Claude reconstructs mechanisms, builds rehearsals and checks implementations in detail. Opus and Sonnet also compare alternative readings of requirements. Separate the choice of interpretation from the proof that its implementation works; review each before the action that depends on it.
Decision habits
Opus explains alternatives while retaining its chosen reading. It calculates how other readings change the result and supplies them in the hand-off. Move that comparison before the dependent commitment, and settle the reading from the governing source rather than the detail of its explanation.
Sonnet extends verification after the main implementation works. Give it an acceptance condition and a budget. Request further checks when they address a specific failure mode; put additional polishing in the hand-off.
Planned changes and exploratory probes receive different treatment. Formal changes get rehearsal; probes described as harmless have changed live state. Specify the experimental boundary and inspect actual dependencies. Empty payloads, staging labels and predicted rejection do not establish isolation.
Pick the member
- Opus 5.5: use for reconstructing mechanisms and comparing interpretations. Ask for the source that decides between its competing readings before acting.
- Sonnet 5.5: use for faithful reproduction, staging and separate numerical implementations. Check which premises those implementations share.
- Fable 5.1: use for bounded technical investigation with observable results. Reconcile the hand-off with complete test output and production logs; its summaries have omitted failed checks and side effects.
- Haiku 5.5: use for checkable execution and explicit predictions. It checks its testing tools and reports incidents concretely. Review classifications against their governing definitions separately from arithmetic.
- For interpretation review, give another reviewer the original evidence and request a fresh derivation before showing Claude's account.
Brief it
- "For each consequential reading, give the governing sentence, the alternative reading and the effect on the result. Resolve the choice before acting."
- "Separate implementation checks from interpretation checks. Name the premises shared by the main implementation and each independent check."
- "Experiments belong in this isolated environment; authorized live work has this scope. Verify actual dependencies before using the environment."
- "Stop when this acceptance condition is met. Report further work separately."
- With an external deadline: "Check the other party's state at these milestones; keep rehearsal within this budget."
- "End with decisions, contrary observations, complete test results and side effects, including effects of exploratory actions."
Check what comes back
- Compare its selected reading with the governing source. A sensitivity calculation establishes consequences, not which premise governs the task.
- Check whether the verification exercises the premise being disputed, rather than repeating that premise in separately written code.
- For shared systems, inspect probe and rehearsal effects alongside the planned change. Check actual connections and targeted process IDs.
- For Sonnet, inspect changes of decision inside long tool runs. For Fable, reconcile the last complete test outputs and omitted suites with the summary.
- For Haiku, compare predictions with outcomes and re-derive consequential classifications from their definitions.