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
All posts
Notes on working with AI
Claude for operations teams
Agents, scheduled tasks and classifiers: how operations teams automate what they can check and hand off what they can't.
4 min read
One agent, many playbooks
Fifteen half-finished AI projects, a few that work some of the time? Stop building agents per task. Give one agent a playbook per job.
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
Not everything needs a language model
Many AI projects are classic prediction problems with a table of past data. A language model is the wrong tool for them, and Claude is most useful as the analyst who builds the right one.
Read the post
4 min read
The most useful thing a company AI can say
A company AI that answers every question has a design defect. The fix is a refusal threshold, cited sentences, and a test set that includes questions it must decline.
Read the post
4 min read
Give your first agent read access and nothing else
An agent is data access, written rules and a scheduler. The design choice that matters most is keeping read permissions apart from action permissions, then proving the rules in a dry run.
Read the post
3 min read
Most business decisions are classification, so stop paying for generation
A model that returns a probability over fixed options routes work better than one that writes a persuasive paragraph. Calibration, latency and cost per decision decide which model you need.
Read the post
4 min read
A classifier is only as good as its rulebook
Automated sorting stalls when the rules are written once. A cascade with a confidence bar turns every escalated case into an edit to a written rubric, which keeps the improvement auditable.
Read the post
3 min read
A one-off task saves an afternoon. A scheduled one saves it every week
Handing Claude a whole job is the part of the demo that gets a reaction. Turning that job into a scheduled task is what gives you the time back every week.
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
The question I ask before any agent gets a real job
Most teams that get an agent working never get past the demo, and the reason is almost never the model. The real question is whether you can ask it, each evening, what it did, and get an answer.
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.