Systems Modernization & AI Development

Earlier I found the one rule that survived a model-offloading experiment: verify the artifact, never the account. The last post turned the same rule on instructions: make the expectation checkable, so the work can fail against it. This one turns it on your own findings. Same discipline, third surface: bookkeeping.... [Read More]

We’re taught to soften feedback so it lands. That’s the part that breaks it. [Read More]

The premise of offloading work to a cheaper model is great. Let the expensive model design. Let the cheap model implement. Put hard gates around the cheap model so it cannot ship something broken. Coverage thresholds, benchmark checks, documentation currency. The gate is supposed to make the model’s reliability irrelevant.... [Read More]

I spent a week comparing Claude Opus and Claude Sonnet on the same spec-driven pipeline, run against pdb_search, a brownfield Python repo of mine. Opus is the stronger model, Sonnet the cheaper one. To score the runs I used a third instance, an Opus 4.8 orchestrator. Same family as one... [Read More]

Previous posts have analysed how Spec Kit handles a small brownfield database (Post 1, Post 2, Post 3). This post is about whether the same pipeline fits legacy modernisation. [Read More]