Find the game before making the move
A six-step orientation method for replacing premature action with a clearer model of the situation, incentives, resources, and next move.
Writing, in reverse
Thoughts
Notes from work, experiments and the life around them.
Automate the job, not the typing
What happened when one engineer pushed software automation on a live product with more than 500 users while retaining judgment and production authority.
I stopped managing agents with memory
A founder-oriented state machine for moving agentic work from scope to production evidence without relying on remembered instructions.
Make the data argue back
A claim-first data framework for preserving meaning across sources, aggregation grains, transformations, and product surfaces.
Two screenshots became a failing data contract
A debugging workflow that normalizes two conflicting claims, traces one value to source, and preserves the finding as an executable contract.
The arithmetic was correct. The grain was wrong.
Why censored detailed rows cannot reconstruct an exact coarse total, and how to make aggregation grain part of the product contract.
Data before code is a graph problem
Model a data-dependent deployment as a commit-ancestry invariant instead of a current-push checklist.
The reviewer rejected my fix twice. Both times it was right.
Why passing tests and a good staging walk did not prove a data-dependent release was safe, and how an independent reviewer found two temporal bugs.
Adopt, adapt, skip
Triaging someone else’s 24 agent skills against our own gates led us to adopt five, adapt six, and skip thirteen. The principle that fell out: a skill teaches behaviour, while a gate enforces it.
The machine that said no to my machine
A data-seeding agent hit half-rotated keys, tried to fetch replacements, and a second AI refused twice. Why a denied request was the system working, and the guardrail I am glad I do not own.
Skills are context compression
Every agent gets the same 1,500-token sermon before its 50-token task. Agent skills fix that by moving proven protocols out of pasted prompts and into versioned SKILL.md files through progressive disclosure and compression by reference.
I ran the robot factory by hand first
A full manual dress-rehearsal of the L5 feedback-to-fix factory turned eight pages of handwritten notes into 24 findings, four tickets, and four parallel build lanes. It taught me what to automate before automating it.
Loops, levels, and building a platform that drives itself
What an agent loop really is, where our platform sits on the L0–L5 ladder, the gates holding it up, and the road to a platform managed by exception.
Our operating model, in autonomy levels
Mapping how we actually run agents onto Addy Osmani’s L0–L5 autonomy levels: where we sit, what evidence makes each level defensible, and the one gap we just closed.
The most expensive tester in my company is me
The most expensive tester in my company is me, catching bugs by eye on staging. Shift-left is the fix: every catch becomes a gate that blocks the bug automatically.
Turn repeated catches into controls
A five-step method for converting recurring human attention into an automated check that proves its own failure path.
I built a check, and the first thing it caught was me
I turned one rule I only enforced by discipline into a gate that fails the build. The first thing it caught was my own half-finished work.
I stopped writing the code. I did not stop being the engineer.
I ship more by typing almost none of it. The workflow, honestly: delegate the building, keep the deciding and the verifying, and treat every “done” as a claim until it is proven.
Sandboxing AI Agents: Running OpenClaw in Docker So It Doesn't Nuke Your Codebase
A walkthrough of setting up OpenClaw, an AI terminal agent, inside a locked-down Docker container with sandboxing, capability dropping, and Ollama connectivity