A to-do list fails in a specific way: you read an item, feel a small wave of resistance, and move on to an easier one. The item wasn't a task — it was a decision you hadn't made yet. “Fix the website” isn't something a human being can do; it's a stack of unmade choices wearing a checkbox.

The fix, from David Allen's Getting Things Done, is to never store decisions as tasks. Ask of every item: what is the very next physical action? Not the goal, not the project — the motion. If the honest answer is “call the vendor,” write that. Deciding happens once, at capture. After that, the list is pure execution.

The anatomy of a next action

A well-formed action has a grammar. Each part answers a question your future self would otherwise have to stop and answer.

Fig. 1Every part closes a question. A task with no open questions has no friction left.

Five tests

Run any list item against these. It only counts as a next action if it passes all five.

  1. 01
    It starts with a verb. Call, email, draft, profile, book. If the first word isn't a motion, it's a project in disguise.
  2. 02
    You can picture yourself doing it. One sitting, one place. If it needs several sessions, break it down until one doesn't.
  3. 03
    The outcome is specific. Not “work on the deck” — “draft the three intro slides.” Specific enough that someone else could do it.
  4. 04
    Nothing is blocking it. You can start right now. If it's waiting on something, the real next action is chasing the dependency.
  5. 05
    Done is obvious. You'll know the moment it's finished, without having to decide whether it counts.

Worked examples

The grammar doesn't change with the job — only the verbs and the objects do. Pick yours.

Profile model loading latency for Llama 70B on A100
Review 3 benchmarking PRs from morning standup
Identify root cause of CUDA OOM error in batch size 32 requests
Update inference API docs with new streaming endpoint parameters
Benchmark throughput on H100 vs A100 for Mistral 7B
Email DevOps team about GPU quota increase for staging
Test model cache eviction with concurrent 50 user load
Profile KV cache memory usage for requests over 8K tokens
Stalls on the listGets done
Improve inference performanceProfile GPU memory usage for batch size 16 on Mistral 7B
Fix the bugFix CUDA timeout error in multi-GPU load balancing (issue #142)
Work on cachingImplement LRU eviction policy for model weight cache
Research deployment optionsCompare H100 vs A100 cost per token for production workload
Update documentationAdd rate limiting examples to API docs streaming section
Launch A/B test for pricing page CTA button color
Document top 3 reasons for signup funnel drop-off at step 2
Draft blog post outline for reducing inference costs with batching
Schedule 5 Twitter posts for new model launch campaign
Set up conversion tracking for free trial to paid transitions
Email 10 AI newsletter editors with partnership pitch template
Add "GPU inference API" keyword to landing page title and H1 tags
Draft referral program mechanics with $50 credit structure for both parties
Stalls on the listGets done
Improve conversionA/B test signup form with 3 vs 5 input fields
Work on SEOAdd meta descriptions to 10 highest traffic blog posts
Grow social mediaPost 3 technical demos on Twitter with benchmark screenshots
Reach out to partnersEmail 5 ML frameworks about integration partnership
Analyze metricsCreate dashboard showing daily active users by plan tier
Draft user stories section for streaming response PRD
Conduct 3 customer interviews and summarize deployment needs in doc
Calculate RICE scores for 12 Q1 feature requests in spreadsheet
Create pricing comparison table for Replicate and Together AI
Define success metrics for new API rate limiting feature
Schedule kickoff meeting with eng and design for model quantization project
Update product roadmap deck with Q4 delivery dates and dependencies
Categorize billing issues from last 50 support tickets into spreadsheet
Stalls on the listGets done
Work on roadmapAdd estimated effort and impact scores to Q2 feature list
Talk to usersSchedule 30-min interviews with 5 users churned last month
Write specsDraft user stories for batch inference API endpoints
Check metricsCompare week-over-week API error rates by endpoint
Research competitorsDocument OpenAI and Anthropic pricing tiers in spreadsheet

The test that catches everything: if you can't picture yourself physically doing it, it isn't a next action — it's a project wearing a task's clothes. Projects are fine. They just don't belong on the list you execute from.

Source The next-action discipline comes from David Allen's Getting Things Done (2001). The book is bigger than this page — contexts, weekly reviews, the two-minute rule — but the single question “what's the next physical action?” carries most of the value.