GTD Purge Skill

Systematically reduce task list bloat by walking through every active task and forcing a keep/drop/demote decision. Based on David Allen's principle that a long list isn't the problem—carrying un-renegotiated commitments is.

When to Use

• Context lists feel overwhelming (>15–20 active items in a single context)
• You're avoiding looking at your lists
• Weekly review keeps skipping the "cull" step
• You notice tasks that have been sitting untouched for weeks
• Total active tasks exceed what you can realistically engage with

Vault Operations

Use the /obsidian skill for all vault reads and writes. Load it before running any commands: → .claude/skills/obsidian/SKILL.md

Key queries for purge:

• Inbox: Dataview eval inbox query
• By context: Dataview eval tag query (e.g. @quick, @deep)
• All open tasks: tasks todo verbose format=json
• Someday/maybe: Dataview eval someday query
• Context distribution: Dataview eval context-counts query
• Stale projects: Dataview eval stale-projects query

Core Principles (GTD-Aligned)

1. Renegotiate, don't just delete. Every task was a commitment. Consciously decide: "Am I still committed to this?" If not, deliberately let it go. Allen calls this renegotiating an agreement with yourself.
2. Someday/Maybe is the pressure valve. If you're not actively committed but don't want to lose it, demote to ⏬ (lowest priority). This keeps active lists lean and trustworthy.
3. Merge duplicates and overlapping tasks. Capture happens fast; duplicates creep in. Combine them into one clear action.
4. Clarify vague tasks or kill them. If a task has sat for weeks and you still can't picture the next physical action, it's not actionable yet. Either rewrite it as a concrete action or drop/demote it.
5. Projects should only surface 1–2 next actions. If your list is bloated because project steps are all visible, ensure blocking chains (⛔/🆔) are correct so only the true next action is unblocked.
6. Completed work should be marked done. Anything you've already done but forgot to check off gets marked done now.
7. Stale recurring tasks need attention. If a recurring task lost its recurrence or is overdue, fix or retire it.

Workflow

Phase 1: Triage Scan

The agent scans all open tasks excluding ⏬ (lowest priority / Someday-Maybe) and groups them into triage categories:

When scanning, always filter out:

• Tasks with ⏬ priority (review these in someday purge mode instead)
• Completed tasks
• Tasks in Templates, Checklists folders
• Tasks with a scheduled date (⏳) don't need a context tag — they will surface when the date arrives. Skip them in orphan/no-context triage.

Use the /obsidian skill's Dataview eval queries for inbox, someday, tag, and all-tasks scans.

Phase 2: Context-by-Context Purge

Work through each context list one at a time. For each context, present all active tasks and ask the user to make a decision on each:

Decision options per task:

Priority ladder (Obsidian Tasks plugin — 6 levels):

🔺 highest → ⏫ high → 🔼 medium → (normal) → 🔽 low → ⏬ lowest

When demoting, move one step down by default — not straight to ⏬. For example: 🔼 → normal, normal → 🔽. Only demote to ⏬ if the user explicitly asks to send something to Someday/Maybe.

Note: To demote from 🔼/⏫/🔺 to normal, use the /obsidian skill's edit-by-match pattern and rewrite the line without the priority emoji.

Context processing order (heaviest lists first):

1. Orphaned daily note tasks (biggest source of bloat)
2. @read (reading lists grow fastest)
3. @quick
4. @deep
5. @batch
6. @pc (legacy — migrate or drop)
7. @home
8. @partner
9. @ai
10. @ponderables
11. @work
12. @out
13. Projects review (stale/stuck projects)

Phase 3: Project Pruning

For each active project:

1. Is this project still a current commitment? → If not, demote entire project to Someday/Maybe or archive to Completed Projects.
2. Does it have a clear, unblocked next action? → If not, define one or put project on hold.
3. Are the dependency chains correct? → Fix any broken ⛔/🆔 references.
4. Are there too many steps visible? → Ensure only 1–2 are unblocked.

Phase 4: Recurring Task Cleanup

Review GTD/Recurring.md:

1. Flag any recurring task that lost its recurrence pattern (like the Weekly Review task).
2. Identify recurring tasks with excessive completed-instance history — these are fine but note the file size.
3. Confirm all active recurrences still make sense.

Phase 5: Summary Report

After purging, produce a summary:

=== Purge Summary ===
Tasks before:  509
Tasks dropped:  XX
Tasks demoted:  XX (→ Someday/Maybe)
Tasks done:     XX
Tasks merged:   XX
Tasks rewritten: XX
Tasks after:   XXX
---
Active projects: XX (was XX)
Stale projects fixed: XX

IMPORTANT: Confirmation & Execution

Batch presentation with user control:

1. Present tasks in batches of 10 per context.
2. For each batch, show numbered tasks with a suggested action based on triage criteria:

   @read (42 tasks — suggesting cuts)
   
   1. Read Zhengdong Wang best facts article          → suggest: keep
   2. Read [The Shape of AI...]                       → suggest: keep (recent, high-quality)
   3. Read New Yorker: why AI didn't transform...     → suggest: demote ⏬ (5 weeks old)
   4. Read [Demystifying evals for AI agents]         → suggest: keep (@ai relevant)
   5. Read Substack article [no title, bare URL]      → suggest: drop (vague, 3 weeks)
   ...
   
   Accept all suggestions? Or type numbers to override (e.g., "3: keep, 5: keep")

3. The user can:
   • Accept all suggestions at once
   • Override specific items by number
   • Skip the entire context ("skip")
   • Stop the session ("stop")
4. After confirmation, execute all actions for that batch before moving to the next.
5. Never delete or demote without the user seeing the suggestion first.
6. Use --yes flag on tools only after the user has confirmed.

Suggestion Heuristics

The agent should use these heuristics when suggesting actions:

Operations

All operations use the /obsidian skill. File paths are vault-relative.

Important: Use the /obsidian skill for all modifications. Do not read files and edit them manually.

Scope: What Gets Purged

In scope (active commitments):

• All open tasks that do NOT have ⏬ (lowest priority)
• These are things you've implicitly said "I'm doing this" — the purge forces you to re-confirm or let go

Out of scope by default:

• Tasks with ⏬ priority (Someday/Maybe) — these are already parked. Purging them defeats the purpose of the pressure valve.
• Completed tasks ([x])
• Tasks in Templates, Checklists

Someday/Maybe purge (separate mode):

• Use someday mode to review the ⏬ backlog specifically
• This is a different mindset: not "am I committed?" but "do I still want this on my radar at all?"
• Heuristics differ: age matters less (they're meant to sit), but items that no longer spark any interest should be dropped
• Suggested cadence: quarterly, or when ⏬ list exceeds ~100 items

Session Management

• Full purge: Walk all active (non-⏬) contexts and projects. Takes 30–60 minutes. Do this quarterly or when the system feels untrusted.
• Quick purge: Target one or two contexts (e.g., just @read and daily notes). Takes 10–15 minutes. Good for weekly review add-on.
• Auto-suggest purge: Agent scans active (non-⏬) tasks only and presents those matching stale/vague/duplicate heuristics. Fastest — just review the flagged items.
• Someday purge: Review ⏬ items only. Different mindset — "do I still want this on my radar?" Drop anything that no longer interests you. Do this quarterly.

Ask the user which mode they want at the start of each session.

Coordination With Other Skills

• Clarify: Orphaned daily note tasks that survive the purge should be processed through Clarify to get proper context tags.
• Organize: Tasks that get rewritten or moved during purge should end up properly tagged and filed.
• Project: Projects identified during purge as stale get either revived (new next action) or archived.
• Review: The purge is a deeper version of the "cull" step in weekly review. Run purge when weekly review alone isn't enough.

Anti-Patterns to Watch For

• Purging too aggressively — If the user drops >50% of tasks, pause and check: are these genuinely not commitments, or is this avoidance? Some tasks are hard but important.
• Demoting everything — Someday/Maybe shouldn't become a second junk drawer. If it grows past ~100 items, it needs its own purge.
• Skipping the rewrite step — A vague task that gets "kept" will just sit there again. Push to rewrite or drop.
• Not fixing the capture habit — If daily notes keep accumulating unprocessed tasks, the real fix is more frequent clarify sessions, not more purges.

Example Session

Agent: Ready to purge. You have 509 open tasks. What mode?
  1. Full purge (all contexts + projects, ~45 min)
  2. Quick purge (pick 1-2 contexts, ~15 min)  
  3. Auto-suggest (only flagged items, ~10 min)

User: 3

Agent: Running triage scan...

Found 47 flagged items across your system:

=== Orphaned Daily Note Tasks (23 items) ===
Batch 1/3:

1. "rubber bands XL" (Jan 2, no context)           → suggest: drop
2. "cribbage @quick" (Jan 2)                        → suggest: keep
3. "[Cool astrophotographer] @quick" (Jan 3)        → suggest: demote ⏬
4. "https://thetopsecretcomedyclub.co.uk" (Jan 3)   → suggest: rewrite "Check Top Secret Comedy Club events @quick"
5. "when I pick up phone..." @deep (Jan 6)          → suggest: rewrite "Draft phone pickup rules @deep" or drop
6. "ready made proteïen @quick" (Jan 12)            → suggest: rewrite "Research ready-made protein options @quick"
7. "change alarm sounds" (Jan 21)                   → suggest: keep (scheduled Feb 7)
8. empty task "- [ ] " (Jan 24)                     → suggest: drop
9. "Check on side project repo" (Jan 25)              → suggest: keep (recurring)
10. "Pi Moon Mission Patch @quick" (Jan 25)         → suggest: keep or demote

Accept all suggestions? Or type numbers to override:

User: accept, but 1: keep (need them for allotment labels)

Agent: Got it. Executing...
✓ 3: Demoted to ⏬
✓ 4: Rewritten → "Check Top Secret Comedy Club events @quick"
✓ 6: Rewritten → "Research ready-made protein options @quick"  
✓ 8: Dropped (empty task)
(7 others kept as-is)

Next batch? (y/n/stop)