How to Make ChatGPT Work With GTD: 6 Jobs It Does Well (And 4 It Ruins)

ChatGPT belongs next to your GTD system, not inside it. That is the honest conclusion, up front: it is genuinely useful for clarifying, planning and interrogating a weekly review, and it is a poor place to keep your actual commitments. If you want ChatGPT work and GTD to sit together without quietly rotting your trust in your lists, the split matters more than the prompts.

Getting Things Done runs on five steps: capture, clarify, organise, reflect, engage. Two of those are thinking problems. Three of them are storage and retrieval problems. A language model is good at thinking problems and structurally bad at storage, because a chat thread has no reliable state, no offline copy and no guarantee that what you told it on Monday is still in the window on Friday.

Where ChatGPT actually fits in the five steps

Capture is a speed problem. Organise is a storage problem. Engage is a retrieval problem under time pressure, usually with a phone in one hand. Clarify and reflect are judgement problems: deciding what a thing means, what the outcome is, and what the very next physical action looks like.

Split it that way and the map draws itself. ChatGPT is strong where the work is wording, judgement and thinking out loud. It is weak everywhere the work is holding state you can bet on.

The trust boundary

Your list container has to be identical at 9am Monday and 4pm Friday. Same items, same order, same wording, no surprises, available on a train with no signal. A chat context window does not meet that bar and was never designed to. Context gets truncated, threads get lost in a sidebar, a regenerated answer quietly drops three items, and you find out during the review that is meant to be the thing you can rely on.

A review you cannot trust is worse than no review, because you walk away from it believing you are current when you are not.

What changes with a connected setup

A custom GPT or agent that writes directly into OmniFocus or Todoist via an API is a different animal. Now the state lives in your task manager, which is correct, and the model is only a keyboard. Better architecture, higher cost: you own an integration, an auth token, and a review burden every time something upstream changes. More on that below, because the exit cost is the part people underestimate.

6 GTD jobs ChatGPT does genuinely well

1. Turning a vague open loop into a physical next action

“Sort out the garage”is not a next action. It is a project pretending to be one, which is why it sits on your list for four months looking reasonable.

Ask the model one question, hard: what does done look like? Then force one physical action with a verb, a place and a tool. “Measure the shelf gap in the garage with the tape from the kitchen drawer”is startable. The classic failure is the model handing back another mini project, so build the rejection rule into the prompt: refuse anything that needs a second decision before you can start.

2. Running the natural planning model on a stuck project

Most stuck projects are stuck at purpose and successful outcome, not at task generation. Nobody is short of tasks.

Run the five phases from David Allen’s natural planning model in one pass: purpose, successful outcome, brainstorm, organise, next actions. Take “plan the team offsite”. The task list you get back will be forgettable. The outcome statement, the one that says what the team is able to do on the Monday afterwards, is the bit worth keeping, and it is usually the bit you skipped. If the model jumps straight to a numbered task list, it is reproducing exactly the reactive planning GTD warns about. Make it answer purpose first or start again.

3. Interrogating a weekly review

This is the highest-value use, and it works because you supply the data. Paste in your project names, nothing else, and ask three questions: which of these have no next action, which have not changed since the last review, and which are actually waiting-for items in disguise.

On a list of around 35 projects, expect a handful of genuine flags, roughly five to eight, and expect to throw some out. That is fine. The model is not deciding anything. It is reading your list back to you slowly, which is precisely what you stop doing yourself once the list gets long.

4. Email: short replies and backlog triage

Email is the one place where an AI layer changes throughput in a way you can feel, because processing an inbox is mostly repeated micro-decisions plus short, low-stakes replies. Two-minute-rule replies drafted for you, in your wording, cleared in a pass.

Keep the decision yours. Action, reference, bin, delegate: you make that call every time. Let the model file things and you no longer know what your inbox contains, and an inbox you do not recognise is an archive with anxiety attached. Same logic that drives the picks in this comparison of Mac email apps for Getting Things Done: throughput is about decisions per minute, not about clever automation.

5. Project start-up checklists for repeatable work

New client onboarding. House move. Conference talk. If you run the same shape of project more than twice a year, get the model to draft the checklist once, edit it hard, then store it as a template in your task manager where it will actually be seen. The model is doing recall work, not judgement work, and it is quick at it.

6. A somebody-else brain for the mind sweep

Staring at a blank page is the worst way to run a mind sweep. Ask the model to walk you through trigger categories one at a time: professional commitments, people you owe replies to, health, home maintenance, finances, things you have been avoiding. Answer out loud or in short lines. Capture the output somewhere permanent, not in the thread. The prompting is the value; the storage is not.

4 GTD jobs ChatGPT quietly ruins

1. Being your list of record

No reliable state. No offline access. Silent context loss. You cannot glance at it, you cannot sort it by context, and you cannot prove to yourself that nothing fell out. Everything else on this list is a nuisance. This one breaks the system at the foundation.

2. Capture

Job one: time to captured. A phone quick-add or a voice memo lands a thought in under five seconds, screen barely on. Opening a chat, typing, waiting on a model and then reading a reply takes tens of seconds and hands you something to process. Capture has to be faster than the thought or you stop doing it, and the ideas you lose are the ones you had crossing a car park.

3. Retrieval from half a memory

Semantic search across old threads is decent. Genuinely. “That thing about the supplier contract in the spring”will often surface. But finding every project tagged @waiting-for that involves Sarah, sorted by age, is not a search problem, it is a structured data problem, and chat history has no structure. Mixed verdict, and the half it fails is the half you need on a Friday afternoon.

4. Exit cost

A bespoke custom GPT plus a no-code backend is the 2020s version of the elaborate setups people once built on Tracks, or hacked together across Palm and Pocket PC lists. Brilliant while they run. Completely dependent on one person maintaining them, and that person is you.

The question is not hypothetical. When an API version changes, or a plan tier moves, or an action stops authenticating, where do your next actions live? If the answer is “behind the integration”, you have built a beautiful shelf of software you never open, and a fortnight of rebuild work you did not budget for.

A working setup alongside your existing system

Three containers, and only three.

• A fast capture tool. Quick-add widget, voice memo, an index card. Under five seconds, always.
• One trusted list container. OmniFocus, Todoist, Things, a notebook. This is the system of record and nothing else gets that job.
• ChatGPT as a thinking layer that writes nothing anywhere without your review.

The workflow is deliberately unglamorous: export or paste your project list, run the review prompt, read the flags, then type the agreed changes back into your task manager yourself. Copy out, copy in. It takes a few minutes and it keeps the model on the correct side of the trust boundary.

Voice capture with an AI assistant is where a fourth silo appears without anyone noticing. Dictating to an assistant on a walk works well, and the transcription quality on current phone assistants is good enough for messy sentences. Route the output into your one inbox. If your voice capture tasks land in an AI assistant’s own list, you now have two inboxes and a review you will skip. The same test applies to any note taking app for GTD: capture speed counts for nothing if the note lands somewhere you never process.

Keep the admin tax at zero. If the AI layer needs weekly fixing, re-authing or tidying, it has failed, and you should drop it without ceremony.

On confidential content: strip client names, salary figures, unreleased plans and anything personal about identifiable colleagues before you paste. Use initials or role labels. Check your employer’s policy, and check the current data controls on your account rather than assuming, since those settings change. The ICO’s guidance on AI and data protection is the sensible UK reference point if you are the person who has to answer for it.

Prompts that hold up over a real week

Clarify. “Here is one item from my inbox. Ask me what done looks like. Then give me exactly one physical next action with a verb, a place and a tool. One line. If it needs a second decision before I can start, say so and ask me the question instead.”

Weekly review. “Here is my projects list. For each, tell me: does it have a clear next action, is it actually a waiting-for, and has the wording changed since last week? Flag stalled and orphaned items only. Do not invent projects. Do not suggest new ones.”

Natural planning. “Project: plan the team offsite. Before any tasks, give me the purpose in one sentence and the successful outcome in one sentence. Wait for my edits. Only then brainstorm.”

Feed it your contexts (@calls, @errands, @laptop), your realistic energy levels and your horizons of focus, and suggestions get specific instead of generic. Then watch for drift, which sets in faster than people expect: invented projects, motivational padding, tasks you never mentioned, a sudden fondness for the phrase “let’s break this down”. When that starts, kill the thread and reopen with the prompt. Long threads decay.

Chat coach, custom GPT, or connected agent?

Plain chat with copy and paste has the lowest exit cost and no setup. It needs discipline, and it is what most people should use.

A custom GPT with fixed instructions buys consistency of tone and rules, so you stop re-teaching it your contexts every Sunday. Still no reliable state, still a fourth place to look.

An agent or API setup writing into OmniFocus or Todoist is the powerful option and the fragile one. Your state stays in a proper app, which is right, but you now own an integration. Decide by rebuild cost: how much of your working week would you have to reconstruct if the tooling changed next quarter?

Analogue remains a fair benchmark, and the old hipster PDA card stack is not a joke. If a stack of index cards and an honest weekly review could carry your commitments, and the AI layer could not, the AI layer is decoration. Could you drop back to paper for a fortnight tomorrow without losing anything? That is the honest test, and the same principle that runs through the earliest chapters of the GTD book: the system exists so your head does not have to hold it.

Frequently Asked Questions

Can ChatGPT replace my GTD app?

No. It has no reliable state, no offline access and no structured retrieval by context or tag, so it cannot be the list you check at 4pm on a Friday. Use it as a thinking layer over a task manager or notebook that holds the actual commitments.

What is the best prompt for a GTD weekly review with ChatGPT?

Paste your project names and ask it to flag three things: projects with no clear next action, projects whose wording has not changed since the last review, and items that are really waiting-fors. Tell it not to invent projects or suggest new ones. On a list of around 35 projects, a handful of useful flags is a good result, and rejecting some is normal.

Is it safe to paste work projects and client names into ChatGPT?

Strip identifying detail first: client names, salaries, unreleased plans, anything personal about named colleagues. Replace them with initials or role labels, which costs you nothing in usefulness. Check your employer’s policy and the current data settings on your account, since those change.

Should I build a custom GPT for GTD or just use plain chat?

Start with plain chat and copy-paste. A custom GPT is worth it only if you are re-typing the same instructions every week and you accept that you now maintain something. The moment it needs fixing more than it saves, drop it.

Does ChatGPT help with capture, or is a notes app still faster?

A notes app is faster, and the gap is not close. Quick-add or a voice memo lands a thought in under five seconds with the phone barely awake, while a chat round trip takes tens of seconds and gives you a reply to read. Compare the options on capture speed rather than feature lists and keep ChatGPT for the clarifying step afterwards.