npx agentspec-cli install 2da5d3a0-cb7d-4d1c-b719-f027854e8cd7 --framework ClaudeCode→ .claude/skills/ct/SKILL.md
# Plan implementation Turn the agreed outcome into a plan that a fresh implementer can use. The normal output is a short section of the existing task record. Depth follows the work, not a fixed document schema. ## 1. Read the task and relevant code Resolve the task with [shared task context](../setup/references/task-context.md). Read its agreed scope, decisions and relevant linked discovery, product or prototype artifacts. Reuse an existing plan; create the minimum record when no task exists. Inspect the affected code, closest existing pattern and test/command sources. Reuse upstream findings that still match the current code and requirements; investigate only missing or changed evidence. Product documents and a separate discovery interview are not prerequisites when the requested behavior and boundaries are already clear. ## 2. Choose the approach and resolve material gaps Choose routine technical details from repository patterns and constraints. Record consequential choices with a short reason. Ask the user only when an unresolved choice materially changes behavior, scope, an external contract, or the safety/reversibility of the work and cannot be settled from the supplied decisions or repository evidence. Keep settled answers settled. Raise new product ideas as follow-ups. Save a genuine blocker and its impact; continue independent planning where possible, but mark the plan ready only when the missing decision no longer blocks implementation. Use /nf when the objective needs actual discovery, not merely because a template field is empty. ## 3. Save the smallest executable plan Define verification before deriving the steps, following the project's testing strategy. Update the task's existing plan section with: - **Changes:** the chosen approach and affected files/modules, linked to the agreed outcome. - **Steps:** concrete actions in dependency order; identify new files as proposed paths. - **Verification:** how each required behavior will be checked, using the project's actual commands and working directories. Distinguish planned checks from executed results. - **Risks and decisions:** only material constraints, unresolved blockers and trade-offs. Reuse acceptance criteria and source IDs where they exist. Connect requirements to steps and checks in this one plan; plain-language references are enough. A separate REQ/TEST matrix or Given/When/Then catalogue is useful only when it makes complex coverage easier to assess. Preserve existing documents and their terminology. For migrations, permissions, cross-system contracts, difficult coverage or phased delivery, read [detailed planning](references/decomposition-guide.md) and add the applicable detail. A separate decomposition is warranted when it needs independent review, no longer fits the task record clearly, or an existing consumer requires it. The detailed template is optional for other tasks. ## 4. Check readiness and hand off Make one final check: the plan covers the agreed outcome, follows inspected code, orders dependencies correctly, has suitable verification, and exposes real blockers. Correct gaps in place. Independent review and splitting follow the detailed guide only when justified by risk or explicitly requested; a routine plan needs neither. Update the task's existing current status and next action in place; retain old states only as dated history. Link the active plan section or file. Report the plan location and any material open decision. Hand a ready plan to `/si [task-path]`; continue implementation when that scope is already authorized. Planning readiness does not claim that tests ran or that a prototype, split or implementation was approved.
Built on the Agent Skills standard
create-implementation-plan
Create a new implementation plan file for new features, refactoring existing code or upgrading packages, design, architecture or infrastructure.
by AgentHub Crawler
writing-plans
Use when you have a spec or requirements for a multi-step task, before touching code
by AgentHub Crawler
simple-task
Implement simple features with best practices. Use when adding small features, bug fixes, or quick enhancements without the full SDD workflow. Invoked via /simple-task <description>.
by AgentHub Crawler
Sign in to leave a comment.
executing-plans
Use when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available
by AgentHub Crawler