| name | execute-from-plan |
| description | Resume and execute plan-driven implementation work with a durable execution-state ledger. |
Execute From Plan
Use this skill when a durable plan should be the execution-control source for implementation and cross-session progress.
Contract
Scope boundary:
- For multi-step or cross-session implementation, prefer a plan as the
execution-control source.
- For plan execution, the default target scope is the whole plan when the user
provides no task, sprint, phase, priority, or stop-condition selector.
- Target scope and execution window are distinct. The agent may propose a staged
execution window only after a suitability assessment and user confirmation.
- Source docs such as
*-discussion-source.md and *-review-source.md should
normally be paired with a sibling plan that links them under Read First.
- Direct source-doc execution is allowed only for bounded single-step work, or
when the user explicitly accepts a direct-execution waiver and that waiver is
recorded in execution state.
Prereqs:
- User explicitly asks to start, continue, resume, or execute work from a durable plan or bounded source doc.
- Execution source exists and is ready enough to proceed, or gaps can be recorded safely before asking for the minimum clarification.
- Target workspace is available, required project preflight has passed, and project rules allow edits.
- A progress record exists or can be created as an execution-state document.
Inputs:
- Execution source path, preferably an execution-ready plan. A source doc such
as an implementation handoff, improvement record, or equivalent project doc is
accepted only when direct execution is bounded or explicitly waived.
- Optional execution-state path; when absent, use the source document's
Execution section or create a sibling
<source-slug>-execution-state.md.
- Optional task, phase, sprint, or priority selector. Selectors intentionally
narrow the target scope; when absent, use the whole plan as the target scope.
- Optional validation commands, scope limits, branch/commit policy, linked issue/PR, and user-provided stop conditions.
Outputs:
- Scoped code, docs, config, test, or workflow changes for the target scope,
continuing through unblocked plan tasks until the target is complete, blocked,
unsafe, validation fails, or a confirmed checkpoint is reached.
- Created or updated execution-state document with current status, task ledger, validation evidence, blockers, and append-only session log.
- Source document updates only when durable facts, decisions, acceptance criteria, execution-state links, or guardrails changed.
- A concise final response summarizing completed work, blocked work, files changed, validation, and next step.
Exit codes:
- N/A (multi-step workflow skill)
Failure modes:
- Source document is missing, unreadable, or does not contain enough execution contract to proceed safely.
- Required task, acceptance criteria, validation plan, or guardrail is ambiguous enough that execution would risk the wrong change.
- Execution state conflicts with the work tree, branch, or source document; inspect and reconcile before editing.
- Work tree contains unrelated dirty changes that overlap the selected task; preserve user changes and ask if needed.
- Validation fails and cannot be fixed within the selected task scope; record the failure and stop without claiming completion.
Execution-Ready Source Contract
Accepted source documents include:
discussion-to-implementation-doc implementation handoffs.
review-to-improvement-doc improvement records with executable backlog.
create-plan or create-dispatch-plan plans.
- Hand-written project docs that carry the same execution contract.
Plan-created source docs under docs/plans/<slug>/ are read-first inputs, not
the preferred execution-control artifact. If a sibling <slug>-plan.md exists,
execute from that plan. If no sibling plan exists and the work is multi-step,
create or request a plan before editing production files.
Required source content:
- Goal or purpose.
- Scope and non-scope, or an equivalent boundary.
- Task ledger, backlog, sprint, phase, or next-task source.
- Acceptance criteria.
- Validation plan.
- Guardrails and known risks.
- Open questions or blockers.
- Execution-state path or enough context to create one.
Do not treat review-evidence.json as the primary source document. Link it as evidence from an improvement record or execution state.
Workflow
-
Resolve and classify the source
- Read the source document first.
- Classify it as
implementation-handoff, improvement-record, plan, or other-execution-doc.
- If the source is a plan-created
*-discussion-source.md or
*-review-source.md, look for a sibling <slug>-plan.md and use that plan
as the execution source when it exists.
- If the source doc is not paired with a plan and the requested work needs
sequencing, dependencies, or multiple tasks, stop and use
create-plan or
create-dispatch-plan instead of starting edits.
- Follow project preflight before reading or editing additional files.
- If the document references
Read First files, read only the files needed for the selected task.
-
Verify execution readiness
- Check for the required source content listed above.
- For direct source-doc execution, verify the work is a bounded single-step
change or record an explicit direct-execution waiver in execution state
using the stable label
Direct source-doc execution waiver.
- Determine the target scope from the user request. If no selector was
provided, record
whole plan as the target scope.
- Assess whether the target scope can be executed in the current run. Consider
plan size, dependency graph, file overlap, dirty worktree overlap,
validation availability, credentials, migrations, external side effects,
rollback risk, and expected runtime.
- If whole-plan execution is unsuitable for one run, summarize the risks and
ask the user to confirm either whole-plan target with staged checkpoints or
a narrower execution window with acceptance gates.
- If fields are missing but the next safe step is obvious, record assumptions in execution state and continue.
- If missing context could change scope, safety, acceptance, or reversibility, ask the minimum clarification and stop.
-
Establish execution state
- Use the source document's execution-state link when present.
- If absent, create a sibling state document named
<source-slug>-execution-state.md and add a link from the source document.
- For plan-created bundles under
docs/plans/<slug>/, prefer
docs/plans/<slug>/<slug>-execution-state.md.
- Keep top sections editable for fast resume.
- Keep
Session Log append-only so prior decisions and validation evidence remain traceable.
-
Resume the target scope
- Inspect
Current State, Target scope, Execution window, Task Ledger,
blockers, and the work tree.
- Pick the next unblocked item inside the target scope and execution window.
- Continue through unblocked items until the target scope is complete,
blocked, unsafe, validation fails, or a confirmed checkpoint is reached.
- For production behavior changes, obtain failing-test evidence or record an explicit waiver before editing production behavior.
- For browser-facing acceptance criteria, use
browser-qa when UI, layout,
JavaScript, DOM interaction, authenticated browser state, console output,
network behavior, screenshots, or deploy-preview inspection matters.
Record browser QA artifact paths, including browser-session.json and
browser-session verify outcome, in execution-state validation evidence
instead of copying large reports into the plan.
- Do not spawn subagents unless the user explicitly requested a delegation mode.
-
Execute within scope
- Read relevant code/docs/tests before editing.
- Make the smallest changes that satisfy the selected task.
- Update the task ledger as tasks move from
todo to in-progress, done, blocked, or accepted-risk.
- If a durable requirement, decision, acceptance criterion, or guardrail changes, update the source document as well.
-
Validate
- Run the source document's validation plan, or the closest project-defined validation for the selected change.
- Record command, result, summary, and artifacts in the execution state.
- If validation fails, fix within scope or mark the task blocked with concrete evidence.
-
Persist and report
- Update
Current State with status, target scope, execution window,
current/next task, last updated timestamp, branch/commit when relevant, and
blocker summary.
- Append a session log entry with read files, changed files, validation, blockers, and next step.
- Final response should link the source document and execution state, summarize work done, and state whether execution can continue.
Execution State Template
# <Topic> Execution State
## Current State
- Status: not-started | in-progress | blocked | validating | complete
- Target scope: whole plan | <task/phase/sprint/priority selector>
- Execution window: whole plan | <confirmed staged window>
- Staged execution confirmation: not applicable | pending | confirmed(<summary>)
- Current task: <task id or short description>
- Next task: <task id or short description>
- Last updated: YYYY-MM-DD HH:mm TZ
- Branch/commit: <optional>
- Source document: `<path>`
- Direct source-doc execution waiver: `not applicable`
## Task Ledger
| ID | Status | Task | Evidence | Notes |
| --- | --- | --- | --- | --- |
| T1 | todo | <task> | <command/path/commit> | <notes> |
## Validation
| Command | Status | Summary | Artifact |
| --- | --- | --- | --- |
| `<command>` | pass/fail/skipped | <summary> | <path> |
## Blockers
- <open question, missing access, failing check, or overlap with unrelated work>
## Session Log
### YYYY-MM-DD HH:mm TZ
- Read:
- Changed:
- Validated:
- Blocked by:
- Next:
Relationship To Nearby Skills
discussion-to-implementation-doc: creates implementation handoffs that can become execution-ready sources.
review-to-improvement-doc: creates improvement records; use this skill after the record has executable backlog and validation gates.
create-plan / create-dispatch-plan: create execution-ready plans; use this skill to resume multi-step implementation from those plans.
execute-plan-parallel: use only when the user explicitly wants plan execution through parallel subagents.
durable-artifact-cleanup: use after execution is complete and the source/progress docs are no longer needed as maintained records.
handoff-session-prompt: use when a fresh session prompt is needed; point it at the source document and execution state instead of
copying the full progress record.