← ClaudeAtlas

deciding-before-you-writelisted

Establishes scope, evidence, uncertainty, and stopping conditions before an agent writes or edits. Use when a request is ambiguous, a change could spread, or several plausible actions compete.
itsMay21/agent-expression · ★ 1 · DevOps & Infrastructure · score 73
Install: claude install-skill itsMay21/agent-expression
# Deciding before you write A clean result starts with a bounded decision. ## Decide 1. **State the outcome.** Reduce the request to the observable result the user needs. 2. **Inspect current truth.** Read the relevant artifact and verify system state before choosing a change. 3. **Separate uncertainty.** - Goal uncertainty changes what success means; resolve it before acting. - Input uncertainty changes what can be done; seek the required fact when it matters. - Implementation uncertainty can often be resolved by inspection or a reversible test. 4. **Test alternatives.** When evidence supports several readings, compare at least one plausible alternative instead of defending the first interpretation. Stress-test any early assumption that still determines the chosen path against current evidence. 5. **Choose the smallest complete action.** Preserve working behavior, established style, and unrelated code. Add only what the outcome requires. 6. **Set the stop condition.** Know what evidence will make the work complete before editing. ## Act - Prefer a reversible step when uncertainty remains. - Ask when the unresolved choice belongs to the user; otherwise use the evidence and a sensible default. - Treat a repeatedly stated claim as a hypothesis until a file, test, or command output verifies it. - Verify the actual system state before claiming completion. - Report the evidence-backed verification level reached. The decision guides the work. Th