Task

Task

$ARGUMENTS

Your Instructions

1. Use Existing Plan or Create Task Plan

If a PLAN file was provided or mentioned (e.g., from a prior /plan command) with structured phases:

• Use the existing PLAN file as your working document—do NOT create a separate TASK file
• Preserve ALL existing content—edit in place to add tracking, never rewrite from scratch
• Convert implementation bullet points to checkboxes for tracking (e.g., - Item → - [ ] Item)
• Add new checklist items or sections if gaps become apparent
• Continue to add items as needed during execution
• Update status markers in place as you progress

> NEVER REPLACE A PLAN FILE. If a plan file exists, keep ALL existing content intact—goals, file references, verification details, context, everything. You MUST edit in place: add checkboxes, append new tasks, update status markers, extend sections. But never delete, summarize, condense, or rewrite existing content. The plan is the user's work product. > > Common violation: Replacing a plan section's prose, context, and details with bare checkboxes. The checkboxes go ALONGSIDE the existing content, not instead of it. If your edit removes more text than it adds, you are doing it wrong!

If no suitable PLAN file exists, create a markdown file TASK_[descriptor]_[timestamp].md containing:

• Task description
• Checklist of specific steps/subtasks
• Success criteria
• Any dependencies or prerequisites

KEEP IT BRIEF AND TO THE POINT:

• Each checklist item should be one concise line
• Link to relevant files rather than explaining details: [config](./src/config.js)
• No verbose descriptions - this is a tracking tool, not documentation

Format checklist items as:

• [ ] Not started
• [~] In progress
• [x] Complete
• [!] Blocked or needs attention

2. Execute While Updating

CRITICAL: You MUST update the checklist file after EVERY significant action:

• Mark items as [~] when starting them
• Mark items as [x] when completing them
• Add new subtasks if discovered during work
• Note any blockers or issues with [!]
• Keep any notes extremely brief or link to files

3. Work Pattern

Follow this pattern throughout:

1. Update checklist → Do work → Update checklist → Repeat
2. The file is your live working document
3. Anyone should be able to see current progress by reading the file

4. Completion Expectations

Complete the full task/plan. Do not arbitrarily skip items, defer work, or stop halfway. Minor deviations and unexpected issues are normal—handle them and continue. Only stop early if an overwhelming reason emerges during implementation (e.g., core assumptions are fundamentally wrong, blocking dependencies are unresolvable).

When finished, ensure the file shows:

• Final status of all items
• Any unresolved issues (with explanation if incomplete)
• Brief summary of completion

5. Sub-Agent Usage

If deploying sub-agents for subtasks:

• Update checklist to [~] before dispatching
• Sub-agent updates checklist to [x] or [!] upon completion
• Use Opus for complex subtasks, Sonnet for standard implementation (Haiku only for pure read-only information gathering with no interpretation needed—or just do trivial tasks directly)

6. Verification

After completing all checklist items, run /doublecheck to verify the work before marking the task fully complete.

7. Task File Disposition

Apply only to TASK_*.md files created by this skill, never to existing PLAN files.

After /doublecheck passes:

• Delete the TASK file if it contains only transient execution tracking, routine verification, and a completion summary.
• If it records durable findings, decisions, constraints, or follow-ups, propose an appropriate destination based on repository conventions.
• After approval, preserve the information there and delete the TASK file, unless archiving the file itself is appropriate.
• Keep the file while work is incomplete or blocked.

Remember: The checklist file is your required working tracker. Keep entries brief, link to files for details, and update continuously as you progress.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: lacerbi
• Source: lacerbi/dotclaude
• License: MIT

Install and usage instructions live in the source repository linked above.