Staff guide • ChatGPT Business + Codex

Get more done without burning through AI credits.

Use the right tool, the right model, the right thread, and the right amount of context. The aim is not to stop AI use — it is to stop paying premium credits for routine work.

Start with the tool picker Copy prompt templates Estimate credit impact

Built from public OpenAI and Microsoft guidance checked on 30 June 2026. Your workspace policy always wins where it is stricter.

The seven credit-smart rules

These are designed to be simple enough for every department, but still useful for developers doing real work.

Recommended default

🧭

1. ChatGPT plans. Codex changes code.

Use ChatGPT Instant for requirements, options, user stories, acceptance criteria, UI text, and test ideas. Use Codex when it needs the repository or must edit/run code.

🪜

2. Use the cheapest capable model.

Company default: use GPT-5.4 or GPT-5.4 mini for routine work. Escalate to GPT-5.5 only for complex debugging, architecture, cross-cutting refactors, or genuinely blocked tasks.

⚡

3. Fast mode is not the default.

Fast mode is useful when speed matters more than credits. Otherwise, leave it off. It gives a speed boost but consumes credits at a higher rate.

🧵

4. One task per thread.

Long chats carry old context. Keep one thread for one active task, then ask for a handoff summary and start a new thread for the next task.

🎯

5. Feed only useful context.

Point Codex to files, folders, selected code, errors, commands, and acceptance criteria. Avoid “look through the whole repo” unless that is truly necessary.

🧪

6. Ask for a plan before edits.

For complex or unclear work, use Plan/Chat mode first. Approve the plan, then allow implementation. This prevents long, expensive wrong turns.

✂️

7. Control output length.

Ask for diffs, summaries, file paths, and test results — not full file dumps. Output tokens are usually much more expensive than input tokens.

🧹

Bonus: keep agent instructions lean.

Shorten AGENTS.md, nest instructions near the relevant code, and disable unused MCP servers. Every automatic instruction and tool can add context.

📊

Bonus: review usage weekly.

Admins should check credits by user, model, surface, and use case. Coach outliers before buying more credits.

The default workflow

This is the recommended flow for both developers and non-developers building scripts, dashboards, or simple internal apps.

1. DefineUse ChatGPT Instant to clarify the outcome, users, constraints, and acceptance criteria.

2. PlanAsk for an implementation plan and risks. Do not write code yet.

3. HandoffPaste a short plan into Codex. Include files, constraints, tests, and model instruction.

4. BuildUse Codex GPT-5.4 / mini first. Keep Fast off. Approve the plan before edits.

5. ResetAsk for a handoff summary. Start a new thread for the next task.

Tool picker

Answer four questions to decide whether to use ChatGPT, Codex chat/plan, or Codex agent mode.

Does the task need repository files?

Will AI edit or run code?

How ambiguous is it?

Business urgency

Use ChatGPT Instant

Start outside Codex. Get the requirements, plan, acceptance criteria, and tests right before using coding credits.

Model: InstantThread: new

Credit impact calculator

This uses public Codex token-based rates for GPT-5.5, GPT-5.4, and GPT-5.4 mini. It is an estimate, not a billing statement.

Estimator

Task size

Codex model

Input tokens

Cached input tokens

Output tokens

Runs/messages per week

Fast mode

Fast mode applies to GPT-5.5 and GPT-5.4 only.

Output discipline

estimated credits per run

estimated credits per week

estimated credits per month

saving vs GPT-5.5 standard

Risk: normal.

Guidance by role

The same rules apply to everyone, but the default workflow changes slightly by job type.

Default

GPT-5.4, medium reasoning, Fast off. Use selected files and clear acceptance criteria.

Escalate

Use GPT-5.5 for complex debugging, architecture, multi-file refactors, performance problems, or when GPT-5.4 is stuck.

Avoid

Parallel cloud runs without a plan, whole-repo prompts, verbose file dumps, unused MCP servers, and “keep going until done”.

Default

Use ChatGPT first to define the app, spreadsheet workflow, automation, or report. Produce a simple build brief.

Then Codex

Ask Codex to build one small part at a time: one page, one script, one function, one bug fix.

Avoid

Asking Codex to “make an app” from a vague idea. Vague prompts create more attempts, more rework, and more credits.

Track

Review weekly usage by model, user, product surface, and task type. Look for repeated high-credit patterns.

Coach

Coach people on prompt hygiene before restricting useful work. Most savings come from better defaults and handoffs.

Govern

Set soft limits, alerts, and an escalation route for GPT-5.5, Pro, Extra High, Fast mode, and long-running automations.

Copy/paste prompt templates

These templates reduce rambling, keep context focused, and make the ChatGPT-to-Codex handoff cleaner.

ChatGPT planning prompt

I need a concise implementation plan for: [outcome].

Do not write code yet. First clarify the goal, users, constraints, data inputs, and acceptance criteria. Ask up to 5 questions only if they materially change the answer.

Output format:
1. Recommended approach
2. Step-by-step plan
3. Risks / assumptions
4. Acceptance criteria
5. Test cases
Keep it under 700 words.

ChatGPT → Codex handoff

Implement the plan below in this repository.

Model policy: use GPT-5.4 unless the task is genuinely blocked or complex enough to justify GPT-5.5. Keep Fast mode off.

Start in Plan/Chat mode. Do not edit files until you have identified the likely files, risks, test command, and minimal diff plan.

Goal: [goal]
Files/areas to inspect first: [paths]
Acceptance criteria: [criteria]
Constraints: [do not rules]
Plan from ChatGPT: [paste concise plan]

Codex targeted bug fix

Bug: [what is broken]
Expected: [expected behaviour]
Actual: [actual behaviour]
Evidence: [error, stack trace, failing test, screenshot]
Relevant files: [paths]

Please diagnose first. Return:
1. Likely cause
2. Minimal fix plan
3. Files you need to edit
4. Test command

Do not rewrite unrelated code. Keep output to the plan and diff summary unless I ask for full files.

Fresh thread handoff

Summarise this thread into a fresh-start handoff.

Include only:
- Goal
- Decisions made
- Important constraints
- Files changed or inspected
- Commands/tests run
- Current status
- Known issues
- Next recommended step

Maximum 15 bullets. Do not include the full conversation.

Small app brief for non-developers

I want to build a simple internal tool for: [business process].

Before writing code, help me make this buildable:
1. Who will use it?
2. What data comes in?
3. What should the user see or do?
4. What output should it produce?
5. What security or approval rules matter?
6. What is the smallest useful first version?

Then produce a one-page Codex build brief with acceptance criteria.

Output control instruction

Be concise. Do not paste entire files unless necessary. Prefer:
- changed file paths
- diff summary
- commands run
- test result
- remaining risks

If you need to show code, show only the changed block.

When to start a new thread

Starting new threads can be efficient, but only when the old context is no longer useful. Continuing can be efficient when Codex genuinely needs the same context.

Start a new thread when…

The topic, feature, bug, or repository area has changed. The thread contains lots of abandoned attempts or irrelevant discussion. You are moving from planning to implementation and only need the final brief. You need a clean second opinion without inherited assumptions.

Keep the same thread when…

Codex still needs the same files, branch, commands, and decisions. The next prompt is a direct follow-up on the same change. You want Codex to apply feedback to the patch it just produced. Restarting would require pasting a large amount of context again.

Model and mode policy

Use this as a simple internal standard. Replace names if your workspace model list changes.

Use case	Default	Escalate only when…	Avoid
Planning, requirements, acceptance criteria	ChatGPT Instant	Use Thinking once for complex trade-offs or ambiguous architecture.	Do not use Codex just to brainstorm.
Small code tweak, test, script, CSS, copy changes	Codex GPT-5.4 mini or GPT-5.4	Use GPT-5.5 only if the smaller model repeatedly misses the issue.	Fast mode and full repo scans.
Normal feature or bug fix	Codex GPT-5.4	GPT-5.5 if multi-file reasoning, hidden dependencies, or hard debugging are involved.	“Keep going” loops without acceptance criteria.
Complex refactor, architecture, security-sensitive work	Plan first, then Codex GPT-5.5 if justified	Fast mode only for urgent blockers where time matters more than credits.	Starting with Fast mode before scope is clear.
Code review	Target specific PR risks and changed files	Ask for deeper review only on high-risk areas.	Repeated broad reviews with no changes.

Credit-drain warning signs

When you spot one of these, pause and reset the prompt before sending another high-cost request.

“Look through the whole repo.”

Point it to likely files first. Use whole-repo exploration only when discovery is the task.

“Make me an app.”

Turn the idea into a one-page brief before Codex writes code.

Long thread, new topic.

Ask for a 15-bullet summary and start fresh.

Full file dumps.

Ask for changed blocks, diffs, and test results instead.

Fast mode for routine tasks.

Fast mode is an exception, not a default.

Multiple parallel agents.

Parallel work is powerful, but expensive when the plan is vague.

Unused MCP servers.

Disable tools you are not using for the task.

GPT-5.5 as first resort.

Use GPT-5.4 or mini first unless the task meets the escalation criteria.

Admin playbook

The goal is to coach better use before buying more credits. Use controls to prevent surprises, not to discourage good AI work.

Weekly usage review

Review top users by credits, not just message count.
Break down usage by model, surface, and Fast mode where available.
Look for repeated GPT-5.5 use on routine tasks.
Identify long-running or parallel Codex jobs that did not ship useful output.
Share one practical coaching tip each week.

Suggested controls

Default guidance: GPT-5.4 / mini, Fast off.
Soft approval route for GPT-5.5 Fast, Pro, Extra High, and large automations.
Credit alerts at 50%, 75%, and 90% of the monthly budget.
Per-user overrides for heavy legitimate users, not blanket restrictions.
Monthly refresh of model names and policy wording.

SharePoint hosting notes

This file is a standalone HTML guide. Some SharePoint tenants restrict custom script or embedded content, so use the route that matches your IT policy.

Best interactive route

Host this HTML as a static intranet asset or approved web app, then embed it in SharePoint with an approved domain or iframe route.

Simplest route

Upload the HTML to a SharePoint document library and link to it from a modern page. Users may need to open it as a file depending on tenant settings.

No-script route

Copy the rules, workflow, model policy, and prompt templates into SharePoint text web parts. Use the PowerPoint/Synthesia version for launch comms.

Sources and assumptions

These links are included so the owner of the SharePoint page can re-check model names, limits, and rates when OpenAI updates the product.

OpenAI Help Center, Codex rate card: token-based credits for GPT-5.5, GPT-5.4, GPT-5.4 mini; typical GPT-5.5 task range; Fast mode note. https://help.openai.com/en/articles/20001106-codex-rate-card
OpenAI Developers, Codex Speed: Fast mode speed and credit multipliers. https://developers.openai.com/codex/speed
OpenAI Developers, Codex Best Practices: Plan mode, reasoning levels, AGENTS.md. https://developers.openai.com/codex/learn/best-practices
OpenAI Developers, Codex Pricing: usage limits, smaller-model guidance, AGENTS.md and MCP context guidance. https://developers.openai.com/codex/pricing
OpenAI Help Center, ChatGPT Business Models & Limits: Instant/Thinking/Pro limits and context windows. https://help.openai.com/en/articles/12003714-chatgpt-business-models-limits
OpenAI Help Center, Tokens: input, output, cached, reasoning tokens and context limits. https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them
Microsoft Learn and Support, SharePoint custom script and embed controls. https://learn.microsoft.com/en-us/sharepoint/allow-or-prevent-custom-script and https://support.microsoft.com/en-us/office/allow-or-restrict-the-ability-to-embed-content-on-sharepoint-pages

Print note: interactive elements are expanded where possible, but the calculator requires JavaScript.