paulnsorensen/extract

Todoist Extract

Identify tasks that aren't really tasks — they're reference material, ideas, bookmarks, or notes masquerading as action items. Extract them into structured markdown the user can save to Dropbox (or wherever they store reference material).

Why This Matters

Todoist is an action system. When it fills up with reference material ("look into X", bookmarked URLs, "idea: Y"), it dilutes the signal. Actionable tasks compete with passive notes for attention. Extracting reference material makes the task list honest — everything left is something to do.

Flow

Step 1: Scan for Candidates

Spawn todoist-fetch to retrieve all tasks in a fresh context window:

Agent(subagent_type: "todoist-fetch", prompt: "Fetch all tasks broadly (all projects, or user-specified scope). Include task IDs, titles, projects, priorities, due dates, creation dates, descriptions, labels, and any URLs in descriptions. Flag: tasks with no verb in title, URL-only tasks, tasks 90+ days old with no due date.")

Score each returned task for "reference-ness":

Detection patterns (a task matching 2+ patterns is a strong candidate):

| Pattern | Signal | Example | |---------|--------|---------| | No verb | Not actionable | Kubernetes best practices | | URL-only | Bookmark | https://example.com/article | | "look into", "check out", "explore" | Passive curiosity | Look into Rust async patterns | | "remember", "note:", "idea:" | Note, not task | Remember: API rate limit is 100/min | | "maybe", "someday", "consider" | Speculative | Maybe try Obsidian for notes | | 90+ days old, no due date | Dormant | Any old undated task | | Project is "Ideas" or "Someday" | Already filed as non-action | — |

Step 2: Present Candidates

Group by type and present:

## Reference Material Candidates

### Bookmarks (X tasks)
1. "https://example.com/rust-patterns" — in project "Learning", created 4 months ago
2. "Check out htmx docs" — in project "Ideas", created 2 months ago

### Notes & Ideas (X tasks)
3. "Idea: build a CLI for todoist cleanup" — in Inbox, created 3 months ago
4. "Remember: work VPN requires cert renewal every 90 days" — in "Work"

### Dormant (X tasks, 90+ days, no date)
5. "Research home solar panel options" — in "Home", created 6 months ago

Extract all of these? Or review one at a time? (A/R)

Use AskUserQuestion. If (R), walk through each and let the user keep or extract.

Step 3: Format as Markdown

Group extracted tasks into a structured reference document:

# Extracted from Todoist — [date]

## Bookmarks
- [Rust async patterns](https://example.com/rust-patterns)
  - Source project: Learning
  - Originally created: 2025-10-15

- htmx docs — check out the getting started guide
  - Source project: Ideas
  - Originally created: 2026-02-01

## Ideas
- Build a CLI for todoist cleanup
  - Context: [task description if any]
  - Originally created: 2026-01-15

## Notes
- Work VPN requires cert renewal every 90 days
  - Source project: Work

## To Research (someday)
- Home solar panel options
  - Source project: Home
  - Originally created: 2025-10-01

Step 4: Save the Document

Write the markdown file to a temp location:

$TMPDIR/todoist-extract-[date].md

Tell the user where it is:

Saved to: $TMPDIR/todoist-extract-2026-04-09.md

Move this file to your preferred reference storage (Dropbox, Obsidian, etc.).
When you confirm it's saved, I'll remove these tasks from Todoist.

Step 5: Clean Up (with confirmation)

Only after the user confirms they've saved the file, remove extracted tasks from Todoist.

Use AskUserQuestion: "Have you saved the reference file? Ready to remove these X tasks from Todoist? (y/n)"

If yes, run through the write pipeline (distill → scribe → QA):

1. Validate reasoning — spawn todoist-distill:

Agent(subagent_type: "todoist-distill", prompt: "Validate deletion of these extracted reference tasks. User confirmed file is saved. Tasks: [task IDs and titles]")

2. Format commands — spawn todoist-scribe:

Agent(subagent_type: "todoist-scribe", prompt: "Format delete-object commands for these extracted tasks: [distill's validated plan]")

3. Verify and execute — spawn todoist-qa:

Agent(subagent_type: "todoist-qa", prompt: "Verify and execute: [scribe's formatted commands]. Original intent: delete extracted reference tasks after user confirmed save.")

Step 6: Summary

## Extraction Complete

| Category | Count |
|----------|-------|
| Bookmarks | X |
| Ideas | X |
| Notes | X |
| To Research | X |
| **Total extracted** | **X** |

Saved to: [path]
Removed from Todoist: X tasks

Research Integration

If the user wants to evaluate whether a candidate is truly reference material vs. actionable:

Skill(skill="todoist-flow:research", args="Is '[task title]' something actionable or reference material? Context: [description]. Help decide if this belongs in a task list or a reference doc.")

Key Principles

• Conservative extraction — when in doubt, keep the task. It's better to leave an ambiguous item than to extract something the user actually needs to act on.
• User confirms before deletion — never remove tasks until the user confirms they've saved the reference file. Data loss from premature deletion is unrecoverable.
• Grouping adds value — random tasks extracted into a flat list aren't useful. Group by type (bookmarks, ideas, notes, research topics) so the reference doc has structure.
• Don't over-process — the markdown output should be clean and readable, not a database export. Include enough context to be useful, not every metadata field Todoist has.