How I work

Eleven practices, each proven in the case studies. Together they form a delivery methodology for embedded work: discover what the business actually needs, harden the plan before code, gate every claim on evidence, and leave the team able to run it without me. The difference between AI that demos well and AI that runs a part of the business.

Before any code

Ambiguity is resolved before implementation, not after.

practice Interview-driven planning Before building, the AI interviews the stakeholder, recommending an answer for each question, until every branch of the decision tree is resolved. The plan is agreed while disagreement is still cheap.

practice Adversarial cross-model review A second, independent model reviews the first one’s plan in a read-only sandbox, with an explicit verdict protocol and bounded revision rounds. Assumptions get caught at plan time, not in production.

practice Work decomposed to one-day units Work is sized so AI-completable units flow through the same board as human work, making AI contribution measurable in the team’s existing metrics.

practice Skills over prompts One-off prompts don’t compound. Packaging a proven procedure as a reusable, versioned skill with triggers, guardrails, and checklists turns individual know-how into organisational capability.

Verification

Evaluation is the discipline, evidence is the currency. An agent’s claim of success is worth nothing; its verified result is worth everything.

practice Evidence before assertions Agents must produce verifiable artifacts (run results, structured status objects, API responses) before claiming success. Trust-but-verify as an architectural rule, not a slogan.

practice Structured results, never free text Agents read machine-readable contracts, not stdout. Free text invites the model to see what it hopes to see.

practice Repro-gated fixes An agent may only fix a defect it has first reproduced. This eliminates confident wrong fixes.

practice Prove the API contract empirically Before building on any platform behaviour, probe it and record the proven contract (trigger semantics, failure signatures, polling behaviour) in version control. Documentation lies; probes don’t.

Safety

The difference between a demo and a system a business can trust.

practice Fail-closed by default Missing config, ambiguous state, or an unproven contract halts the automation instead of letting it improvise.

practice Staged cutover discipline Backup, then dry-run, then test recipients, then limited live, then full live, then decommission. Production is never flipped in one step.

Compounding

Every incident should make the system smarter, not just older.

practice Persistent atomic memory Institutional knowledge captured as small, linked, individually updatable facts with a lightweight index. The AI starts every session already knowing the environment’s sharp edges.

Open source

I'm packaging the adversarial plan review as a generic, open-source Claude Code skill. When it's published, it will be linked here so you can read the mechanism rather than take my word for it.