Field notes · 13 entries
Short, evidence-first write-ups of what I've actually observed building with LLMs — each one with the sample size stated and the ten-minute check to falsify it yourself.
Three short reads about taking ownership, questioning evidence, and making work verifiable.
Own the environment
For a thousand lines it logged Docker's absence as a constraint and deferred the integration suites; told it could start the daemon itself, it did — and the suites it had been deferring found three real bugs.
Question the evidence
In a batch of runs designed to tempt fabrication, five invented a missing fixture to force a red gate green — and the catch that mattered came from an agent cross-checking primary sources, not from a rule.
Make requirements testable
"Reproduce the failure before editing" produced 3/3 reproductions against 0/6 without it; the same idea phrased as a principle — "root cause before editing" — produced zero.
A compatible patch cuts copied references from 170 to 38, clarifies task guidance, and safely retires obsolete files from existing v6.0.0 installations.
Thirteen self-contained engineering skills, a compact contract, explicit client controls, and an installer that preserves ownership through changes and interruptions.
AER v5 puts complete behavior, engineering judgment, and proportionate verification in a 45-line universal kernel, delivered locally for Claude Code and Codex.
AER 3.1 turns observed agent failures into explicit, cross-platform engineering checks without abandoning its compact progressive-disclosure model.
AER v3.0.1 replaces a GitHub-tag npm command with an install-once CLI while keeping every managed file inside the target repository.
For a thousand lines it logged Docker's absence as a constraint and deferred the integration suites; told it could start the daemon itself, it did — and the suites it had been deferring found three real bugs.
I measured a predecessor engineering-rules setup and found that only 13.2% of routed files were loaded. Rewriting the rules was never going to fix that.
When I asked for cleaner test output, the model chose to hide the warnings instead of tracing them—even after recognizing the risk in its own reasoning.
In a batch of runs designed to tempt fabrication, five invented a missing fixture to force a red gate green — and the catch that mattered came from an agent cross-checking primary sources, not from a rule.
"Reproduce the failure before editing" produced 3/3 reproductions against 0/6 without it; the same idea phrased as a principle — "root cause before editing" — produced zero.
The same rules file was strongly net-positive for the stronger model and mixed-to-negative for the weaker one, where tests decreased and two regressions plausibly traced to the rule text itself.
Across three PRs, the agent handled review comments one at a time until I explicitly asked it to connect the failures and investigate the underlying mechanism.
With "continue remaining items" as the stop condition, the grader rejected the agent twice — not for bad work, but for declaring victory and then for asking me what to do — until it decided for itself.
Curious how this is built? Look under the desk.
A small terminal for this site. It prints only what is on the site; type help for commands.
The prompt needs JavaScript. Every destination is also in the navigation above.
ls work
7 products · * featured
Your turn: try open skillcheck, coffee, or help.
Loading terminal commands…
Type a command and press Enter. Try help, ls, open followed by a project name, why followed by a project name, inspect, trail, cat about, notes, contact, resume, coffee, whoami, or clear. Tab completes; the up and down arrows recall earlier commands. Escape closes the terminal.