Lightweight Task Workflow Skill for Claude
FOLLOW THE STATE MACHINE IN SKILL.MD. When user says 'continue': (1) FIRST: Run pwd, (2) Announce STATE: CHECK_STATUS, (3) Read .claude/session.md to check Status field, (4) Route based on Status. NEVER auto-advance tasks. NEVER use TodoWrite. NEVER create git commits.
Install
one command, takes just this skill from the repositorynpx skills add https://github.com/NTCoding/claude-skillz --skill lightweight-task-workflowThe instruction itself
20 sections, as written by the authorLightweight Task Workflow
🚨 CRITICAL: YOU MUST FOLLOW THE STATE MACHINE BELOW 🚨
🚨 EVERY SINGLE MESSAGE MUST START WITH: 🔵 STATE: [STATE_NAME] 🚨
NOT JUST THE FIRST MESSAGE. EVERY. SINGLE. MESSAGE.
When you read a file - prefix with state.
When you run a command - prefix with state.
When you explain something - prefix with state.
When you ask a question - prefix with state.
Example:
🔵 STATE: WORKING
Reading requirements.md...
🔵 STATE: WORKING
I can see the requirements specify...
🔵 STATE: WORKING
Now running tests...
🔵 STATE: WORKING
Test results show...
This skill is a persistent todo list based on 3 files in .claude/: tasks.md (checklist), requirements.md (specs), session.md (current state).
When user says "continue", you MUST:
- Run
pwdto check current working directory - Announce
🔵 STATE: CHECK_STATUS - Read
.claude/session.mdfrom the current project directory - Follow the state machine below based on the Status field
STATE MACHINE:
user: "continue"
↓
┌────────────────┐
┌───│ CHECK_STATUS │←──────────┬──────────┐
│ │ Read session.md│ │ │
│ └────────┬───────┘ │ │
│ │ │ │
Status= │ │ Status= │ │
"Complete" │ │ "in progress" │ │
│ │ │ │
↓ ↓ │ │
┌───────────┐ ┌──────────────┐ │ │
│ AWAITING_ │ │ WORKING │←────┐ │ │
│ COMMIT │ │ │ │ │ │
│ │ │ Read: │ │ │ │
│ Ask │ │ requirements │ │ │ │
│ permission│ │ tasks.md │ │ │ │
│ STOP │ │ │ │ │ │
└─────┬─────┘ │ Write: │ │ │ │
│ │ session.md │ │ │ │
user: yes │ └──────┬───────┘ │ │ │
│ │ │ │ │
│ │ task done │ │ │
│ │ │ │ │
│ ↓ │ │ │
│ ┌──────────────┐ │ │ │
│ │ VERIFY │ │ │ │
│ │ │ │ │ │
│ │ Run steps │ │ │ │
│ │ from │─────┘ │ │
│ │ requirements │ fail │ │
│ └──────┬───────┘ │ │
│ │ │ │
│ │ pass │ │
│ │ │ │
│ ↓ │ │
│ ┌──────────────┐ │ │
│ │ COMPLETE │ │ │
│ │ │ │ │
│ │ Write: │ │ │
│ │ session.md │ │ │
│ │ Status= │─────────┘ │
│ │ "Complete" │ │
│ └──────────────┘ │
│ │
↓ │
┌──────────────────┐ │
│ MARK_TASK_ │ │
│ COMPLETE │ │
│ │ │
│ Write: tasks [x] │ │
│ Write: session.md│──────────────────────────────┘
│ (next task) │
└──────────────────┘
🚨 STATE DEFINITIONS - FOLLOW EXACTLY 🚨
CHECK_STATUS:
ACTIONS:
1. Run pwd
2. Read .claude/session.md
3. Look at Status field
4. IF Status="Complete" OR "ready to commit" → Go to AWAITING_COMMIT
5. IF Status="in progress" OR missing → Go to WORKING
DO NOT: Read other files, launch agents, do anything except route
IF ERROR: STOP and tell user what failed
AWAITING_COMMIT:
ACTIONS:
1. Say: "Task X is complete. May I mark Task X as complete in tasks.md?"
2. STOP - wait for user response
3. IF user says yes → Go to MARK_TASK_COMPLETE
4. IF user says no → STOP, await further instruction
DO NOT: Read files, launch agents, work on next task, do anything except ask permission and STOP
IF ERROR: STOP and tell user what failed
MARK_TASK_COMPLETE:
ACTIONS:
1. Write tasks.md: Change [ ] to [x] for current task
2. Write session.md: Update to next task with Status="in progress"
3. Go to CHECK_STATUS
DO NOT: Read other files, launch agents, research next task
IF ERROR (e.g., plan mode, can't write): Say "I cannot edit files: [reason]" and STOP
NEVER try alternative actions if write fails
WORKING:
REMINDER: EVERY message in this state must start with: 🔵 STATE: WORKING
ACTIONS:
1. Read requirements.md
2. Read tasks.md
3. Work on current task
4. Update session.md after TDD cycles
5. When task done → Go to VERIFY
EVERY message you send while WORKING must have the state prefix.
When you read a file → prefix with state
When you run tests → prefix with state
When you explain results → prefix with state
DO NOT: Skip to next task, work on multiple tasks
IF ERROR: Document in session.md as blocker, STOP
VERIFY:
REMINDER: EVERY message in this state must start with: 🔵 STATE: VERIFY
ACTIONS:
1. Read Verification section from requirements.md
2. Run all verification commands
3. IF all pass → Go to COMPLETE
4. IF any fail → Go to WORKING (treat as blocker)
EVERY message you send while VERIFYING must have the state prefix.
DO NOT: Skip verification, claim complete without running checks
IF ERROR running verification: STOP and tell user
COMPLETE:
ACTIONS:
1. Write session.md: Set Status="Complete"
2. Go to CHECK_STATUS
DO NOT: Read files, launch agents, ask permission (that happens in AWAITING_COMMIT)
IF ERROR writing: STOP and tell user
CRITICAL: State Announcements
ALL messages MUST be prefixed with your current state.
Format:
**🔵 STATE: [STATE_NAME]**
[Your message here]
When transitioning:
**🟢 TRANSITION: [STATE_A] → [STATE_B]**
Example:
**🔵 STATE: CHECK_STATUS**
Reading session.md to check current task status...
**🟢 TRANSITION: CHECK_STATUS → AWAITING_COMMIT**
**🔵 STATE: AWAITING_COMMIT**
Task 2 is complete and ready for you to commit. May I mark Task 2 as complete in tasks.md?
When to Use This Skill
Activate when the user:
- Says "create a plan", "setup tasks", "new task list"
- Says "continue", "continue plan", "resume work", "where were we"
- Is working on multi-step projects that span multiple sessions
⚠️ CRITICAL: Task Management System
THIS SKILL REPLACES Claude Code's built-in TodoWrite functionality.
NEVER use the following tools:
- ❌ TodoWrite
- ❌ TodoRead
- ❌ Any built-in todo/task tracking features
ALWAYS use this skill's files instead:
- ✅
.claude/tasks.mdfor task checklists - ✅
.claude/requirements.mdfor plans and implementation specs - ✅
.claude/session.mdfor session context and recovery
Why this matters: Using TodoWrite creates workflow conflicts. The built-in todo system stores tasks in internal state (not visible as files), causing the plan to be lost in chat history instead of persisted in .claude/requirements.md. This prevents session continuity and defeats the purpose of this skill.
If you find yourself wanting to use TodoWrite, STOP and use this skill's files instead.
Files This Skill Manages
IMPORTANT PATH GUIDANCE:
- This skill's definition files (SKILL.md, CLAUDE.md, README.md) live in
~/.claude/skills/lightweight-task-workflow/ - The task files are created in THE PROJECT'S .claude/ directory, NOT the skill directory
- Example: If working on project
<project-root>/, task files go in<project-root>/.claude/ - Always use relative paths from the project root:
./.claude/tasks.md,./.claude/requirements.md,./.claude/session.md - NEVER read from
~/.claude/skills/lightweight-task-workflow/tasks.md(that's the skill directory, not the project directory)
.claude/tasks.md - the task checklist (in the PROJECT directory)
- [ ] Task 1: Extract UserService
- [x] Task 2: Add tests
.claude/requirements.md - implementation specs and guidelines
## Global Guidelines
- Preserve existing API contracts - no breaking changes
- Add logging for error cases
- Follow repository's existing code style
## Verification & Definition of Done
Before marking any task complete, the following must pass:
- `npm test` - all tests must pass
- `npm run lint` - no lint errors
- `npm run build` - build must succeed
## Task 1: Extract UserService
- Move all user-related methods from AppService to new UserService
- Keep existing method signatures for backward compatibility
- Update dependency injection in app.module.ts
## Task 2: Add tests
- Cover happy path and error cases
- Include edge cases for null/undefined inputs
- Mock external dependencies
.claude/session.md - session recovery context
**Current Task:** Task 3
**Status:** in progress
## What's Done
- Task 1: Extracted UserService (commit a1b2c3d)
- Task 2: Added tests (commit e4f5g6h)
## Next Steps
1. Finish Task 3: Update documentation
## Context
- Using yarn for builds
- Commands: ./verify.sh
Behavior
When User Says "Create a Plan" or "Setup Tasks"
FIRST: Remember you are NOT using TodoWrite. Use this skill's .claude/ files exclusively.
- Ask user to describe their tasks
- Ask user about implementation approach:
- "What requirements or guidelines should I know for implementing these tasks?"
- "Are there testing/quality standards I should follow?"
- "Any architectural constraints or patterns to follow?"
- "What verification must pass before marking a task complete?" (e.g., npm test, ./verify.sh, build, lint, manual review)
- Capture their answers - these become the implementation specs
- Create
.claude/tasks.mdwith numbered checklist - format:- [ ] Task 1: [exact user wording],- [ ] Task 2: [exact user wording], etc. - Create
.claude/requirements.mdwith: - Global Guidelines section (testing, commands, patterns)
- Verification & Definition of Done section (commands/checks that must pass)
- Per-task requirements (one section per task with specs)
- Create
.claude/session.mdinitialized to Task 1 with Status="in progress" - Confirm setup complete
After setup: You enter the state machine at CHECK_STATUS with session.md showing Task 1 Status="in progress"
When User Says "Continue" or "Resume"
Start at CHECK_STATUS state: Read session.md and route based on Status field. Follow the state machine at the top of this file.
What to Track in requirements.md
Include: Global guidelines, Verification & Definition of Done (user can edit anytime), per-task requirements, learnings/edge cases discovered.
NOT for: Progress notes, debugging notes, code changes (that's git).
What to Track in session.md
Update at 4 triggers: (1) Start task, (2) End of TDD cycle (one line), (3) Hit blocker, (4) Complete task.
Include: Current task/status, completed tasks with commits, brief progress notes, blockers, next steps.
NOT for: Every change, file paths, verbose explanations.
Anti-Patterns: What NOT to Do
❌ WRONG: Investigating Codebase to Figure Out Progress
User: "continue"
Claude: *Reads tasks.md*
Claude: "Let me investigate the codebase to understand what's already been done"
Claude: *Searches through 10+ files, runs git log, checks test files*
Claude: *Wastes 2 minutes and 5000 tokens figuring out current state*
Claude: "I can see Task 1 was completed, let me start on Task 2..."
Problem: Wasted time and tokens. All that information was already in session.md.
✅ RIGHT: Reading session.md to Know Current State
User: "continue"
Claude: *Runs pwd*
Claude: "🔵 STATE: CHECK_STATUS"
Claude: *Reads session.md FIRST*
Claude: "Status shows 'in progress'. Routing to WORKING."
Claude: "🟢 TRANSITION: CHECK_STATUS → WORKING"
Claude: "🔵 STATE: WORKING"
Claude: "Continuing Task 2: Add email validation..."
Result: Instant context, no wasted time, exactly where to resume. That's the whole point of session.md.
❌ WRONG: Skipping Verification
Claude: *Completes task implementation*
Claude: "Task 1 is complete and ready for you to commit"
Claude: "May I mark this task as complete?"
*User commits and deploys*
*Build breaks in CI - test failures, lint errors discovered*
Problem: Introduced regressions, broken build, wasted time debugging issues that should have been caught before claiming "complete."
✅ RIGHT: Running Verification Before Completion
Claude: *Completes task implementation*
Claude: *Reads requirements.md Verification section*
Claude: *Runs npm test* → All pass ✅
Claude: *Runs npm run lint* → All pass ✅
Claude: *Runs npm run build* → Success ✅
Claude: *Updates session.md: "Task 1 complete - all verification passed"*
Claude: "Task 1 is complete, all verification passed (tests/lint/build), ready for you to commit"
Result: Confidence that task is truly complete, no regressions introduced, ready for production.
❌ WRONG: Creating Git Commits
Claude: *Completes task implementation*
Claude: *Runs git add .*
Claude: *Runs git commit -m "Add UTF-8 correction table"*
Claude: "I've committed the changes"
Problem: User loses control over commits - can't review changes, adjust commit message, or stage selectively.
✅ RIGHT: Handing Off for User to Commit
Claude: *Completes task implementation*
Claude: *Updates .claude/session.md: "Task 1 complete - UTF-8 correction ready for commit"*
Claude: "Task 1 is complete and ready for you to commit. The changes include..."
Claude: "May I mark this task as complete in tasks.md?"
Result: User reviews changes, creates commit with their preferred message and staging, maintains full git control.
❌ WRONG: Auto-Advancing to Next Task
Claude: "Task 1 is complete, verification passed, ready for you to commit"
Claude: "May I mark this task as complete?"
Claude: "Let me explore the codebase to understand Task 2: Add email validation..."
Claude: *Launches Plan agent for Task 2*
Problem: User doesn't have time to review Task 1, commit changes, or decide when to proceed. Claude rushes ahead without permission.
✅ RIGHT: Stopping After Task Complete
Claude: "Task 1 is complete, verification passed, ready for you to commit"
Claude: "May I mark this task as complete in tasks.md?"
Claude: *Waits for user response*
User: *Reviews changes, creates commit*
User: "continue"
Claude: *Reads session.md, sees Status="Complete", routes to AWAITING_COMMIT*
Claude: "Task 1 is complete. May I mark it [x]?"
User: "yes"
Claude: *Updates tasks.md, updates session.md to Task 2 Status="in progress"*
User: "continue"
Claude: *Reads session.md, sees Task 2 Status="in progress", routes to WORKING*
Result: User controls the pace, reviews and commits when ready, decides when to proceed to next task.
Troubleshooting: Common Path Mistakes
Symptom: "Error reading file" when trying to continue tasks
Likely cause: You're looking in the skill directory instead of the project directory
Fix:
- Run
pwdto check current working directory - Look for
.claude/subdirectory in the project root - Read from
./.claude/tasks.md, NOT~/.claude/skills/lightweight-task-workflow/tasks.md - Remember: skill definition ≠ task files
Example:
- ❌ WRONG: Reading
~/.claude/skills/lightweight-task-workflow/tasks.md(skill directory) - ✅ RIGHT: Reading
./.claude/tasks.mdor<project-root>/.claude/tasks.md(project directory)
Important Rules
- ALWAYS prefix EVERY SINGLE MESSAGE with your state - Not just when entering a state. EVERY message. When you read a file, when you run a command, when you explain something - ALL messages start with
🔵 STATE: [STATE_NAME] - ALWAYS run pwd first - Check current working directory before reading files
- ALWAYS follow the state machine - Start at CHECK_STATUS, route based on Status field from session.md
- NEVER use TodoWrite, TodoRead, or Claude Code's built-in todo features - This skill replaces them entirely
- NEVER create git commits - User handles all commits
- NEVER auto-advance to next task - STOP and wait for user
- ALWAYS run verification from requirements.md before claiming complete
- ALWAYS read task files from PROJECT's
.claude/directory, not skill directory - Preserve user's exact wording when creating tasks
- Always ask permission before marking tasks complete
- Update requirements.md when discovering new constraints
Other skills for the same job
different authors, same section of the catalogue
Complete development kit for Microsoft 365 Copilot declarative agents with three comprehensive workflows (basic, advanced, validation), TypeSpec support, and Microsoft 365 Agents Toolkit integration
Format and structurally validate local treatment-plan documentation after clinical decisions have already been supplied and verified by authorized licensed professionals. Use for source traceability, clinician-authored intervention records, goals and checkpoints, shared-decision records, reconciliation handoffs, and release gates—not for clinical decision-making.
> provider/change budget/修改卖家/修改预算/draft/草稿/我的任务/my tasks/what am I working on/关闭/取消任务/决策列表/decision list/指定服务商/browse (sender.role = COUNTERPARTY, not you); (3) literal "Read the okx-ai skill" (or legacy "Read the okx-agent-task skill") in the envelope.
Automate payer review of prior authorization (PA) requests. This skill should be used when users say "Review this PA request", "Process prior authorization for [procedure]", "Assess medical necessity", "Generate PA decision", or when processing clinical documentation for coverage policy validation and authorization decisions.
Expert in designing and building autonomous AI agents. Masters tool use, memory systems, planning strategies, and multi-agent orchestration.
Autonomous agents are AI systems that can independently decompose goals, plan actions, execute tools, and self-correct without constant human guidance. The challenge isn't making them capable - it's making them reliable. Every extra decision multiplies failure probability.
Orchestrates design workflows by routing work through brainstorming, multi-agent review, and execution readiness in the correct order.
Structured persuasion for tech leads, PMs, and founders—not activity logs. Five scenarios (kickoff, status update, wrap-up, investor pitch, solution selling) on one 5-part framework (Hook→Context→Proposal→Evidence→Ask). AI prompts for missing materials and audience context; pre-submit checklist. Claude Code plugin; Cursor, Codex, and chat via prompts.
How to use it
Take ntcoding/lightweight-task-workflow 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.