Strong takes on AI. Made drinkable.
The latest pours — new entries, revised verdicts, honest corrections. If something in here changed, it’s on this shelf with a date on it.
June numbers in; verdict revised to "Worth it — budget for review."
08 Jul 2026A straight answer on adoption: fix verification, then let envy work.
24 Jun 2026No jargon, no evangelism. The questions people ask at lunch, answered like a human.
Ship faster, spend less. What a one-person shop squeezes out of these tools.
The management layer nobody ships in the SDK. The stuff I do for a living.
Reviews get written on launch day and never again — these get re-driven. Verdict first, one thing worth knowing, then what shone and what bit. When the facts move, the verdict moves.
The questions asked at lunch, on Slack, and by my mother. Short answers here, long versions linked.
No. Autocomplete finishes your sentence; this takes the ticket. You describe an outcome, it reads the codebase, plans, writes, runs and tests — then hands you a change to review. The skill it demands isn’t typing, it’s delegation.Long version → the Claude Code drive
Same engine class, different steering wheel. A chatbot answers; an agent acts — it runs tools, edits files, checks its own work in a loop. The interesting shift isn’t smarter answers, it’s models that do things and report back.See → Lexicon: Agent
The honest version: it’s already taking tasks, and jobs are bundles of tasks, so the bundle changes. On my team the writing shrank and the reviewing grew. The people at risk are the ones whose whole job was the part that got easy. Move toward judgment.Long version → an essay in The Human Layer
You can, and the number will go up, and it will mean nothing. Story points assume a human wrote the story. When one engineer plus three agents outships a squad, the interesting numbers are review time and defect escape, not velocity.Long version → the measurement playbook
Good news: refusal is a trust problem, and trust problems are solvable. Someone got burned by a confident wrong answer once. Don’t mandate usage — fix the verification step so being wrong is cheap, then let the fast people be visibly fast. Adoption follows envy.Long version → the rollout playbook
A human, every time, and it should be obvious who. Once agents write real volume, review ownership becomes your governance question — not "can we trust the AI" but "whose name is on the merge." Decide it before month six, because month six decides it for you.Long version → the delegation playbook
Depends who’s counting. If the alternative is a six-month procurement fight, the licence you already own wins by default. If you can choose freely, benchmark it against your own tickets for two weeks first. Most teams never run that test. Be the team that does.Long version → the Copilot drive
Close to it. A free-tier model for the cheap bulk work and a paid one only for the final pass gets you a long way at near-zero marginal cost. The trick isn’t the free tier — it’s knowing which step is cheap and which step you should pay for.See → the Gemini drive
The long-form, load-bearing guides — written to still be right next year. When one stops being right, it gets edited, not replaced by a fresher hot take.
The hard part isn’t building the agent — it’s managing one. Briefing, checkpoints, review ownership, and the moment you take the work back.
What to instrument when your sprint tooling assumes humans did the typing — and how to explain the new numbers upstairs.
Trust calibration, the verification reflex, and teaching people when not to use the tool. Everything here survived a real rollout.
The words people use to sound clever about AI, defined so you don’t have to nod along. The italic aside is the “with milk” bit — what it means when it hits your actual work.
The dotted border means cut along the line. Print them, pin them, paste them into your team wiki — attribution appreciated, not required.
The five lines every brief needs before an agent touches your codebase. Fits on a sticky note.
Get the sheet →ReviewSpec-level review with targeted spot checks. Caught everything the old way did, at a third of the time.
Get the sheet →ChoosingOne page, four questions, no benchmarks. The best model your team won’t verify loses to the one it will.
Get the sheet →Essays on the main page lean on this place the way a feature leans on its reporting. Follow the trail either direction.
India's payments rail is the most-cited case study in global fintech. Nearly everyone citing it is learning the wrong lesson — because the lesson isn't the rail, it's the preconditions.
Draws on the Copilot drive
Every failed AI rollout I've seen was a management failure wearing a model's clothes.
Draws on the delegation playbook · Claude Code drive
It had the best engagement numbers in the app. It was also quietly destroying retention.
Draws on the Collab System drive
The whole place on one shelf — for the days you come to look something up, not to browse. Type to filter across all of it.