-
Keeping a Codebase Legible to Models
Ask a model to write a lock-free ring buffer in an empty directory and you will have working code in fifteen seconds. Ask the same model to ship a routine feature inside a three-year-old repository and you will often watch it spend twenty turns reading the wrong files, editing a dead helper, and handing back a pull request that takes a senior engineer an hour to untangle.
Read article → -
Beyond Chat Scrollback: Storing Plans and Agent Runs in One XML Tree
Why plans in Jira and execution in ephemeral chat windows drift apart, and how project.xml keeps tasks, dependencies, and sub-agent transcripts in one place.
Read article → -
Counting What the Model Pays For: tc and tokdump
A single lockfile or inline SVG can quietly burn 40% of a context window on every turn. Two small Unix tools make o200k_base token weight visible in local shells and CI.
Read article → -
What does ‘redacted’ mean? Designing a sentinel for computer-use telemetry
Why the obvious redaction markers fail for computer-use data, and what a better one looks like.
Read article →
Writing
Notes and empirical benchmarks on agentic software engineering, repository legibility, and developer tools.