Experiments

What happens if I try this?

Updated 5 September 2026. Tests of AI code review and test conversion, alongside earlier workflow notes. These are observations from my own workflow, not controlled benchmarks.

Testing a PHPUnit-to-Pest conversion

Question
Could background AI work carry out a large test conversion that is worth taking through further review?
Setup
PHPUnit, Pest, and background AI work
Constraint
A passing suite is one check. It does not establish that the conversion is ready for adoption.
Observed result
About 600,000 lines changed over four days of background AI work, and the full suite passes. This is a feasibility result, not a shipped migration.

Next step: Further validation and an adoption decision are still pending.

Related: Engineering practices at Tillo

How much context does an AI reviewer need?

Question
How does limiting Harbourmaster to the diff affect review cost and useful findings?
Setup
TypeScript, OpenCode, a review panel, verification agent, and coordinator
Constraint
Cost and runtime need to be considered alongside accuracy and useful findings.
Observed result
The initial version used too many tokens exploring the codebase. Moving to diff-only review reduced cost and runtime, but narrower context also reduced accuracy and findings.

Next step: Keep checking findings when choosing the context and models for a review.

Related: The AI review tool

Agent-assisted delivery loops

Question
Testing where multi-agent workflows genuinely improve engineering output, and where orchestration overhead outweighs the gains.
Setup
Codex, Cursor, repo checklists, and parallel agent handoffs
Constraint
Useful speed only counts if review quality, integration discipline, and ownership stay intact.
Observed result
Parallel agents help most with repo orientation, first-draft implementation, and comparison work. The failure mode is orchestration theatre: too many agents, vague briefs, and more review debt than progress.

Next step: Keep the loop small: one clear owner, explicit checks, and tight handoffs instead of open-ended autonomy.

Related: The Velocity Paradox

Specification-first building

Question
Tested whether clearer intent, quality gates, and better prompts could improve architecture and communication when using AI tools.
Setup
Prompt scaffolds, acceptance criteria, and lightweight decision logs
Constraint
AI punishes vague intent immediately. Bad briefs turn into plausible but generic output.
Observed result
The biggest gain is not raw speed. It is forcing clearer thinking earlier, which improves architecture notes, pull requests, and conversations with humans as well as tools.

Next step: Carry the useful parts into future work: acceptance criteria, small decision records, and explicit checks.

Related: What I am focused on now

Product-system sketches

Question
Used small tools and interface sketches to expose where product behaviour and backend constraints did not line up.
Setup
Small Astro pages, interface notes, and API-shape experiments
Constraint
Interesting product ideas usually fail at the joins between interface, workflow, and backend reality.
Observed result
The strongest sketches are the ones that expose permissions, trust, handoffs, and maintenance load early. If those stay vague, the concept is not ready.

Next step: Use permissions, trust, handoffs, and maintenance load as an early checklist for future sketches.

Related: Selected work

How to read these notes

Each entry records a narrow question and the constraint that kept the test honest. A concluded entry stays here so its date and result do not drift into a claim about current work.

A useful result can be negative. If a workflow adds more review than it saves, that is evidence to make the loop smaller or stop it.