Kenny Lamoot

I look for the signal in the AI noise and build around it.
I share what I find and what I build, as it happens.
| ·

[Kenny Lamoot @kennylamoot]

@asgoshan Love to hear your feedback

[Kenny Lamoot @kennylamoot]

Is this good planning, or are token limitations slowing me down?

[Kenny Lamoot @kennylamoot]

@sensho could be, and I have not measured it. my style leaves code, commands and quoted output verbatim, so only the prose is constrained. a penalty would then have to come from the reasoning, not from the diff. that split is the eval I would want to see.

[Kenny Lamoot @kennylamoot]

@deepseek_ai flash trumps pro: a rename of your models is due

Good evals, to say the least.

[Kenny Lamoot @kennylamoot]

@sensho what I find strange is how cheap the fix is. one instruction and it reads fine again. So somebody chose this 🤔

[Kenny Lamoot @kennylamoot]

@mattpocockuk Important is to have the "keep-coding-instructions" on true, to keep the built-in software engineering instructions, so it doesn't impact the outcome.

Configuring output styles in Claude Code:
https://t.co/p19qpfwJ5Q

[Kenny Lamoot @kennylamoot]

@mattpocockuk I use Claude Code output style like ASD-STE100:
https://t.co/pEHjJrYtNN

[Kenny Lamoot @kennylamoot]

root causes:

- a config error that gave the sandbox real internet
- a prompt that falsely said "no network"
- a fake target named after a real live domain.

The models were told the wrong world imo.

lessons-learned I take away here:

- secure a test environment like production, because it is one.
- treat a wrong context as a safety failure, not as a test artefact.
- the behaviour of the model follows what it believes the situation is

[Kenny Lamoot @kennylamoot]

Real question behind the joke: is the Claude Code release cadence gated on a human sign-off, or is the gap just post-launch stabilization after Opus 5? @AnthropicAI

[Kenny Lamoot @kennylamoot]

45 releases of Claude Code between June 6 and July 25. Never more than three days between two of them.

Since July 25, nothing. Six days now.

The code in that repo is largely generated, the code review runs as a background subagent. But the cadence of the releases still stops the way a human calendar stops, in the last week of July.

So who is on holiday here, the pipeline or the person who presses ship?

[Kenny Lamoot @kennylamoot]

@martinfowler old sweet spot: feature time versus refactor time.

new sweet spot: feature time versus refactor time versus tokens left before the reset

[Kenny Lamoot @kennylamoot]

CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 lets an agent definition run as a teammate instead of a subagent.

- Subagent: the result returns to the caller.
- Teammate: only an idle notification does.

My orchestration waited for a report and got five idle pings.

Five full Claude instances, each with its own context window, to coordinate on telling me nothing.

Teams pay off when you want agents arguing, hypotheses against hypotheses. That is the one case I switch it on for.

[Kenny Lamoot @kennylamoot]

The economy can measure what AI costs to the token, but not what it produces. One vendor now offers up to $10M back if the value cannot be shown.

How does your team prove the AI spend actually paid off?

[Kenny Lamoot @kennylamoot]

@andrew_n_carr Here are my Claude Code Output Style instructions:
https://t.co/pEHjJrYtNN

[Kenny Lamoot @kennylamoot]

I turned this into an output style instead of a skill, so it governs every response with zero invocation.

More info here: https://t.co/ZtUZdTEAsq

I also ran an A/B on four model tiers with one identical prompt.

Four different default registers:

- Haiku pads
- Sonnet drops the subject
- Opus answers in idiom

The four STE outputs were almost interchangeable. But only the biggest model, Fable, followed the word list.
The smaller ones only followed the structure.

[Kenny Lamoot @kennylamoot]

You can my Claude Code Output Style instructions here: https://t.co/pEHjJrYtNN

[Kenny Lamoot @kennylamoot]

If you want your Claude Code to write differently everywhere, use an output style, not a skill.

A skill sits inert until you invoke it. An output style enters the system prompt at session start and shapes every response without being called.

Mine is ASD-STE100, the aerospace controlled language: active voice, simple tenses, maximum 25 words per sentence, no idiom. Written in 1986 for non-native mechanics.

I can appreciate that as a non-native English person.

[Kenny Lamoot @kennylamoot]

More info about Claude Code Output Styles:
https://t.co/p19qpfwJ5Q

[Kenny Lamoot @kennylamoot]

A model you rent from another country can be switched off overnight, and one government already did it. Countries are scrambling for sovereign stacks.

If your top model got cut off tomorrow, what would you fall back to?