The stuff I'd tell you over coffee — short, useful, no fluff. Playbooks, notes and clips on building AI content systems that stay honest.
The House Rules
If you can't write the number that means it worked, you're not building — you're shopping.
1 month ago
The House Rules
If it speaks in your name, it is you speaking. The disclaimer is a note to yourself.
1 month ago
The 2-Minute Test1 month ago
One reader who checks
A AU$440,000 consultancy report cleared every internal review with a fabricated court quote in it. Then one academic actually opened the sources, and the firm refunded the fee. Your reputation is one diligent reader away from the truth, and that reader always shows up eventually. Today: be that reader for your own last deliverable. Open one source. See if it says what you claimed it says.
The Blueprint1 month ago
The receipt gate
Bind every factual claim to a row that points at evidence — a log line, a run id, a URL. Publishing calls a preflight that resolves the reference and refuses if it is missing, unresolvable, or unverified by a human. Verification is a separate act by a named person, never the author, never the model. Nothing with an unverified claim can reach a reader, because the code will not emit it.
Get the structure
The Ledger1 month ago95%of enterprise GenAI pilots returned zero measurable P&L impact
MIT's Project NANDA reviewed over 300 disclosed deployments. Not low return — zero. The pilots that worked were scoped to one workflow with one number attached to it before anyone started building.
The House Rules
Review is a person promising to be careful forever. A gate is a machine that refuses.
1 month ago
The Teardown1 month ago
The reading list that wasn't
01 / 06Slide graphic
01 — The exhibit
In 2025 the Chicago Sun-Times published a summer reading list. Ten of the fifteen books did not exist. Real authors, invented titles, plausible blurbs — printed in a real newspaper.
02 / 06Slide graphic
02 — The tell: it reads perfectly
There is no grammatical seam, no weird phrasing, nothing a copy editor would catch. The text is exactly as fluent as the five real entries, because fluency and truth are produced by different parts of the process.
03 / 06Slide graphic
03 — The tell: plausible, not checkable
Each fake pairs a real author with a title that sounds like something they'd write. It's built to survive a skim and collapse under one search. Nobody ran the search.
04 / 06Slide graphic
04 — Why it fools people
A reading list is a low-stakes artifact, so it inherits the trust of the masthead around it. The format says 'someone chose these.' The pipeline never had a step where anyone chose anything.
05 / 06Slide graphic
05 — The fix
Titles are checkable against a catalogue. Bind every title to an ISBN lookup and make the pipeline refuse to emit an entry that doesn't resolve. Not a prompt asking for real books — a gate that drops unresolvable rows before they reach a page.
06 / 06Slide graphic
06 — The receipt
Attach the gate log from the run that dropped the unresolvable rows. If there's no log, there's no claim.
The 2-Minute Test1 month ago
Verification is the job now
Canadian courts have logged at least 167 filings containing AI-invented citations, and US courts issued over $145,000 in hallucination sanctions in Q1 2026 alone. These were lawyers — people whose entire profession is checking things. They didn't skip verification because they're lazy. They skipped it because the output looked done. Today: pick the step in your workflow where output looks finished but nothing has checked it, and put a check there.
The 2-Minute Test1 month ago
The polish is the tell
The Chicago Sun-Times ran a summer reading list where 10 of the 15 books did not exist. Titles, authors, blurbs — all invented, all published. Nothing looked wrong. That's the problem: fabrication doesn't arrive looking like fabrication, it arrives looking like a reading list. Polish is not evidence. Today: take the last thing your pipeline published and try to verify one specific fact in it from the source. If you can't, that's not a content problem, it's a pipeline problem.
The House Rules
Real measurements are ugly. If every number ends in a zero, nobody measured anything.
1 month ago
The House Rules
Anyone can generate. The value is in what you refuse to ship.
1 month ago
The House Rules
A demo proves it can happen once. A system proves it can't stop happening.
1 month ago
The House Rules
If you can't point at the log line that proves it, you don't get to say it out loud.
1 month ago
The House Rules
A rule that lives in a prompt is a wish. A rule that lives in code is a rule.
1 month ago
The 2-Minute Test1 month ago
Nobody remembers your AI ad
Svedka ran a fully AI-generated Super Bowl spot in 2026. Post-game surveys found 7% of viewers could name the brand — among the worst recall ever recorded for the slot. The pixels were fine. The ad was cheap, fast, and completely forgettable, because generation optimises for looking right, not for being remembered. Today: ask of your last AI-made asset — if you stripped the logo off, would anyone know it was yours?
PlaybookThis week
5 tells that your AI content is quietly lying
01 / 05Slide graphic
01 — Confident, no source
If it states a number or a name with total confidence and can't point to where it came from, treat it as fiction until proven otherwise.
02 / 05Slide graphic
02 — Rounded everything
Real data is messy. Suspiciously clean stats (‘over 10,000 users’, ‘3x growth’) are usually the model reaching for a plausible shape.
03 / 05Slide graphic
03 — Timeless voice
It reads like it could've been written any year. Real editorial has a now to it — a reference, a tension, a point of view.
04 / 05Slide graphic
04 — Every brand sounds the same
Drop the logo and you can't tell whose it is. That's drift. A voice spec is the fix, not more prompting.
05 / 05Slide graphic
05 — No one would disagree
If a take is safe enough that literally nobody would push back, it isn't a take. Ship a point of view or don't ship.
Note3 days ago
The 2-minute brand voice test
Paste three paragraphs of your best writing and three of your worst into a doc. The delta between them — not the average — is your actual voice. Spec the delta.
2:14Clip preview · add real media
Watch5 days ago
Making AI video look like a phone at 2am, not a movie trailer
Principle
If a system can't show where a fact came from, it doesn't get to publish it.
1 week ago
Field data1 week ago0hallucinated venue fields, across 5 cities
Every claim traces to a fetched source. The verification spine is the product — the content is the by-product.
Template2 weeks ago
The carousel structure I use for every post
Hook that names a tension → three slides that pay it off → one slide that reframes → a close that earns the follow. Nine slides, never ten.
Get the structure
Note2 weeks ago
Stop prompting for ‘engaging’
‘Engaging’, ‘compelling’, ‘punchy’ are noise words a model averages into mush. Give it a constraint instead: one idea, 40 words, no adjectives. Constraints read as voice.
Playbook3 weeks ago
How to make one content engine serve two brands
01 / 03Slide graphic
01 — Separate the spec from the system
The pipeline is the constant. Voice, palette, pillars and gates are the variables. If they're tangled, you built a one-off.
02 / 03Slide graphic
02 — Re-skin, don't rebuild
A second brand should take a fraction of the time. If it doesn't, the first build wasn't really an architecture.
03 / 03Slide graphic
03 — Gates travel, taste doesn't
Verification rules port cleanly. Editorial taste has to be re-specced per brand — that's the part worth your actual attention.
1:31Clip preview · add real media
Watch3 weeks ago
A 90-second teardown of a post that looks AI and how to fix it
Principle
Finished beats clever. Ship what someone can post, not a draft they have to fix.