Bokili
Tool Lessons4 min read

How to Build a Repeatable Workflow With ChatGPT Projects

Turn ChatGPT Projects into durable work systems by separating instructions, sources, history and review—not just collecting chats in a folder.

Bokili Editorial· Verified August 10, 2026
ShareX
Scattered chats and reference documents flowing into one organised project workspace and emerging as a consistent deliverable.

The expensive part of a recurring ChatGPT task is often not the answer. It is the restart: finding the right files, repeating the audience and tone, explaining what changed, and correcting assumptions that were settled last week. A ChatGPT Project can remove much of that setup—but only if you design it as a working system rather than a folder for miscellaneous chats.

The goal is not to make ChatGPT remember everything. It is to keep the right context together: stable instructions, a controlled source set, the history of the work and a clear output contract. That turns a useful conversation into a repeatable workflow.

A project should hold rules, evidence and work history

OpenAI’s current Projects documentation describes a project as a workspace that groups chats, reference files and project instructions. Project memory can use chats and files from the project, and moved chats inherit its instructions and file context. You can also save a useful response back into the project as a source. These features solve different problems: instructions state how to work; files supply evidence; chats preserve the path already taken.

Loose collection of chatsDesigned recurring workflow
Purpose“Everything about this topic”One recurring outcome for one audience
InstructionsScattered across promptsStable project-level operating rules
SourcesUploaded whenever someone remembersSmall, named set with dates and owners
HistoryOld threads are hard to interpretDecisions and changes are recorded
ReviewTrust the latest answerCheck claims against the current sources

Use the CONTEXT design test

CONTEXT: seven choices before the first recurring run

1

C — Concrete outcome

Name the deliverable: for example, a weekly competitor brief for a commercial leadership meeting.

2

O — Owner and audience

State who requests, reviews and uses the result. This controls depth, language and acceptable uncertainty.

3

N — Named sources

Keep the source boundary explicit. Label each file with a date and remove superseded versions.

4

T — Task instructions

Write stable rules for format, tone, evidence and what the assistant must not infer.

5

E — Evolving history

Use chats to retain previous runs, decisions and feedback. Do not hide key rules only inside an old thread.

6

X — eXamination step

Require source checks for figures, quotations and consequential claims before the output is used.

7

T — Tidy-up rhythm

Schedule a short source and instruction review so stale context does not become permanent context.

The distinction between project instructions and memory matters. OpenAI’s Memory FAQ says explicit guidance belongs in custom instructions, while memory draws relevant detail from conversations and sources. For a business workflow, do not rely on memory to infer a non-negotiable rule. Write the rule.

Worked example: a weekly competitor brief

Imagine a product marketing manager who prepares a Friday briefing on three competitors. The recurring sources are an approved positioning document, a feature matrix, the previous briefing and a list of official competitor pages. The output must fit one page, separate observed changes from interpretation and include direct links for every new claim.

Reading is a start. Practice makes it stick.

Start learning

Set up the workflow once

  1. 1

    1. Create a narrowly named project

    Use “Weekly competitor brief”, not “Market research”. A narrow name reinforces a narrow job.

  2. 2

    2. Add stable reference material

    Upload only approved internal context and record the date or version in each filename.

  3. 3

    3. Add project instructions

    Define the audience, structure, evidence standard, exclusions and review step.

  4. 4

    4. Run one baseline brief

    Correct the output and save the final structure or exemplar back into the project as a source.

  5. 5

    5. Start a fresh chat for each edition

    Keep each run inspectable while allowing the project to use relevant context from earlier work.

Project instructions for the weekly brief
You support a product marketing manager preparing a one-page Friday competitor brief for commercial leaders. Use only the sources available in this project and current official competitor pages when web search is enabled. Separate: (1) verified changes, (2) likely implications, and (3) open questions. Link every new external claim. Never treat an older feature matrix as current evidence. End with three items a human reviewer must check.
A consistent brief with evidence, interpretation and open questions kept visibly separate.

Adapt the source and privacy rules to your organisation. Do not upload material that your policy does not permit.

Three maintenance habits prevent context decay

  • Date sources and remove superseded files instead of leaving contradictory versions in place.
  • Put stable rules in project instructions; put run-specific requests in the current chat.
  • After each run, record decisions and unresolved questions rather than saving every intermediate answer as a source.

A project can make work more consistent, but it can also make an old assumption more persistent. OpenAI notes that project memory behaviour depends on the project’s memory setting and account or workspace configuration. Check those boundaries before using a project for sensitive or team work, and keep consequential decisions subject to human review.

The useful principle

Persistence is not the same as correctness. A good project preserves the workflow and makes its evidence easier to inspect.

Try it in ten minutes

Turn one repeated prompt into a project
  1. Choose a task you repeat at least monthly and write its exact deliverable in one sentence.
  2. Create a project with that deliverable as its name.
  3. Add no more than three current reference files.
  4. Write five project rules: audience, format, evidence, exclusions and review.
  5. Run the task once with a small, real input.
  6. Mark one answer that came from a source and one that was inferred.
  7. Revise the instructions so the distinction is clearer next time.

The best recurring AI workflow does not begin with a clever prompt. It begins with a clean boundary around the work. When instructions, sources and history each have a defined role, ChatGPT Projects can reduce repeated setup without hiding how the answer was produced.

Bokili helps employees practise the behaviours behind durable AI work: setting context, controlling sources, verifying important claims and improving a workflow through short, realistic missions.

Sources

  1. Projects in ChatGPT — OpenAI Help Center
  2. Memory FAQ — OpenAI Help Center
ShareX

Reading is a start. Practice makes it stick.

Bokili turns skills like this into ten-minute missions for your whole team, with instant feedback and progress you can see.

Start learning

Keep reading

All articles