Maya Chen

From asking about code to implementing and verifying real changes.

In this example, independent implementation tasks include requirements, changes, post-change checks, and an explicit record of unresolved issues.

Using connected tools and checking what actually happened.

Connected-tool results are read back and checked for completeness, with errors followed by recovery and another verification.

Applying reusable procedures and improving them through use.

Task-specific skills are applied and reviewed. A revised procedure is visible, but its subsequent use is not fully established.

Running automatic checks at the moments that matter.

Hooks perform automatic checks at useful moments. Follow-up evidence for exception handling is still missing.

Giving an agent the right information, and correcting it when needed.

Outdated context is corrected against source material, and the correction changes both the next decision and the delivered result.

Carrying useful knowledge across sessions and applying it again.

Saved rules inform work in another session. A validity review exists, but a later reapplication of the reviewed version remains unverified.

Designing feedback, retries, and clear stopping conditions.

A later execution reads persisted progress, respects retry and stop rules, and verifies its outcome separately from the generator’s own claim.

Connecting stages with explicit inputs, outputs, and dependencies.

Outputs flow between stages and some inputs are validated. Recovery and whole-flow validation are not completely demonstrated.

Dividing work between agents and checking the combined result.

Specialized agents return work that is integrated and checked for conflicts. Complete omission and duplication checks are still limited.

Keeping work separate and delivering it without unrelated changes.

Separate workspaces keep edits isolated. The delivery is reviewed, with incomplete evidence of the original workspace’s preserved state.

Choosing models for a reason and checking the effect of that choice.

Models are selected by role and used as assigned. Comparable evidence has not yet connected those choices to subsequent outcomes.

Evaluating AI outputs against criteria, then checking the evaluation itself.

Multiple AI executions are scored against defined criteria, grading errors are reviewed, and the corrected results are aggregated.