I would rather give ChatGPT a clear acceptance test than an impressive job title. “Act as a world-class strategist” leaves the meaning of world-class entirely unresolved. “Cite each account claim and flag missing budget” gives the reviewer something to check.
The habit I want is a reusable brief. Stable preferences live once; changing facts arrive with the task. That keeps prompting from becoming a very elaborate form of copy and paste.
Put each instruction at the right level
Personal preferences include response length, terminology and a writing style you consistently want. Project requirements describe a particular body of work, such as the audience for a launch and the approved source of positioning. Task instructions specify today's deliverable and inputs.
Mixing the three creates prompt clutter. A global preference for short answers may be useful, but today's request for a detailed research report should make the required depth explicit. A campaign date belongs with the campaign, not in personal instructions that follow you into unrelated work.
I would sort an existing long prompt into these categories before rewriting it. Much of the length often comes from carrying setup that should have a durable home. The remaining task brief becomes shorter because it is more specific, not because the standards disappeared.
TipKeep temporary campaign facts out of global personal preferences.
Write preferences a person can maintain
Useful custom instructions are concrete: write in plain English, avoid unsupported numbers, distinguish facts from recommendations and ask a focused question when a missing detail changes the result. These are behaviors someone can inspect in the output.
Avoid collecting dozens of contradictory style demands. A request to be exhaustive, extremely brief, conversational and formal in every answer is a negotiation nobody has resolved. Decide which preferences are defaults and which should vary by task. The model should not be the first person asked to reconcile the organization's writing strategy.
Open the available personalization settings and review what is already there before adding more. In Codex, personal instructions can be represented through global AGENTS.md guidance. The controls differ by surface, so keep the scope of the change clear.
TipRead the existing preferences first; adding a second rule may preserve the contradiction.
Use the five-part task brief
I use five fields: outcome, audience, sources, constraints and acceptance checks. For an account brief, the outcome is one page, the audience is the AE, the sources are dated notes, the constraints exclude unsupported claims and the checks require a citation for each account fact.
This is enough structure to help without turning the request into a programming language. Write naturally. If the output can take several valid forms, name the choice that matters to you: a document, a spreadsheet, a short answer or a proposed repository change.
The acceptance checks deserve particular attention. “Make it accurate” is an aspiration. “Reconcile the total to the source export and show excluded rows” is a procedure. The latter tells both the agent and the reviewer how to recognize a problem.
TipUse a check that could actually fail on a plausible but wrong answer.
Give examples that reveal the standard
An example should show a decision, not merely a polished sentence. For lead scoring, include a record with missing employee count and show why it stays unknown. For copy, include a claim that should be removed because the proof source does not support it.
One positive and one difficult example can teach more than a page of tone adjectives. Keep the explanation with the example so the agent can generalize the rule. “This is concise” is less useful than “this keeps the buying concern, removes an unsupported performance claim and ends with one specific next step.”
Use clearly labeled fictional examples when you do not have approved customer material. The examples in this hub follow that approach. They illustrate a method and expected reasoning without pretending to be measured outcomes from a real deployment.
TipExplain why the example passes, not only what it says.
Ask for clarification where it changes the work
A task can be underspecified in ways that matter. If the report's audience could be finance or sales leadership, the required detail may change. If a spreadsheet mixes currencies, the calculation cannot be interpreted correctly without a conversion rule. Ask the agent to identify those decisions early.
Do not require a long interview before every simple request. A small rewrite can proceed on reasonable assumptions. The useful distinction is whether the missing detail changes the correctness or purpose of the result. If it does, resolve it; if it only affects presentation, a stated assumption may be enough.
For Cedar Metrics, a campaign analysis should ask about the definition of a qualified lead when the export only says “lead.” It can choose a clean table layout on its own. That is the level of judgment I would expect from a competent colleague.
TipTell the agent which assumptions it may make and which need an answer.
Revise the cause of the failure
When a draft misses the mark, name the failure precisely. An unsupported claim needs an evidence correction. A bloated opening needs an editorial correction. A file that cannot be opened needs an access fix. These problems do not all improve by making the prompt more emphatic.
Give feedback tied to the output: “The second paragraph says finance approved the purchase, but the notes only mention a future meeting. Correct that claim and inspect the remaining brief for the same inference.” This is actionable because it identifies the error and the evidence.
After the revision, decide whether the lesson belongs in reusable guidance. If it is a recurring requirement, add it to the relevant instruction layer. If it is specific to one account, keep it in the task. Otherwise your permanent instructions become a museum of every mistake ever observed.
TipFix the reusable rule only when the correction should generalize.
Turn a good prompt into a repeatable workflow
A prompt becomes reusable when its inputs and expected outputs are clear enough for another person to supply them. Replace account-specific facts with named input requirements, keep a concrete example and preserve the acceptance checks. Do not remove the quality standard while making the wording generic.
For a campaign brief, require a current positioning source, an audience description and approved evidence. If any is missing, the workflow should say what it can still produce and what remains blocked. That is a more reliable template than one which always returns a confident complete brief.
When the workflow has several steps or supporting resources, consider a skill. When it needs connected tools and distribution, consider a plugin. The skills and MCP guide explains those choices. Reuse should make the process inspectable, not hide it behind a memorable prompt name.
TipKeep an example input and accepted output beside the reusable instructions.
Where prompting advice becomes superstition
The recurring trap is adding intensity instead of information. Repeatedly asking the model to think harder does not supply the missing account note or settle a conflicting business rule. Use available reasoning controls when appropriate, but diagnose the evidence first.
Another trap is treating a single successful answer as proof that the prompt is robust. Try a sparse input, a contradictory source and a request outside the intended scope. A good workflow should preserve uncertainty and decline unsupported conclusions while still producing useful partial work.
I would keep the brief short enough to read, the examples specific enough to teach and the checks strict enough to catch a real error. That is a process a team can improve together. Which sentence in your current prompt actually tells the reviewer what finished means?
How to set it up
Review existing instructions
Inspect personal and project guidance for contradictions, obsolete facts and unnecessary repetition.
Write the task brief
Use this structure: outcome, audience, sources, constraints and acceptance checks.
Example: “Create a one-page Cedar Metrics launch brief for sales. Use positioning-2026-09-22.md and approved-proof.md. Separate facts, recommendations and unknowns. Cite each quantitative claim and return a draft for review.”
Add a difficult example
Include one missing-data or unsupported-claim case and explain the expected treatment.
Test reuse
Run the instructions with a second input. Save the accepted examples and move stable requirements to the appropriate instruction layer.
Frequently asked questions
What belongs in custom instructions?
Stable preferences that should apply across relevant chats, such as communication style and evidence habits.
What belongs in a task prompt?
The specific outcome, audience, inputs, constraints and checks for the current job.
Should I use roleplay?
A role can provide context, but concrete sources and acceptance criteria do more work than an elaborate fictional title.
How long should a prompt be?
Long enough to specify the decisions that matter. Remove repeated setup and conflicting instructions.
Can examples improve the result?
Yes, especially when they explain why a difficult case should be handled a certain way.
Should ChatGPT always ask questions first?
Only when missing information materially changes the result. Simple tasks can proceed with reasonable stated assumptions.
When should I make a skill?
When a proven workflow needs reusable steps, supporting resources or consistent checks across repeated tasks.
How do I test a prompt?
Use representative, sparse and conflicting inputs, then score factual errors, omissions and review effort.
Sources & further reading
ChatGPT and Codex change quickly. This page was last reviewed September 22, 2026; verify time-sensitive details against the official docs above before relying on them.
Related GTM workflows
Use these existing playbooks to explore the business workflow. Adapt their tool-specific steps to your chosen environment and check the result.