by Chris Hunter

Share

A stack of blank audit questionnaires with empty checkboxes on a dark desk beside a red pen and a brass lamp, with the title "Build log, Nov 7-Nov 11: 36 old audits said "No" to questions nobody answered"

TL;DR

  • 36 old audits counted unanswered questions as a real “No.” They now show as not confirmed, and staff can re-run any audit on honest inputs.
  • A web page was too long for my AI to read. Most of it was page-builder styling, and the fix cut that audit from about 203K tokens to about 30K.
  • Friday (my chief of staff) can answer me inside the shared brain now, and every reply is scanned for leaks first.

This one covers November 7 through 11. Most of the fixes this week started the same way: measure the thing first, then decide what to change. More than once the number pointed somewhere I didn’t expect.

Friday (my AI chief of staff) - My agents can answer me now

Last week I wrote that the two-way version of my shared brain was designed but not built. It’s built. Plan to live took one day, Saturday, November 7.

Until now the brain was one-way. I dropped notes in, and my agents read them. Friday (my chief of staff) now has a channel back to me. A note I leave as a request waits under a stable ID until it’s answered, so it can’t age out or get lost. Friday’s replies land in their own folder, each one scanned for leaks before it’s written. When I mark a reply “got it,” it clears.

A new “waiting on Chris” page lists every decision that needs me, with the options and a recommendation. Open requests show at the start of every session and in the end-of-day check.

Before this, a decision I’d made once could come back, and a request I sent from my phone could sit unseen. Edited, duplicate and missing requests now get reconciled instead of dropped.

4,382 tests passed at ship, and the review was a pass. Two gaps: replies go to one recipient for now, and one leak-checker test file still hangs on this machine’s older Python.

Capture is only half of a memory. If your team can read what you said but can’t answer you in the same place, you end up routing every reply yourself.

AI Marketing Audit - The page looked normal, so I measured it

My AI Marketing Audit reads a business’s website and hands the text to the AI to score. On November 9 one audit failed: “prompt is too long: 203,253 tokens,” against a limit of 200,000.

The page looked ordinary. Measured, 376K of its 535K characters were style and class attributes from the page builder. The visible text was 13K.

A cleaner now strips that markup and the inline graphics. It keeps the tracking tags, because once the scripts are gone they’re the only proof a site tracks anything. Then a prompt budget works out how much room is left and cuts the page on a clean boundary, with a note telling the AI that a cut page isn’t proof something is missing.

Review caught two mistakes in my first cut. It dropped the tracking tags, and a self-closing graphic tag swallowed everything after it. Both are fixed.

The same audit went from about 203K tokens to about 30K. 43 new checks use the real 577K page as a test file, and the suite passed 61 of 61. It was live about 54 minutes after the failure. Last week’s stuck-audit guard marked it failed instead of leaving it spinning, which is how I saw it at all.

Still open: an odd closing tag on an inline graphic can still eat content up to the next one (logged, low severity). Nobody has checked the results page with a staff login yet.

An AI step that works on 99 pages and fails on the 100th is a reliability bug, not a fluke. Measure what you fed it before you touch the prompt.

AI Marketing Department - Stop patching edges

My marketing engine shipped four releases on November 9 and 10.

The big one is about photos. Every photo carries hidden location data from the phone that took it. Every upload path now strips all of it and checks the final file. If it can’t prove the strip worked, nothing uploads. That release took 3 review rounds.

The smaller fixes include a retired image model replaced as the default and a reading rule that fails closed, so a bad read can’t wipe a data bank. The SEO audit also treats schema features the search engine has retired as information, not a finding, and flags dead business-profile buttons and links.

The lesson came from output getting cut off when a script exits. The first fix added a wait-for-it step in 237 places. Review showed one missed spot made the guard fail open. I threw it out for one rule: every entry script sets blocking output first, and shared code never does. It’s proven on Windows in the build pipeline.

The November 9 bug sweep confirmed 3 of 18 raw findings (2 high, 1 medium), fixed in these releases. 9 commits in all.

Still open: the strip needs a free photo tool on every machine, and the Windows machine needs it next. The readiness check now says when it’s missing.

When review keeps finding edge cases, stop patching the edges. Change the mechanism so the edge can’t exist.

AI Marketing Audit - The report became a plan, and the meter came first

The audit got three new pieces on November 10.

The cost guard came first. Before wiring a paid search-data service into audits, I measured how it was already used. Audits never called it. The only caller was one admin button, with one recorded call ever. So the guard went in log-only. It estimates every call before it’s made and writes each one to a ledger, with caps per run and per day. It shipped with 40 checks and the suite at 62 of 62, and it passed review.

Next, the report became a plan. Each finding gets a severity, a one-line fix, an owner and an evidence label, from measured down to no data. Findings sort into This week, Next 30 days and Days 30 to 90, and the score only counts parts that have data. It’s live but switched off, staff-only, until my team and I review it.

Last, a competitor panel: you against your three nearest listings on the map. It caches results and won’t make a paid call that would run over its estimate. It shipped with 74 checks and the suite at 67 of 67. Review found 4 low issues, all fixed, and it’s on for staff now.

Two things broke. The “find your rank” button was broken on production for lack of a place ID column, which exists now. And the first panel deploy aborted before changing any code, because my own check miscounted the columns. Fixed and re-run.

Still open: the code that reads ranking results hasn’t seen a real map reply.

Find out what a paid tool will cost you before you wire it in. Measuring first is why this guard could ship without blocking a thing.

AI Video Editor - It won’t render a cutaway I haven’t approved

The AI video editor learned to cover talking-head video with B-roll and a music bed. A demo on one real episode came out at 84 seconds: 8 cutaways (3 captures of public web pages, 5 text cards) and a bed that ducks under the voice.

By the end of the day the demo was a build-out. A cue drafter reads the transcript and proposes cutaways. I refine and approve them, and the renderer refuses to run on a cue list I haven’t approved. I approved 18 for the demo. A music step picks a track and checks it for vocals with a local speech tool, and every render logs the license of the track it used.

The first pick got cut short by vocals. The picker now prefers a track that covers the whole hook, and 13 tracks with singing went into a “vocals OK only” pile, leaving 9 clean instrumentals. The quality check also flagged captions 0.19 seconds off, most likely the music confusing the speech detector.

Phases 1 and 2 of a 5-phase plan took one afternoon, 4:25 to 5:35 PM CT: 6 feature commits, 477 tests after phase 1. It’s a demo branch, not merged or pushed. Next is wiring it into long and short videos and running the caption check on the voice track alone.

Generated media needs a human gate you can’t forget. Approval has to be a state the machine checks, not a note on a card.

AI Marketing Audit - 36 old audits said “No” to questions nobody answered

Last week’s post was about a blank answer that counted as a gap. This is its older cousin. Before October 31, the staff form pre-selected “No,” so every question nobody touched got saved as a real “No.”

On November 10 I traced how far it reached: 36 staff audits from January to October, with those false answers feeding the reports and the AI’s gap list.

The fix went live that morning. Those answers now sit in their own box, “Not confirmed (answer not reliably recorded before 10/31),” with a mark in the PDF. They never count as a gap, and scores don’t change. Old audits carry a note above the AI recommendations until someone re-runs them.

A new staff “Re-run analysis” button redoes an audit on the honest inputs, with a per-audit lock and a 10-minute cooldown. A bad reply changes nothing. It updates the score in the CRM only where the audit was already sent, and never over a newer assessment.

A control re-run on the test audit moved the score by 0.2 and turned 25 answers into “unconfirmed” instead of gaps. Both pieces passed review after a fix round. The rollback was proven to restore the exact row counts before the final apply.

The first deploy of the button rolled itself back on my own smoke test. The test was wrong, not the feature, and the rollback did its job.

A default that looks like an answer is the most expensive kind of bug. You find out when someone shows a customer a gap that never existed.

Friday - One home for every fact

Friday diagnosed this one before building anything. One fact was getting written four times. A decision batch lived in a scratch folder and was lost on restart. Timers inside a session died with it, and the same procedures kept getting rediscovered.

Friday now keeps one shared event log, a line per shipped, deployed, decided or blocked event, written once and read everywhere. A work registry tells a session when another one is already in the same project. Recipes turn a procedure that took a dozen tool calls to rediscover into one reusable file. There are 27 so far. On its first run, a handoff archiver moved 67 old session notes out of Friday’s live folder.

The always-loaded rules file went from 23,357 bytes to 10,173, a 56% cut and the second trim after the November 3 token diet. Eight tool-specific rule sets moved into skills that load only when needed.

Still open: the Windows machine needs the installer re-run to get the new skills, and the check that reconciles my task board against the event log hasn’t run live.

Shared state beats a shared inbox. Give every fact one home, and nobody has to reconcile four copies.

AI Builders Labs - A re-import that could knock a lead backward

AI Builders Labs’ lead tooling got its weekly bug sweep on November 7: 6 confirmed findings (1 high, 3 medium, 2 low), all fixed the same day.

The high one: re-importing a list could knock a lead that had already moved forward back a stage. A lead’s status now never moves backward on a re-import. The other five were smaller, one of them a stale memory note. 135 tests pass, though there’s no live data to check them against yet.

The sweep also produced a rule: contact details never live in tracked drafts. The contacts file is excluded from version control everywhere, and the client template says so.

A tool that quietly resets your pipeline costs you follow-ups you never knew you owed.

Need help building something like this?

Everything I write about here runs in my own business first. If the AI job landed on your plate, start with an assessment: where AI will make or save you the most money, and what it'd take to run it right.

About the Author: Chris Hunter

I'm Chris Hunter, a fractional Chief AI Officer. I've spent 23 years in marketing and co-founded Marketing Heroes, an agency for roofers and home-service companies. I've spent the last four years building AI into the work, and today an AI system I built runs a big piece of it: an AI sales rep, an executive assistant, and a marketing department that ships content and runs ads every week. One operator and those agents now cover two of our six production seats, and the agents cost about $460 a month to run. Now I do the same for other companies, so the founder doesn't have to be the head of AI on top of running everything else. Before any of this I was an Air Force crew chief, working on the first stealth fighter under zero-defect standards. That's still how I build. If it breaks in front of a customer, it wasn't done.

Share

Related Posts