Run deadline-bound or scope-sensitive agent tasks with explicit done conditions, protected verification time, evidence gates, and safe-stop rules. Use for rapid implementation, debugging, integration, delegated work, feasibility gates, or tasks where scope drift, false completion, repeated failure, or deadline overruns are material risks.
npx skills add https://github.com/wenjimods/timeboxed-execution --skill timeboxed-execution
Use timeboxes to make agent work bounded, visible, and verifiable. The goal is not to make the agent rush. The goal is to stop scope drift and repeated failure early while preserving enough budget to prove that the result works.
The governing contract is:
bounded goal + explicit done condition + evidence gate + deadline + stop rule
A partial result remains partial. A passed test is evidence for what it tested, not proof that the whole task is complete.
Use this skill when:
Do not add process to a genuinely simple task. For a small lookup, edit, or check, use:
one goal → shortest verified path → one check → stop
Percentages are defaults, not laws. The invariant is that acceptance receives protected time.
Use explicit states so speed cannot hide an incomplete result:
verified_complete — every declared exit condition passed with claim-matched evidence.partial_verified — some bounded outcomes passed, but the main goal or a material risk remains open.blocked — a named prerequisite or external decision prevents further valid progress.failed — the attempted result was checked and did not satisfy the declared exit condition.stopped_unverified — the deadline arrived before the result could be accepted; no success may be inferred.“Implemented,” “tests passed,” “UI updated,” and “works in the current runtime” are separate claims. Promote the task to verified_complete only when the evidence required by the original done condition exists.
When states overlap, classify the main declared goal first. Use verified_complete only when every exit condition passes; use partial_verified when at least one bounded outcome is accepted but the main goal remains incomplete, and name any blocker, failed check, or deadline separately. Use blocked, failed, or stopped_unverified when no bounded outcome has been accepted and that condition best explains why the goal did not complete.
Before acting, state or internally record:
Completion criterion: another agent could later classify an observed outcome as verified_complete, partial_verified, blocked, failed, or stopped_unverified using the contract and obtained evidence.
Resolve the smallest fact that can invalidate the whole plan:
If the prerequisite fails, stop at the gate. Do not compensate by building an unofficial detour unless that detour is separately approved.
Completion criterion: the prerequisite gate is explicitly Go or No-Go with evidence; any resulting task blocker is recorded separately. Go and No-Go are gate outcomes, not Completion States.
Create phases only for work that cannot safely fit one short loop. Each phase must leave a result that can be judged even if the project stops immediately afterward.
Good phase exits:
Bad phase labels:
These describe activity, not an accepted state.
Completion criterion: every phase has one outcome, one evidence gate, and one local deadline.
Work on the critical path only:
When a path fails and its prerequisites have not changed, do not repeat it. Record the blocker and switch strategy or stop.
Do not run a sleeping timer process and repeatedly poll it. Record the start and deadline once, work continuously, and check the clock at necessary phase boundaries or acceptance without polling.
Completion criterion: every action either reduces a named uncertainty or advances the defined artifact toward acceptance.
Delegation does not suspend the deadline.
A practical split for a 10-minute implementation phase is about 6 minutes for worker execution and 4 minutes for parent review, correction, and evidence. Adjust to task risk.
Completion criterion: the parent can still accept or reject the result before its own deadline.
At scope freeze:
At acceptance mode:
Report five items:
State: applicable Completion State
Completed: outcomes that satisfy the contract
Evidence: exact tests, probes, readbacks, or source checks
Blocked/unfinished: what is not accepted and why
Remaining: bounded next step and realistic time needed
Completion criterion: the report distinguishes verified completion from implementation progress.
At the hard deadline, stop all new implementation, retries, exploration, and non-safety external mutations. Perform only the bounded actions necessary to pause safely, make the state observable, and preserve handoff evidence.
Safe-stop actions may include:
Do not use safe-stop as hidden overtime for feature work or optional testing.
Request an extension only when:
Any extension must be explicit before the current hard deadline, approved where required, and establish a new hard deadline. Otherwise, apply safe-stop.
Use this format:
Current accepted state:
Remaining items:
Why the estimate was short:
Extension requested:
Acceptance reserved:
Risk if denied:
If the contract says a phase timeout stops the whole project, do not borrow later-phase time. If unused time may roll forward, do so only under the original project rule.
Completion criterion: work is stopped or resumed only under an explicit, bounded, and correctly approved rule.
Choose evidence that matches the claim:
| Claim | Minimum useful evidence |
|---|---|
| A bug is fixed | failing reproduction before, passing regression after |
| Declared code change satisfies its done condition | diff readback plus tests or probes that cover the declared done condition |
| Integration works | real boundary call and readback of the target state |
| UI changed backend state | real UI action plus backend or target readback |
| A setting affects the current runtime | one real subsequent operation through that runtime |
| A setting persists | restart or fresh-session readback when persistence is part of the claim |
| Performance improved | same-path before/after under materially comparable conditions; repeat or report variability when magnitude matters |
| Permission exists | current live capability check, not historical state |
| External write succeeded | read back the exact record or remote object |
| Research conclusion is current | authoritative source or sources appropriate to the claim, checked during the task |
Never use activity counts as a substitute for these gates.
Evaluate the workflow over several real tasks, not one successful run.
Track:
A good timebox may produce an honest No-Go or partial result. That is a quality improvement when it prevents unsafe changes, false claims, or hours of uncontrolled work.
For anonymized validation examples and a small replay rubric, read references/validated-patterns.md.
10. Tool-count optimization. Fewer calls are good only when correctness and evidence remain intact.
Take wenjimods/timeboxed-execution from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.