3 min read
The average score hid the break
Folding one subagent in cut latency and broke a process that had been passing, in one of two runs. How to test any AI change, question by question.
Read the post
All posts
Notes on working with AI
Claude Code for engineering teams
Claude Code, agents and token bills: how engineering teams run Claude on real code and keep the costs in view.
3 min read
Don't make Claude read every row
One job took 150,000 tokens, then about 2,000 once the agent wrote code instead of reading everything. Why that works, and a version you can use today.
Read the post
5 min read
Why AI websites all look the same
Same purple gradient, same three cards. Why AI-built pages look alike, and the design skills for Claude Code that check their own work in a browser.
Read the post
4 min read
More effort is not more care
Max effort on a snooze button gets you a scheduling engine. Why turning the dial up backfires, and where to put the effort instead.
Read the post
5 min read
Pick your 40 lines on purpose
Your agent hands you 600 lines and you read 40. Anthropic's Claude Code team built a system for choosing which 40, and it works for any AI output.
Read the post
3 min read
Understand the software before you clone it
Agents can now explain how an app works without its source code. The valuable output is the mechanism and its evidence, not the clone. Here is the line.
Read the post
8 min read
Which Claude Code framework should your team use? Start with none
No independent test shows Superpowers, ECC, Spec Kit or GSD beating plain Claude Code. What the evidence says, and which one to try for which problem.
Read the post
3 min read
A plan you can answer beats a plan you can read
The slow part of building with Claude is no longer producing the work. It's reviewing the plan. A plan shaped as decisions you answer gets reviewed; a long text file gets skimmed.
Read the post
3 min read
A video made with code is a function of time
When Claude builds a video in code, every frame is a function of time. That is why each frame can be checked, and why a storyboard before rendering saves hours.
Read the post
3 min read
Add a camera and a referee before you add a second agent
An agent you can't see is one you can't trust, and one you can't grade can't improve. A trace and a fixed test set close most of that gap.
Read the post
3 min read
The small model is the workhorse
Most work in an AI workflow is small, routine calls. Paying top-model prices for them is the main waste. Use a big model to plan and small models to do the volume.
Read the post
2 min read
Your setup costs tokens every session
Everything always loaded in Claude Code is re-read every turn. Audit your instructions, skills and connectors like a subscription bill.
Read the post
3 min read
The moment Claude says it can't check
Ask Claude to check today's price and it might tell you it can't, its knowledge stops at training. A connector fixes that, but only if you pick the one built for the task you actually have, not all five at once.
Read the post
4 min read
MCP just learned the trick your phone company never did
Every MCP connector used to work like a call center that forgets you the second you're transferred. That's changing, and it means more of the software you already use is about to plug into Claude.
Read the post
4 min read
The instruction file getting shorter, not longer
An instruction file that grows one frustrated rule at a time eventually fights itself. The newer models do better with a short one that trusts judgment instead.
Read the post
4 min read
Automate what you can check. Hand off what you can't.
A calendar slot is a fact you can check. A customer's judgment call isn't. The gap between them is the whole question worth asking before you automate anything in your business.
Read the post
3 min read
When specializing your agents helps, and when it's just another meeting
Split work across specialist agents and each one gets sharper at its slice. Add enough of them without enough distinct work to justify it, and you've built a meeting, not a team. The rule is the same one that decides when an event planner needs a lighting specialist.
Read the post
3 min read
A dev team in a folder, and the question it doesn't answer
Three saves landed in the same folder this week, and none of them taught a technique. Together they say something about what capability looks like now. It arrives in bulk, and installing it is not the same as knowing which three pieces you will actually use.
Read the post
4 min read
Loop engineering is just the loop you already run, with an agent in your seat
Everyone is posting diagrams about loop engineering like it's a new discipline. It isn't. Here's the actual anatomy, the one idea in it worth taking seriously, and the honest test for whether you should build one at all.
Read the post
3 min read
The mechanic behind every surprising Claude Code bill
We run enough scheduled Claude Code sessions now to actually watch the token bill, and the biggest driver of it has nothing to do with how much work got done. It's a caching mechanic almost nobody explains, and it can turn one word into a real charge.
Read the post
Claude training for teams
Want your team working this way?
This is the material we teach in workshops, at your office or online. Tell us what you're trying to solve, and we'll come back with a plan.