Patterns
What I've learned building software with AI coding agents, written up as patterns you can use.
It's a kind of personal digital archaeology. I go back through the transcripts of my own sessions with coding agents, the saved record of everything they and I did, and look for the things I keep doing. The ones that hold up get written up here, then built into how I work, and I go round again. Some of them might be useful to you.
An AI agent drafts each page from my notes, and each page says whether I've edited it yet. Every one has an example you can run and a skill, a short set of instructions you can give your own coding agent.
Checks you can trust
Automated checks that hold a line in a codebase, and how to tell whether they are really looking.
The ratchet
Switch on a rule over code that already breaks it, and let the count of breakages only fall.
A check you have never seen fail is not a check
Test-driven development's "watch it fail first", applied to lint rules, build jobs and the checks agents write.
Checking what the browser actually drew
Measure where the browser put each piece of the page, for the layout bugs that code checks and screenshot tests miss.
Running agents
Running many coding agents at once: what each one needs from you, and what their logs can teach you.
Agents say how they stopped
Each time an agent stops, it gives one label saying what it asks of you, so a list of idle sessions shows who needs you.
What your agents get wrong is a dataset
Log the failed commands your agents quietly recovered from, rank them, and fix each trap once.
Software that hunts its own bugs overnight
An agent looks for bugs in the day's changes, a second one tries to disprove each, and only the survivors are filed.
Follow-ups that wait on an event
The steps only a person can do after a change is merged, each filed with the event it waits for.
Mining your own transcripts
Read your own agent sessions back for what keeps recurring, write it up, turn it into rules, and check the rules held.
Lists and pages that stay true
Keeping documents and lists in step with the code, so they fail a check when they stop being true.
Documentation where every sentence has evidence
List each factual sentence beside its evidence, so changing either one fails a check.
Manifests for everything
One typed list of every environment variable, or every log event, that fails when it and the code disagree either way.
Journeys as typed data, prose generated
Declare the paths people take through a product in code, against real events, and generate the written version.
Judgement and taste
Where a machine can score the work but a person has to decide.
The machine scores, only a human rejects
A model scores a list too long to read but never removes a thing, and its verdict counts only as far as its evidence survives.
Taste as a record
Agents write down how you react to a design, in your words. Only you can turn a reaction into a rule.
The cold reader
A reviewer that sees only the rendered page lists every name and "the X" a first-time visitor couldn't work out.
Your writing voice, captured and checked
Keep what an agent drafts in your name sounding like you: a measured profile, a check run until it passes, a cold read.