love sneaking realistic details onto desk screens, so spelling out that fence early saved us from our own cleverness.

Four build scripts that don't rely on polite prompts

Markdown instructions tell well-meaning agents where the lines sit, and a dedicated privacy-reviewer persona checked uncommitted docs before they hit Git. Even so, prompts are just advice. Real peace of mind came from four deterministic checks in the build pipeline:

  1. Pixel-only media exports: Every rendered AVIF, JPEG, and video clip strips all EXIF tags, encoder comments, and file metadata on export. On my Mac, the upload script scans outgoing manifests before pushing and aborts immediately if a local file path or macOS username shows up anywhere:
golden send: STOP. <file> holds a path or a user name of this machine. Nothing was sent.
  1. Zero raw contact details in public files: Early on, I asked for a contact form so my personal email address wouldn't sit out in the HTML for scrapers, then realized later that my email was still sitting inside my resume PDFs. Now a build-time scanner reads every text file a build serves and fails the build instantly if a phone number or email address ever slips in. The PDFs left the static site and are served only after a human check, and direct messages stay gated behind human verification on /api/contact, so visiting hiring-manager agents can freely read /llms.txt and resume.md without being able to spam my inbox.
  2. An out-of-repo blocklist: A private word list of internal codenames and sensitive terms lives outside the Git repository itself (because committing a secret blocklist to Git leaks the exact words you're hiding). Every release build scans all output text against that external file.
  3. Explicit hold gates on draft copy: Whenever I pasted in a fresh brainstorm doc from another chat that accidentally reintroduced something I'd ruled off the public site earlier, like internal team-size numbers, the copy script automatically flagged those specific lines on hold and blocked public release builds until I explicitly ruled keep or cut.

Locked down before build number one

Most side projects deploy wide open on a random .vercel.app preview URL from commit one and bolt on privacy settings at the end if someone remembers.

Here, authentication went onto the Vercel project before the very first deployment ran, covering staging previews, the private phone review tools, and production alike until launch day. The one time I wanted a temporary link so a second AI assistant could Tear Down our layout and positioning, the lead generated a single 23-hour expiring token, listed plainly which unreviewed draft lines were visible on that build, and let me make the call with eyes open.

Once you wire deterministic privacy gates into one personal repo, leaky defaults start bugging you everywhere else. In CrashCalc, a CI test enforces zero network imports inside the shared forensic physics core so case numbers physically cannot leave an investigator's phone. In Messina OS, cloud sessions started from my phone never touch a single home LAN key or token. And in our wedding camera, GPS EXIF coordinates get stripped inside the guest's browser before a photo ever uploads, while next-morning gallery reminders use a one-tap .ics calendar file instead of harvesting guests' phone numbers.

Good fences don't make personal software feel stiff or guarded. They're the whole reason you can invite strangers in to poke around without worrying about what got left on the desk.