AI Assistant GTD Workflow: A 5-Stage Build

The best AI assistant GTD workflow is the one where the assistant never holds a single task. That is the honest conclusion, up front, and the rest of this piece is just the build: five stages, one handoff each, a prompt pattern where it helps, the failure mode that breaks it, and the one stage where the right move is to shut the window.

If you already run OmniFocus with custom perspectives, or Things, or Todoist, nothing here asks you to move. The assistant drafts and questions. Your task manager decides and holds. Get that boundary wrong and the system stops being trustworthy inside a fortnight.

Why most AI productivity advice breaks a GTD system

The common failure is quiet and quick. You start asking ChatGPT or Claude to “help me plan the week”, it produces a tidy list, you leave that list in the thread, and now you have two inboxes. One gets reviewed. The other sits in a conversation history nobody opens again.

GTD runs on completeness. The five stages David Allen sets out, capture, clarify, organise, reflect and engage, only work because every open loop lands somewhere you trust and every commitment appears on one set of lists. An assistant that keeps three of your commitments in a chat transcript has made your list incomplete. An incomplete list is one your brain stops believing, and once that happens you are back to holding things in your head while paying a subscription for the privilege.

So the working rule for this whole article: AI drafts and questions, your task manager decides and holds. Everything below is an application of that single line.

What has changed recently is worth naming, because it is what makes any of this viable. Persistent memory means the assistant can carry project background between sessions. Voice modes mean capture no longer requires typing. Shared, editable document surfaces such as ChatGPT Pages mean generated material can live somewhere with a URL instead of scrolling away up a thread. All three improvements help with reference material. None of them makes a chat window a list manager.

Stage 1: capture, where an AI assistant earns its place fastest

Job one: time to captured. Not how clever the tool is, not how good the transcription sounds in a demo video, just how many seconds pass between the thought arriving and the item sitting in the one inbox you actually process.

Under about five seconds, you will keep doing it under pressure. Over that, you will skip it while walking between meetings, which is precisely when the useful ideas turn up. Treat five seconds as a practical working benchmark rather than an official rule, but test yourself against it honestly. Phone locked, screen off, hands busy. That is the real test condition, not a calm desk.

Voice capture into an assistant works genuinely well, on one condition: the route ends somewhere else. A spoken note that triggers an email to your task manager’s inbox address, or fires a Shortcut on iOS, or drops through a share sheet into Things, has captured something. A spoken note that lives in the transcript has not. It has made a recording of an idea and then buried it.

Run the ten-item test. Over one working day, capture ten things by whatever mix of voice, phone and paper you normally use. At your next weekly review, count how many arrived in the single inbox. Fewer than ten and you have a broken route, not a discipline problem, and fixing the route takes twenty minutes once. The longer version of that build is in the piece on capturing with an AI assistant in under five seconds.

Stage 2: clarify vague open loops with the right prompt pattern

Clarify is where most inboxes rot. You have an item that says “website”. Six days later it still says “website”, because deciding what it means costs more thought than the two seconds it took to write.

This is the stage where an assistant does real work, provided you make it ask GTD’s questions rather than its own. The question set is fixed: is this actionable? If not, is it rubbish, reference, or someday/maybe? If it is, what is the successful outcome, and what is the very next physical, visible action?

A prompt pattern that holds up:

• “Here is a raw inbox item: [item]. Ask me whether it is actionable. If it is, ask me what done looks like in one sentence, then propose three possible next actions, each starting with a physical verb. Do not restate the item as the action. Flag anything you have assumed.”

That last clause matters more than the rest. Without it, the assistant will cheerfully invent plausible-sounding commitments you never agreed to, and they look exactly like the ones you did.

Worked through, “website”came out as “Draft the three questions for the designer about the pricing page”, under the project “New pricing page live”, tagged @Mac once it reached the task manager. Verb first, visible, doable in one sitting. The assistant proposed it. You approved it, line by line, and typed the context yourself.

For messier projects, the natural planning model makes a far better prompt skeleton than any “be my productivity coach”instruction: purpose, principles, outcome, brainstorm, organise, next action, in that order. Ask for the stages explicitly and the output stops being a generic project plan and starts being yours. Budget a few minutes per tangled project. Clarifying ten simple inbox items this way is slower than clarifying them yourself, so do not bother; save it for the three that have been staring at you for a fortnight.

Stage 3: organise without letting the assistant own your lists

Contexts, energy levels, waiting-for status, due dates: all of that stays in the task manager. AI output gets pasted in. It never syncs from.

Make the paste painless by asking for a paste-ready format. One action per line, verb first, project name in brackets, context tag at the end. Something like:

• Draft three questions for the designer about pricing [New pricing page live] @Mac
• Ring the accountant about the quarterly filing deadline [Q4 accounts filed] @Calls

A block in that shape drops into OmniFocus, Things or Todoist without re-typing, and more importantly it forces you to read each line as a commitment before it lands. Re-typing is not the enemy. Unexamined additions are.

The honest home for most AI output is project support material: briefs, research notes, meeting agendas, comparison tables, first-draft checklists. Reference you consult while working the project, stored next to the project, not scattered through the action list. Treat every generated bullet as a task and your lists inflate with work you never committed to, which is the fastest route to a list you avoid opening.

One rule on storage, and it matters more with AI than with ordinary notes, because generated reference piles up fast. If you cannot get it out as plain text or Markdown, it is not storage, it is rented memory. Evernote’s change of ownership and its tightened free tier were a reminder of how quickly a storage decision can be taken out of your hands, which is why the migration costs of the main alternatives are worth knowing before you accumulate three hundred generated documents, not after.

Stage 4: use AI to sharpen the weekly review, not to replace it

Reflect is where an assistant helps most and can quietly do the most damage.

What it can honestly do: read a week’s captured notes and meeting scribbles and pull out candidate open loops you wrote down but never turned into projects; scan a project list and flag the ones with no available next action; suggest the awkward question to ask about an item that has not moved in a month. All three are summarisation and pattern-spotting, which is what the technology is actually good at.

What it cannot do is feel the small flinch you get when you scroll past the project you have been avoiding. It will produce a neat, well-formatted summary and leave you feeling reviewed, which is worse than not reviewing at all, because you now have false confidence.

A pairing that works: open a custom perspective filtered to projects with no available next action, and run an AI pass over the week’s captured notes at the same time. The perspective catches structural stalls. The AI pass catches outcomes you mentioned in a note and never promoted to a project. Then you do the part neither of them can: decide what leaves the system. Dropping a commitment is a judgement call about what you are willing to disappoint people over. Keep it human, keep it in your own head, and do not outsource it to something optimising for a tidy answer.

Stage 5: engage, and the four moments to close the chat window

Doing the work is where AI costs you most. Every query is a context switch out of the action you chose, and the switch is cheap individually and expensive in aggregate.

Four jobs to keep off the assistant while you are engaged:

1. Choosing what to do next. Asking an assistant how to prioritise tasks produces a confident ordering based on none of the information that matters: your energy, the time you actually have, the person waiting on you, what you can live with not doing. Context, time, energy, priority. That is your call, made against your own lists.
2. Writing in your own voice to people who know you. They can tell. The relationship cost outweighs the eight minutes saved.
3. Anything with confidential or client data. Check your organisation’s policy and the tool’s data handling before pasting; the ICO’s guidance on AI and data protection is the sensible starting point for UK readers, and most enterprise agreements have their own line on this.
4. Anything whose accuracy you cannot check yourself. If you could not spot a wrong figure or an invented citation, you are not verifying, you are hoping.

The exception is low-energy list work. Twenty minutes left, nothing in the tank, a dull first draft due: that is the moment a generated starting point beats a blank page, because editing something mediocre takes less willpower than creating something from nothing. The longer breakdown of which jobs survive this test is in the piece on the six jobs ChatGPT does well in GTD and the four it ruins.

Then set a rule, the same way you set a context. The assistant is open during clarify, organise and review blocks. It is shut during execution blocks. Write it on the wall if you need to.

Putting the AI assistant GTD workflow together: a one-week setup plan

Build it in order, one piece a day. Skipping to the clever prompts before the capture route works is the standard mistake.

• Days 1 and 2: capture route only. Voice to assistant to email or Shortcut to your one inbox. Run the ten-item test on day two.
• Day 3: save your clarify prompt somewhere you can reach in two taps. Run it on the five oldest items in the inbox.
• Day 4: set the paste-ready output format and paste one block into your task manager. Check nothing arrived that you did not agree to.
• Day 5: move one project’s generated material into project support, wherever that lives for you, and confirm you can export it as plain text.
• Days 6 and 7: one weekly review with the stalled-projects perspective and an AI pass over the week’s notes. Do the clean-out yourself.

What good looks like at the end of the week: one inbox, one trusted list, and every piece of AI output sitting in reference rather than on an action list. If a generated bullet is on your next actions list and you cannot remember agreeing to it, the boundary has slipped.

Give it a fortnight. If a stage is not better than it was before, remove AI from that stage and keep the rest. That is a valid outcome, not a failure, and for a lot of people the honest result is that capture and clarify improve, organise stays manual and engage gets quieter. Fewer moving parts, wired together properly, beat a clever setup you stop maintaining by the second month.

If you want the next piece of the picture, read up on building a personal productivity system that lasts, because the boundary described here only holds inside a system you already trust.

Frequently Asked Questions

Can an AI assistant replace my task manager in a GTD workflow?

No. A chat history is not a complete, reviewable list, and GTD only works when every commitment lives in one place you check. The moment half your actions sit in a thread and half in an app, your brain stops trusting either. Keep the assistant as a drafting and questioning tool, and let OmniFocus, Things or Todoist hold the commitments.

Which GTD stages does an AI assistant actually help with?

Capture, clarify and reflect, with reflect the stage where it helps most, provided the decisions about what to drop stay yours. Capture gains from fast voice entry, as long as the route ends in your real inbox. Clarify gains because the assistant will patiently ask the outcome and next action questions on items you have been avoiding. Reflect gains from summarising a week of notes into candidate open loops, though the decisions about what to drop stay yours.

What prompt turns a vague open loop into a proper next action?

Give it the raw item and require three things: a yes or no on whether it is actionable, one sentence describing what done looks like, and three candidate next actions that each begin with a physical verb. Add “do not restate the item as the action”and “flag anything you have assumed”. That is how “website”becomes “Draft the three questions for the designer about the pricing page”.

Is it safe to put work projects into an AI assistant?

It depends entirely on the data and your organisation’s policy, so check both before pasting. Client information, personal data and anything under an NDA need a tool and a contract that permit it; the ICO publishes guidance on AI and data protection for UK organisations. For general planning and non-sensitive drafting, the risk is usually low, but verify rather than assume.

How do I stop the chat window becoming a second inbox?

Give every useful output a destination before you generate it. Actions get pasted into the task manager; reference gets saved into project support material; everything else gets closed. A short end-of-session habit works well: before you shut the window, ask yourself what in here is a commitment, move it, and then let the thread go.