I have been building a GTD system around a fairly simple division of responsibilities:
-
GTD is the ontology. It defines what things are: an Inbox item, Next Action, project, Waiting For, deferred action, Someday/Maybe possibility, calendar commitment, reference material, completed outcome or abandoned commitment.
-
OmniFocus is the system of record and logic. It records what I have committed to, the current state of that work, its context, genuine dates, review cycle and history.
-
ChatGPT is the coach and working partner. It helps me clarify, review and carry out the work. It is not allowed to invent my commitments, decide my priorities or silently reorganise my life.
ChatGPT can still do quite a lot. It can research, write emails, create documents, operate websites and work alongside me through a task. The distinction is that I retain the judgement. ChatGPT may recommend what to do and carry out an agreed action, but it does not get to decide that I have made a new commitment.
Capturing everything
I have two main automatic capture routes.
All of my email accounts are accessed through IMAP connections. In my system:
Unread means unprocessed.
If an email is unread, it is still something I need to deal with. Reading or discussing it with ChatGPT does not make it processed. It stays unread until the action needed to deal with it has actually been carried out and verified.
For voice capture, I speak to Siri. Siri creates an item in Apple Reminders, and that is transferred automatically into the OmniFocus Inbox. This gives me a very low-friction way to capture something before I forget it.
From there, everything is clarified and organised through a set of ChatGPT Skills. OpenAI describes Skills as reusable instructions and supporting resources for work you repeat, which is exactly how I am using them here: Save workflows as skills.
1. Inbox Triage to OmniFocus
This Skill processes unread email, starting with the oldest message and working through them one at a time.
For each message, ChatGPT shows me:
-
the subject;
-
a short plain-English summary;
-
its recommended GTD decision; and
-
a question asking whether I want to do that or something different.
Before making its recommendation, it considers whether the message relates to any of my Active or On Hold OmniFocus projects, existing actions or Waiting For items.
The decision process is conventional GTD:
-
Is it actionable?
-
If not, can it simply be marked read, or does it genuinely belong in Someday/Maybe?
-
If it is actionable, can the real action be completed in about two minutes?
-
If so, do it now.
-
If it will take longer, does it belong in an existing project, as a standalone action, or as a new multi-action project?
-
Has the email changed or completed something already recorded in OmniFocus?
A two-minute action is not treated as complete merely because ChatGPT tried it. The real outcome must be checked. If a reply creates a continuing open loop, for example, the existing project may need updating or a Waiting For action may still be required.
For work taking longer than two minutes, creating or updating the correct OmniFocus record is the processing action. Once that record has been approved, written and verified, the email can be marked read. I do not have to finish the future work before the email counts as processed.
Emails remain unread throughout any unfinished workflow. A saved draft, failed action, uncertain send or unverified OmniFocus write leaves the source email unread.
When an email needs a reply, ChatGPT writes the actual email. It shows me the complete final message, including recipients, subject, attachments and body, before I approve that exact version. If it is sent, ChatGPT verifies the matching message in Sent before recording it as sent or completing any dependent OmniFocus action.
2. OmniFocus Inbox Coach
This Skill processes the items that have reached my OmniFocus Inbox, including the things I captured through Siri.
It has two distinct phases: Clarify, then Organise.
For an unclear item, ChatGPT works out:
-
What is this?
-
Is it actionable?
-
What outcome do I want?
-
What is the next physical or visible action?
-
Am I actually committed to doing it?
Nothing is changed in OmniFocus while the meaning is still uncertain.
A Next Action must be something a person could observe me doing, specific enough to begin without another thinking session, available in its stated context and not really a project in disguise.
If completing the proposed action would not complete the intended outcome, it probably belongs in a project. Conversely, a single action is not inflated into a project simply because more planning could be added.
ChatGPT checks for existing projects and duplicate actions before proposing anything new. It then shows me the exact proposed destination, wording, tags, dates and flag state. I approve the proposal before it writes anything, and the result is read back and checked afterwards.
3. OmniFocus Project Review Coach
The review workflow begins with the project or Single Actions list with the oldest due review date, unless I name something specific.
It presents one project at a time, rather than giving me an enormous diagnostic report. The opening view includes the outcome, current Next Action, meaningful recent activity, deadlines or availability and one recommendation.
The first important question is usually whether I still want the outcome at all. The valid answer may be:
-
keep it Active;
-
move it On Hold;
-
reactivate it;
-
complete it because the outcome is already true; or
-
drop it because it is no longer a commitment.
For each project, the Skill checks that the title describes a short, recognisable achieved outcome. It also checks every current action for short, clear, scannable wording and a real observable finish point.
It looks for missing Next Actions, vague or overloaded actions, stale Waiting For items, false deadlines, inappropriate defer dates, stale flags, duplicates, blocked projects that still appear available and projects that may already be complete.
It also recommends an appropriate review interval. A fast-moving financial or health project may need weekly review, while a stable or deliberately paused project may only need reviewing monthly or quarterly.
Projects with credible financial, health, housing, legal or family consequences receive a stronger recommendation for Active status where action is genuinely possible. That does not allow ChatGPT to invent urgency, deadlines or flags. The final decision remains mine.
Once the agreed changes have been made and verified, the Skill can literally mark the project reviewed in OmniFocus and verify that its review date advanced.
4. Do My GTD Projects
This is the working mode of the Project Coach. Its purpose is to help me make real progress, not just improve the task list.
ChatGPT reads the live project, its outcome, actions, notes and previous continuity record. It then helps in one of four ways:
-
ChatGPT can do the work;
-
we can do it together;
-
I must perform a physical, private or identity-bound step; or
-
somebody else must act, making it a Waiting For situation.
Where tools allow, ChatGPT should actually perform the work: research the issue, write the document, prepare the email, open the website or carry out the agreed steps. For account, identity or judgement-dependent work, we co-drive it.
An OmniFocus action is not completed because an attempt ran. The actual result must be verified: the email is in Sent, the form was submitted, the file exists, the booking is confirmed or the requested output was produced.
OmniFocus is then updated to reflect what really happened. Any new Next Action must arise from the real outcome. If it would represent a new commitment or judgement call, ChatGPT recommends it and asks me rather than silently creating it.
Recording decisions and history
The conversation with ChatGPT is not the authoritative record. OmniFocus is.
Project and action titles stay short, clear and easy to scan. Details belong in notes rather than being repeated in long task names.
A project note has two broad parts:
-
Stable project information: the intended outcome, definition of done, background, constraints, important links and supporting information.
-
A chronological work record at the very end of the note.
The final section is called AI continuity. New dated entries are appended at the bottom. They briefly record:
-
what was decided;
-
what was actually done;
-
where it was done, when that matters;
-
the verified result;
-
important references or confirmation numbers;
-
whether communication was prepared, drafted, sent or answered;
-
unresolved dependencies; and
-
the exact point from which to resume.
This preserves the useful history without turning OmniFocus into a transcript or document repository. A prepared email is never recorded as sent, and an attempted action is never recorded as successful without evidence.
Because this information lives in OmniFocus, I can resume the project myself or with a different ChatGPT session without relying on the previous conversation being available.
5. GTD System Audit
The final Skill is deliberately separate from Inbox processing and the normal project review.
Its question is:
Can I still trust this system?
It is read-only by default. It checks the whole system for things such as:
-
Active projects without usable Next Actions;
-
vague or overlong actions;
-
projects disguised as actions;
-
duplicates;
-
missing or unexpected context tags;
-
stale Waiting For items;
-
inappropriate due or defer dates;
-
old flags whose urgency needs reconsidering;
-
false availability or false completion;
-
review-cycle problems;
-
recurring-task anomalies; and
-
broken or excessively verbose continuity records.
It reports a prioritised set of findings, beginning with something like:
I found 13 things worth reviewing. I changed nothing.
If I choose to fix them, we work through the findings one at a time. Each object is read again, the exact change is proposed, I approve it, the smallest change is made, and the result is verified.
Keeping the model simple
I have deliberately resisted adding AI confidence scores, energy ratings, elaborate priority matrices, excessive contexts or automatically generated project plans.
The underlying model remains:
-
projects and actions;
-
genuine GTD state;
-
a small, existing vocabulary of context tags;
-
real due and defer dates; and
-
a simple, user-owned flag for urgency.
That separation is what makes the system work for me. GTD provides the meaning, OmniFocus holds the truth, and ChatGPT helps me think and act without quietly taking ownership of my life.
I am exporting the Skills so anyone interested can inspect or adapt them here:
