Access granted · practical field guide

ChatGPT Work operating guide

ChatGPT Work is most useful when you give it a finished-work brief: the outcome, the source material, the boundaries and the review standard. This guide shows you how to turn that brief into a reliable workflow.

You will finish with

one reusable outcome prompt
one well-scoped Project
one source and permission map
one reviewed deliverable
one repeatable workflow or task
one handoff note for a teammate

Quick answer or brainstorm

Chat

A first draft

Multi-step research or business artifact

Work

A reviewed document, sheet, deck or report

Repo edits, shell commands or tests

Codex

A verified software change

0/12 complete

Step 01

Chat, Work or Codex?

Choose the surface by the shape of the work. The right choice makes the source boundary and review path visible from the first step.

Need	Use	Example
quick answer or brainstorm	Chat	Give me three angles for this memo.
multi-step research or finished business artifact	Work	Research the market and create a cited briefing.
repository edits, shell commands, tests or PR evidence	Codex	Update the resource renderer and run checks.

Do this now

1.Name the final artifact.
2.List the approved sources.
3.Choose the surface and review owner.

Realistic example

Resource-page QA

The source of truth is React/TypeScript and completion needs a diff and build verification, so the task belongs in Codex—not Work.

Expected output: A verified code change.

Copyable prompt

I need to complete: [task]. The final artifact should be: [document/spreadsheet/presentation/report/site]. The sources are: [files/websites/apps]. The work may or may not require: [browser/app actions/code/repository access]. Recommend Chat, Work or Codex. Explain the boundary, the first step and the review required.

Verification or handoff

Write down the condition that would require you to move this task to Codex.

ChatGPT Work and Codex Projects in ChatGPT Deep Research ChatGPT Apps Scheduled Tasks

Mark complete

Step 02

Define the deliverable before asking for work

Describe the thing you want to receive, not just the topic. Start with the messy brief in your own words, then turn it into a deliverable the agent and reviewer can judge.

Do this now

1.Capture the raw request, audience and decision before polishing it.
2.Name audience, decision, inputs and format.
3.Set length, tone, evidence standard, deadline and reviewer.

Realistic example

GTM briefing

Raw brief: we need to decide how to launch a new AI resource. The resulting contract asks for a two-page GTM briefing, current source links, risks and three launch recommendations for the founder and marketing lead.

Expected output: Every current product claim linked; assumptions marked.

Copyable prompt

Here is the messy version of the request: [notes, voice transcript or bullets]. Create a deliverable brief. Ask only questions that change the output. Separate facts, assumptions and recommendations. Return the brief first; do not start researching or drafting yet.

Verification or handoff

A teammate should be able to use the brief to assess the artifact without rereading the chat.

Mark complete

Step 03

Build a strong outcome prompt

A strong Work prompt has five parts: outcome, context, constraints, process and review. Match reasoning effort to the task: routine classification and formatting need less than ambiguous research or high-stakes synthesis.

Part	What it controls
Outcome	what must exist at the end
Context	approved sources and facts
Constraints	what must not happen
Process	plan, evidence and review pause
Review	how quality will be judged

Do this now

1.Replace the vague task with a specific artifact.
2.Name only the sources that matter.
3.Ask for assumptions, a final checklist and an appropriately scoped research approach.

Realistic example

Weak versus strong

Weak: ‘Research competitors and make a report.’ Strong: ‘Compare five competitors for a founder deciding whether to launch, cite every current claim and pause before finalizing.’

Expected output: A reviewable, audience-specific briefing.

Copyable prompt

Outcome: Create [artifact] for [audience] so they can [decision/action]. Context: Use only [files, Project instructions, connected apps, approved websites]. Constraints: Do not invent facts. Keep [tone/length/format]. Ask before taking external actions or using sources outside this boundary. Process: First propose a short plan and the level of research needed. Then gather and organize evidence. Draft a usable first version. Pause for review before finalizing. Review: Include a source list, open questions, assumptions and a final checklist.

Verification or handoff

Before drafting, confirm the requested artifact, source boundary, external-action rule and level of effort match the decision at stake.

Mark complete

Step 04

Use Projects as context

A Project is a container for an ongoing body of work—not a magic memory. Keep its instructions and files current enough to be trusted.

Do this now

1.Name the Project by outcome.
2.Add one authoritative source-of-truth document.
3.Create instructions, then start one chat per deliverable.

Realistic example

Keep product truth current

Use the repository’s PRODUCT-TRUTH and current product tracker as explicit sources; keep the resource manifest as the durable authority for outbound resource URLs.

Expected output: A Project another teammate can audit.

Copyable prompt

Project purpose: [purpose]
Primary audience: [audience]
Current source of truth: [source]
Preferred output style: [style]
Definitions and terminology: [terms]
Facts that must be cited: [facts]
Files that are authoritative: [files]
What to do when context conflicts: [rule]
What requires approval: [actions]

Verification or handoff

Archive or label stale material and record a Project owner and next review date.

ChatGPT Work and Codex Projects in ChatGPT Deep Research ChatGPT Apps Scheduled Tasks

Mark complete

Step 05

Connect sources deliberately

Every source is a permission decision and an evidence decision. Use the authoritative source for each fact and show conflicts rather than silently choosing.

Source	Best use	Check before relying on it
uploaded file	private reference	version and owner
Project instruction	durable rules	whether it is current
connected app	live business context	scope and freshness
public web	current facts	citation, date and quality
deep research	multi-source synthesis	plan and cited output

Do this now

1.Map each question to one source.
2.Check read/write permissions and freshness.
3.Define how conflicts will be shown.

Realistic example

Source map for a launch

Use the resource page for product copy, Search Console for current queries, competitor pages for market framing and customer comments as labeled qualitative evidence.

Expected output: A source list attached to the final brief.

Copyable prompt

Before using connected sources, list which source will answer which question. Prefer the authoritative source for each fact. If sources conflict, show the conflict instead of silently choosing. Cite current claims and label any inference.

Verification or handoff

Reject any result that makes a current claim without a source or a clearly labeled inference.

ChatGPT Work and Codex Projects in ChatGPT Deep Research ChatGPT Apps Scheduled Tasks

Mark complete

Step 06

Create documents, sheets and decks

Ask for an artifact with visible structure and a reviewable first version. A downloadable file is not proof of correctness.

Do this now

1.Choose document, spreadsheet or presentation.
2.Provide the expected sections and evidence standard.
3.Approve the outline before final formatting.

Realistic example

Three artifacts

Document: two-page decision memo. Spreadsheet: raw data, assumptions, calculations, summary and formula notes. Presentation: ten slides, one message per slide, speaker notes and source footer.

Expected output: A structured first version plus review checklist.

Copyable prompt

Create a [document/spreadsheet/presentation] with [sections]. Use headings, tables or visuals where they improve scanning. Flag missing or inconsistent data instead of filling it silently. Return an outline and source mapping for approval before finalizing.

Verification or handoff

Review source mapping, calculations, citations and audience fit before sharing the artifact.

Mark complete

Step 07

Set browser and cloud boundaries

Let Work browse or act only inside a boundary you can explain. Drafting an external action is different from executing it, and a denied action must not be worked around through another app or browser surface.

Do this now

1.List allowed websites and connected sources.
2.Mark reading, drafting and execution as separate steps.
3.Name the exact confirmation point for every external side effect.

Realistic example

A safe research run

Allow reading approved sites and Drive files, but require the final email, form submission or shared-file edit to be shown for approval first. If a connector cannot perform the action, stop rather than switching to browser control to bypass the boundary.

Expected output: A traceable approval boundary.

Copyable prompt

Use only these websites and connected sources: [list]. Do not log in to new services, send messages, publish content, edit shared files or submit forms without showing me the final action and asking for approval. Do not bypass a missing permission or connector restriction through another surface.

Verification or handoff

Check availability and permissions at run time; desktop files, cloud browser, apps and workspace settings can vary.

Mark complete

Step 08

Use long-running work safely

Long-running work needs checkpoints, not one giant prompt. Make progress, sources, risks, material updates and pause points visible.

plan → source inventory → first findings → draft → review pause → revision → final handoff

Do this now

1.Split work into checkpoints.
2.Request a source inventory before synthesis.
3.Define a material-update rule and pause before external actions and finalization.

Realistic example

Deep research briefing

First approve the research plan, then inspect the cited findings and open questions before asking for the final recommendation. Report only a decision, source or risk change—not a stream of routine activity.

Expected output: A cited report with a visible review pause.

Copyable prompt

Break this work into checkpoints. Before each major phase, summarize progress, sources and open questions. State what counts as a material update, when to remain quiet and the stop condition. Pause before external actions and before finalizing the deliverable. If the task cannot continue because a source or permission is missing, explain the blocker instead of improvising.

Verification or handoff

Each checkpoint should state completed work, sources used, decisions needed, risks, next checkpoint and the condition for stopping.

ChatGPT Work and Codex Projects in ChatGPT Deep Research ChatGPT Apps Scheduled Tasks

Mark complete

Step 09

Automate only after review

First prove the workflow manually. Then make the repeatable parts automatic with a stated runtime source boundary, owner, cadence and stop condition.

Do this now

1.Run one successful manual version.
2.Extract a reusable prompt or template.
3.Verify scheduled runtime access, quiet reporting and the stop condition before relying on it.

Realistic example

Weekly leadership update

After a reviewed manual update works, schedule a draft using approved sources that flags assumptions, reports only material changes and never sends anything externally.

Expected output: A draft ready for human review.

Copyable prompt

At [time/frequency], prepare [artifact] using [approved sources]. Include [required sections]. Report only these material changes: [criteria]. If there are no material changes, return [brief status or no update]. Stop when [condition] is met. Do not take external actions. Show sources, assumptions and items requiring approval.

Verification or handoff

Do not assume a scheduled workflow can access every Project file or connected source available in an interactive chat; verify current behavior, runtime inputs, owner and stop condition.

Mark complete

Step 10

Teams and enterprise governance

Shared workflows need owners, permissions and review standards. Enterprise adoption is not everyone receiving the same prompt.

Role	Responsibility
workflow owner	keeps instructions and sources current
contributor	works within approved boundaries
reviewer	checks evidence and output quality
workspace admin	controls apps, permissions and availability
data/security owner	approves sensitive-data handling

Do this now

1.Name workflow owner and intended users.
2.Allowlist sources and external actions.
3.Set review cadence and a retirement plan.

Realistic example

Rollout worksheet

Record workspace, Project/template, approved apps, sensitive-data rules, source standard, audit evidence, approval points, owner and rollback plan.

Expected output: A governed pilot that can be expanded deliberately.

Copyable prompt

Workflow owner: [owner]
Intended users: [users]
Workspace: [workspace]
Project or template: [asset]
Approved apps: [apps]
Approved sources: [sources]
Allowed write actions: [actions]
Human approval points: [points]
Sensitive-data rules: [rules]
Source/citation standard: [standard]
Audit evidence: [evidence]
Review cadence: [cadence]
Rollback or retirement plan: [plan]

Verification or handoff

A workflow is ready only when it has a named owner, approved sources, explicit permissions and a reviewable output.

ChatGPT Work and Codex Projects in ChatGPT Deep Research ChatGPT Apps Scheduled Tasks

Mark complete

Step 11

Troubleshoot rollout friction

When Work disappoints you, diagnose the boundary before rewriting the prompt. Most problems are vague outcomes, weak sources, missing access or a task that is too large.

Symptom	Likely cause	First fix
output is generic	outcome or source boundary is vague	add audience, decision and authoritative inputs
citations are weak	source policy is unspecified	require source mapping and current links
file is unavailable	surface, permission or file-limit constraint	verify access; process smaller batches
task stops halfway	scope too large or no checkpoints	split phases and require updates
scheduled result lacks context	runtime access differs	verify runtime inputs

Do this now

1.Identify the exact failure symptom.
2.Check outcome, sources, permissions and scope in that order.
3.Move orchestration outside chat when the platform boundary is the blocker.

Realistic example

Community reports

Users report friction with large sequential file or image batches. Treat these as symptoms to investigate, not product guarantees.

Expected output: A smaller, evidenced rerun or a clear escalation.

Copyable prompt

Diagnose this failed workflow. Separate outcome ambiguity, source quality, permission limits, platform availability and task scope. Propose the smallest test that distinguishes the likely causes. Do not claim a product limitation without evidence.

Verification or handoff

Label community reports as user-reported experience, not authoritative product behavior.

Community report: practical Project structures Community report: review before automating recurring work

Mark complete

Step 12

Handoff and maintenance

A finished artifact is useful only when someone can trust and reuse it. Preserve the source list, assumptions, decisions and the prompt or template that made it possible.

Do this now

1.Link the artifact and name its audience.
2.List sources, assumptions and unresolved questions.
3.Assign next owner, review date and template maintenance.

Realistic example

Handoff to Codex

When a recommended next step requires a repository edit, shell command, test or release evidence, hand off the task with its outcome contract and source map.

Expected output: A clear boundary between knowledge work and software work.

Copyable prompt

Prepare a handoff for a teammate who did not see this conversation. Include the final artifact, purpose, sources, assumptions, unresolved questions, decisions, next owner and review date. Identify which Project instructions or templates should be updated.

Verification or handoff

Confirm the next person can locate the artifact, understand its provenance and know whether the next step belongs in Work or Codex.

Mark complete

Finish with a handoff

Make the next run easier.

Record what changed, why it changed, the checks or sources used, remaining limits, the next owner and the next review date. A useful workflow is one another person can continue without redoing the discovery.

Back to the tutorial overview Codex tutorial Claude Code tutorial AI Work Cycle Context Engineering Daily Overview Second Brain