AI Coding-Agent Work Plan Template

Plan from outcomes and constraints

Restate the user-visible outcome, exclusions, and acceptance criteria before listing implementation steps. Split work at independently verifiable boundaries: research, contract or schema, implementation, tests, documentation, review, release, and operational evidence. Show dependencies and identify which steps may safely run in parallel.

A useful plan is neither a transcript of imagined commands nor a promise that unknown work is already understood. Mark assumptions, discovery questions, irreversible actions, external writes, and decisions that require a person. A read-only plan mode can help separate exploration from implementation, but tool enforcement still depends on the selected product and environment.

Record the executable plan

Keep each step small enough to verify and large enough to deliver an observable result. Attach an owner, dependency, evidence requirement, and rollback or recovery path to every material change.

Outcome: [USER-VISIBLE RESULT]
In scope: [FEATURES, PATHS, SYSTEMS]
Out of scope: [EXPLICIT EXCLUSIONS]
Acceptance criteria: [TESTABLE CRITERION -> EVIDENCE]
Baseline: [BRANCH, VERSION, CURRENT CHECK RESULTS]
Assumptions: [ASSUMPTION -> VALIDATION]

Steps:
1. [BOUNDED STEP]
   Owner: [HUMAN | AGENT ROLE]
   Depends on: [NONE | STEP ID]
   Files/systems: [TARGETS]
   Verification: [COMMAND, REVIEW, OR OBSERVATION]
   Recovery: [REVERT, RESTORE, OR COMPENSATING ACTION]

Approval gates: [ACTION -> AUTHORIZED APPROVER]
External writes: [SYSTEM -> INTENDED CHANGE]
Risks and mitigations: [RISK -> CONTROL]
Definition of done: [ALL REQUIRED EVIDENCE AND RELEASE STATE]

Expected evidence and verification

The expected evidence is a plan that maps every acceptance criterion to one or more steps and maps every step to a verification result, without silently expanding scope.

Review the plan against repository instructions and the actual permission boundary. Confirm that destructive, security-sensitive, production, publishing, and external communication actions have explicit gates. Update the plan when evidence disproves an assumption; preserve the decision and reason instead of silently rewriting history.