Notes
Written when something was learned.
Dated and opinionated. Allowed to be wrong later; a later note says so. Filter by who it is for.
Filtering happens on the page: pieces marked "for both" appear under either audience. Prefer a subject to an audience? Browse by topic →
Oct 28, 2026For companiesDraft
Check the data before you sell the map
I pulled three public data portals for my own region into one AI-coding session to find products local businesses would pay for. Joining the data was the easy part. The checks afterwards killed two of the best-sounding ideas, moved a third, and showed that "nobody does this here" was wrong. The four checks to run before you build anything on someone else's data.
Oct 28, 2026For companiesDraft
The report that lists its own gaps
I built a property due-diligence report from public data: one address in, nine sections out, about forty queries in thirty seconds. I picked the test site blind, and it turned out to have a live application for more than 1,300 homes. The report held up. The most useful page was the list of eight things it couldn't answer. What that taught me about any report an AI tool helps you produce.
Oct 21, 2026For companiesDraft
My website said "security" sixty times and "revenue" eight. Owners buy the other way round.
A services site written by a technical person sells protection. An owner buys gains first: hours back, costs down, work won. I counted the outcome words on my own owner-facing pages and found the imbalance. What I changed, and how to run the same count on any AI proposal put in front of you.
Oct 21, 2026For companiesDraft
"Are we aiming too high?" The sentences were fine. The shelf was wrong.
When I asked whether my own site sounded too expert for the owners it's for, the instinct was to rewrite. The plain-English pass had already been done. The problem was which subjects dominated, which pieces were labelled for everyone, and which expert words had leaked onto the introductory pages. A five-step audit to run before touching a sentence.
Oct 14, 2026For companiesDraft
"Done" is a ticked checklist, not a feeling
Software projects drag on because the client is asked "is this how you want it?" instead of "does it work, as specified?" Only the second has an answer. One yes/no test per scope requirement, on a page the client ticks, and every "could it also" goes somewhere else.
Oct 14, 2026For companiesDraft
The page that lets a software project end
Every "could it also" raised during a build needs a home, or the phase never closes. One page holds everything raised but not in scope, in ordered waves with the ordering rules written at the top, and every item on it is a change request.
Oct 7, 2026For companiesDraft
Five companies that got AI wrong, with the receipts
Real companies, their own words, and the one decision each of them skipped. A coding agent that deleted production and its backups, an email bot that confirmed a typo, 45 jobs cut and reinstated, a chatbot a Canadian tribunal said was the company, and $31,150 in costs for invented case law.
Oct 7, 2026For companiesDraft
I nearly built an AI assistant that reports your staff to you
An AI chat for employers that tells management when someone pastes client data in or goes looking at things they shouldn't. I did the homework before building it. The big vendors already sell it, it watches the wrong place, and the reporting is the part lawyers push back on. What a small firm should want instead.
Sep 24, 2026For companies
Canada's privacy finding on ChatGPT isn't about your business. These rules are.
The May 2026 finding against OpenAI is about how OpenAI trained its models. The rules for a business whose staff use AI were published by the same regulators in 2023, and most small firms have never read them. What they say, in plain English, and how they fit on one page.
Sep 24, 2026For AI-codersShipping inside a company
Your AI coding tool is part of your supply chain
The tool that writes your code also installs code: plugins, extensions, servers. In September 2026 a flaw in four major coding agents showed that a pinned plugin wasn't really pinned. What that means for anyone coding with AI inside a company, and four checks that cover it.
Sep 11, 2026For companies
AI writes instant legacy code
Code written with AI is being thrown away or rewritten within weeks at twice the old rate, and the share of code being refactored is falling. The draft is faster. Owning it is not.
Sep 11, 2026For companies
Fast food architecture
AI is a very good drafter and a poor architect. Treating its output as finished engineering is like living on fast food: it works today and costs you later. The fix is a place the reasoning lives that the model can read back.
Sep 11, 2026For companies
The build is the cheap part
A tool built in 48 hours feels free until the first API change. The initial build was always the small share of the cost of owning software, and AI just made that share smaller.
Sep 11, 2026For companies
The tool nobody in IT knows about
Shadow IT used to be a spreadsheet. Now it is a working app that runs part of the business, built by one person with an AI tool, and that person is the only one who can change it.
Sep 5, 2026For companies
A chatbot for everyone is not an AI strategy
It is the most common AI decision a company makes, and the most common place it stops. What the licence buys, what it does not, and the four things it quietly leaves undone.
Sep 5, 2026For AI-codersShipping inside a company
Why I put the authorization rule in the database, not the app
Multi-tenant ticketing with row-level security, and why "you can only see your own company's tickets" belongs below the application code.
Sep 5, 2026For AI-codersYour first weeks
Make the tool interview you before it builds anything
The cheapest review happens before the code exists. Three sentences that turn an AI coding tool from an eager builder into a careful interviewer, and when to use them.
Sep 5, 2026For bothRunning a team or several apps
Porting the product, not the platform
Moving a CRM off a no-code tool means documenting every screen and automation first, then rebuilding the flows the team already knows, not the workarounds the platform forced.
Sep 5, 2026For companies
Seven questions to ask anyone selling you AI
The vendor has a demo. You need answers. Seven questions that separate an AI feature from an AI liability, and what a good answer sounds like.
Sep 5, 2026For AI-codersYour first weeks
The AI writes the code fast. Reading it is the job.
What changed in my day when the typing stopped being the bottleneck, and the three places the time went instead.
Sep 5, 2026For AI-codersShipping inside a company
A security pass that runs whether or not anyone remembers
Scheduled scans, nightly dependency matching, findings as tickets, fixes as reviewed pull requests. The shape of a security practice for a small team that ships with AI.
Sep 5, 2026For companies
The app is the visible part
Companies that build or buy an AI feature usually stop at the feature. Three invisible parts decide whether it is an asset or a liability, and none of them is optional.
Sep 5, 2026For companiesYour first weeks
The coding standard is a file the tool reads
Three people building with AI produce three different codebases unless the standard lives somewhere the tool reads it every time. Not a wiki page. A file in the repo.
Sep 5, 2026For companies
The employee who can now build apps
Someone in your company who is not an engineer has built something the business now relies on. It is an opportunity and the biggest new risk in most small firms. What to require before you rely on it.
Sep 5, 2026For AI-codersYour first weeks
Three design directions in an evening, and why the client picked the one I did not expect
A method for getting a real design choice out of a non-designer: stable option names, an honest tradeoff on each, and a fresh-eyes critic pass before anyone sees it.
Sep 5, 2026For AI-codersRunning a team or several apps
An intake step that asks two questions before a human reads the ticket
Using a model to clarify vague support requests against a map of the client's systems, with a person still approving every ticket.
Sep 5, 2026For companies
What a company keeps when the coder leaves
Git keeps the code. It does not keep the reasoning, the gotcha that cost a day, or why the last approach was abandoned. That is a different kind of backup, and most companies do not have it.
Sep 5, 2026For bothShipping inside a company
Where the rule lives
"You can only see your own company's tickets" is one sentence. Where a team chooses to enforce it decides whether the next feature can forget it. A story about moving a rule below the code, told for people who will never write a database policy.
Sep 5, 2026For companies
Your staff already have an AI plan. You just have not read it.
In most companies without a policy, AI is already in daily use on whatever data is to hand. That is not a discipline problem. It is the plan you get by not writing one.