Richard Adleta | Notes

Stored eval cases raise the cost of every change

2026-10-13T00:00:00+00:00

Nine days ago I put a testing standard on my skills: a change had to beat a no-skill run on the eval’s mean score, with no case scoring worse than before. It came with 43 stored cases. Today there are 61, 354 files and 2.0 MB across 22 skills. Agents add edge cases and more tests for anything they do, and every new case raised the token cost of every later change.

The scores were a huge waste to maintain. Yesterday a build was flagged on four cases a few hundredths below the old version, 0.89 against 0.96 at the widest. Each flag was a question to me, and I ruled all four run variance.

So today I deleted all 61 cases. The tooling stays for a run I ask for. Evals are a one off way to check things, not a form to dev against.

The terminal output style was shaping Claude’s writing

2026-10-13T00:00:00+00:00

My terminal output style was leaking into everything Claude wrote for me. It asks for one idea per sentence, short paragraphs and no idioms; my voice rule asks for varied rhythm, light idioms and humor. The two are for different audiences, and the style was winning.

The output style is loaded in the main session, so I can’t rely on the main session to draft copy. Three sub-agents asked to quote the style from their own context all reported it absent, while all three quoted my personal instructions file correctly. That’s the agents’ own report, unchecked against the harness, and a sub-agent draft did read much more natural.

So letters, emails, replies, posts, READMEs, PR descriptions and docs now go to a drafting sub-agent, with a separate reviewer agent cold-reading each draft. The main session shows the draft as written and sends any change back to the drafter.

Don’t tell the agent it’s running out of context

2026-08-24T00:00:00+00:00

Tell an agent its context is filling and it rushes. It can’t gauge what’s left or what matters, so it closes everything and leaves work undone. They’re goal oriented: give them an ending and they take it early.

Today an agent proposed a hook to warn it on every prompt past 300k tokens. I said no. That 300k is my line in the sand, not the agent’s. I watch the meter, stop at a clean point between 300k and 400k, hand off and restart.

I’m watching for a quieter way, a fact at a checkpoint, never an alarm.

Turn off your agent’s context warnings and watch the number yourself.