File-based planning workflows for Cowork — Prevent context loss on complex tasks.
Cowork launched Jan 12, 2026. This is the first community resource for using file-based planning to handle multi-step tasks without hitting context limits.
Cowork (and Claude Code) has a ~200K token context window. When you're working on complex tasks — organizing thousands of files, processing receipts, generating reports — you'll hit this limit mid-task.
What happens:
- Context fills up after 50-150 operations
- Cowork loses track of what it was doing
- You have to start over
- Progress is lost
The solution: File-based planning. Use markdown files as "working memory on disk."
Instead of keeping everything in context, Cowork writes progress to three persistent files:
- task_plan.md — What phases exist, what's complete, what's next
- progress_log.md — Detailed log of every action taken
- decisions.md — Why choices were made (prevents re-deciding)
When context fills up, you start a fresh Cowork session and tell it:
Read task_plan.md, progress_log.md, and decisions.md
Continue from Phase 3. Do NOT re-do completed work.
Cowork restores context from the files and picks up exactly where it left off.
Ready-to-use markdown templates for any task:
- task_plan.md — Phase breakdown, progress tracking, decisions
- progress_log.md — Session logs, file tracking, issues
- decisions.md — Rationale, naming conventions, user preferences
Complete workflow examples showing the planning pattern in action:
- 01-organize-downloads.md — Organize 6 months of downloads by type and date
- 02-expense-report.md — Create spreadsheet from receipt images with OCR
Each workflow includes:
- Initial prompt to give Cowork
- Expected phases with context risk warnings
- Planning file examples
- Session handoff instructions
- Expected outcomes
- Start Cowork with planning files:
I need help [your task description].
Please:
1. [What you want done]
2. [Any specific requirements]
Use these planning files to track progress:
- task_plan.md
- progress_log.md
- decisions.md
-
Cowork creates the planning files and works through phases
-
If context fills up (Cowork slows down or stops):
Start a fresh Cowork session and say:
Read task_plan.md, progress_log.md, and decisions.md
Continue from Phase [current phase]. Do NOT re-do completed work.
Check the workflows/ folder for ready-to-use prompts for:
- Organizing files
- Processing receipts
- Generating reports
- [More coming soon]
From Anthropic's engineering team (Feb 2024):
"We use markdown files as a form of working memory on disk. When context gets full, we have agents read these files to restore state."
This is the same pattern used internally by Anthropic engineers for complex agent tasks.
Key insight: Context is temporary. Files are permanent.
You: "Organize my 300 files"
Cowork: [Processes 100 files]
Cowork: [Context fills up]
Cowork: [Stops or slows down]
You: "Keep going?"
Cowork: "What files? I don't remember what we were doing."
You: "Organize my 300 files using planning files"
Cowork: [Creates task_plan.md, progress_log.md, decisions.md]
Cowork: [Processes 100 files, logs everything]
Cowork: "Context at 80%, recommend switching sessions"
You: [New session] "Read the planning files and continue"
Cowork: [Restores context] "Continuing from file 101..."
Cowork: [Completes all 300 files]
Use file-based planning for tasks that involve:
- 100+ operations (file moves, API calls, processing steps)
- Multiple phases (discovery, processing, verification)
- Data extraction (OCR, parsing, categorization)
- Long research (multi-source investigation)
- Complex builds (multiple dependencies, testing phases)
Don't use it for:
- Quick one-off tasks
- Simple questions
- Single file edits
- Ask for planning files upfront — Don't wait until context fills
- Check context every 50 operations — Watch for slowdowns
- Log decisions immediately — Don't rely on memory
- Use git checkpoints — Commit planning files after each phase
- Test resume prompts — Make sure files have enough detail
This repo launched the day after Cowork's announcement. We're just getting started.
Ways to contribute:
- Share your workflows — Open a PR with a new workflow example
- Improve templates — Suggest better structure or wording
- Report issues — What didn't work? What was confusing?
- Start discussions — Ideas for automation? Integration patterns?
Discussion topics we need to explore:
- Should Cowork auto-create planning files for tasks >X operations?
- Can we automate context checks?
- How do we handle file conflicts across sessions?
- Integration with existing project management tools?
Q: Does this work with Claude Code too? A: Yes! Claude Code and Cowork are built on the same SDK. These templates work for both.
Q: Can Cowork read these files automatically? A: Not yet. You need to tell Cowork to read them in your prompt. (But we're exploring automation.)
Q: What if I forget to switch sessions? A: Cowork will slow down or start losing track. The planning files act as a safety net — you can always recover.
Q: Do I need to learn markdown? A: No. Cowork creates and updates the files. You just need to read them occasionally.
Q: Can I modify the templates? A: Absolutely! They're starting points. Adapt them to your needs.
- Cowork announcement (Jan 12, 2026)
- Claude Agent SDK
- planning-with-files skill (Claude Code version)
MIT License — Use freely, share widely, improve together.
- Discussions: Share your workflows, ask questions, propose ideas
- Issues: Report problems or suggest improvements
- Twitter/X: @OthmanAdi (if you have updates)
Built by @OthmanAdi — First community resource for Cowork file-based planning.
🚀 Let's build the Cowork workflow library together.