Finish Todos

• 🔧 tasks: Set up environments, install dependencies, run scripts, organize files
• 📝 tasks: Write drafts, create templates, generate documents
• 💻 tasks: Write code, fix bugs, build pipelines (if the codebase is local)
• 📊 tasks: Organize data, create charts, analyze results (if data is accessible)
• 📖 tasks: Summarize papers, literature review (can use web search)
• 🔧 system tasks: Run /weekly-plan, update Things/Notion

🟡 Claude Can Assist

• 💬 tasks: Draft messages/emails, but the user needs to send them
• 🧪 tasks: Write experiment scripts, but the user needs to run them on HPC
• 📝 writing tasks that require human judgment (e.g., articulating novel contributions in the method section)
• Tasks that require reviewing human work

🔴 Requires Human

• 💬 In-person meetings, discussions
• 💪 Workouts, life events
• Tasks requiring physical action
• Core research tasks requiring human creative decisions

Step 3: Present Categorized Results

Format:

🟢 I can do these for you:
1. 💻 [HarmonyAgent] Fix cSAMPLEID mapping bug
   → Plan: Read code to locate bug, fix and test
2. 📝 [RLVR] Write method section first draft
   → Plan: Draft based on existing notes, you review

🟡 I can assist with:
3. 💬 [ToMAS] Schedule sync with Peiwen
   → Plan: Draft the message for you, you send it
4. 🧪 [RLVR] Run SFT training
   → Plan: Write the launch script, you submit to HPC

🔴 You need to do these yourself:
5. 💪 Workout
6. 💬 Lab meeting

Step 4: Execute

Ask the user: "Which ones should I work on? You can pick by number, or say 'do all the green ones'."

After the user chooses:

• 🟢 tasks: Execute directly, report results when done
• 🟡 tasks: Complete the assistable portion, tell the user the remaining steps
• Completed tasks: Mark as done via Things URL Scheme (requires auth-token)
• Brief status report after each task is completed

Step 5: Wrap Up

After all selected tasks are done, summarize:

✅ Completed 3 tasks:
- 💻 Fixed cSAMPLEID mapping bug → Fixed and passed tests
- 📝 Wrote method section first draft → Draft at [location], needs your review
- 💬 Drafted message for Peiwen → Ready for you to send

Remaining Today tasks: 2 (require you to complete personally)

Decision Criteria

To determine whether Claude can complete a task, consider:

1. Is the data/code accessible? Local filesystem, GitHub repo → yes. Data on HPC → no (unless SSH is available).
2. Does it require human judgment? Pure execution → yes. Requires creative decisions → assist only.
3. Does it require human interaction? Sending messages/meetings → assist/no. Independent work → yes.
4. Does it require physical action? Workouts/running/in-person meetings → no.
5. Is there a matching skill? Supported by a skill (Calendar, Notion, Things, Gmail) → higher confidence.

Available Skills Arsenal

Currently available skills (will expand over time):

Important Notes

1. Confirm before executing: Show the plan for each task and wait for user confirmation before proceeding
2. Don't overpromise: If unsure whether a task can be completed, categorize as 🟡 rather than 🟢
3. Confirm before marking complete: Before marking a task as done in Things, show the user the result and get confirmation
4. Stay humble: Some tasks look doable but won't be done well (e.g., writing requiring deep domain knowledge) — be honest about it
5. Log learnings: If a new automatable pattern is discovered, suggest the user create a new skill