Guide

ChatGPT custom instructions examples for work, with a template

Good custom instructions for work are five short parts: your role and what a good answer does for you, who reads the output and the tone, what the AI must never do and what to do instead, your default format, and one situation where it tends to go wrong. Everything else, like a report template or a client's details, belongs in a project or a skill, because custom instructions apply to every chat. Below are four examples in that shape and a template to copy.

Short answer

  • Role and goal: two sentences.
  • Audience and tone: who reads it, two tone rules.
  • Never, and instead: the promises it must not make.
  • Format and one edge case. Under 15 lines in all.

The template

Role and goal: I'm a [role] at [type of company]. A good answer
[what it lets me do].

Audience and tone: Usually read by [who]. [Tone rule]. [Tone rule].

Never, and instead: Never [risky thing]; instead, [what to do].

Format: Default to [length and shape].

Edge case: If I [common situation], [what to do].

Four examples

Account manager

Role and goal: I manage 12 B2B clients for a packaging supplier. A good answer is ready to send to a client's buyer.
Audience and tone: Buyers and their managers. Plain and warm; no exclamation marks.
Never, and instead: Never promise delivery dates or discounts; say I'll confirm with operations.
Format: Under 150 words, the ask in the last line.
Edge case: If I paste a complaint, draft the reply and list what to check before I send it.

Operations lead

Role and goal: I run operations at a 40-person manufacturer. A good answer turns a messy situation into decisions and owners.
Audience and tone: Shift leads and the managing director. Short sentences; numbers before adjectives.
Never, and instead: Never guess a stock level or lead time; mark it [check] and say where to look.
Format: Bullets: what happened, decision needed, owner, by when.
Edge case: If two requests conflict, list both and ask me which wins.

Recruiter

Role and goal: I recruit for a 60-person software company. A good answer saves me writing time without deciding anything about a candidate.
Audience and tone: Candidates and hiring managers. Direct and friendly; no "rockstar", no "fast-paced family".
Never, and instead: Never rank or score candidates; give me the questions to decide with.
Format: Job posts 300 to 600 words; outreach under 120 words.
Edge case: If I paste a CV, remind me to remove the name and contact details first.

Founder

Role and goal: I run a 9-person design studio. A good answer helps me decide fast and tells me what it's unsure of.
Audience and tone: Mostly me; sometimes clients and the team. Blunt; no filler.
Never, and instead: Never invent figures; ask me for them.
Format: The answer in the first line, then at most five bullets.
Edge case: If I ask for a legal or tax view, give the questions to ask a professional.

Illustrations: made-up roles and businesses.

A worked example: trimming a page of instructions

Illustration: a made-up business. Marta, an account manager, had a full page of custom instructions: her role, the company's product list, three client names with their quirks, a report template and twenty tone rules. Answers to unrelated questions kept mentioning her clients.

She cut it to the five parts in the first example above. The product list went into a project; each client's quirks went into that client's project; the report template became a skill. Her custom instructions dropped from 60 lines to 8, and a question about a spreadsheet formula no longer came back with a note about a client's delivery window.

Write yours

  1. Copy the template and fill the five parts for your job.
  2. Cut every line that isn't true in every chat; move it to a project or a skill.
  3. Name the promise it must never make, and what to say instead.
  4. Test it on three different requests: a reply, a summary, an unrelated question.
  5. Add a rule only when a test fails, and keep the whole thing under 15 lines.

Not sure what belongs where? See custom instructions vs Projects vs skills.

Next step: one set of instructions for the team

Custom instructions are personal; they can't be shared. When the whole team should answer clients the same way, the shared part belongs where everyone's AI reads it: getting your team to use AI the same way covers that, and sharing ChatGPT custom instructions with a team covers the ChatGPT side.

Skills for this

All skills

Questions

What should I put in ChatGPT custom instructions for work?

Only what's true in every chat: your role and what a good answer does for you, who usually reads the output and the tone, what the AI must never do and what to do instead, the default format, and one edge case. Task rules, like a report template, belong in a project or a skill.

Can you give an example of custom instructions?

For an account manager: 'I manage 12 B2B clients for a packaging supplier; a good answer is ready to send to a client's buyer. Plain, warm, no exclamation marks. Never promise delivery dates; say I'll confirm with operations. Default to under 150 words with the ask in the last line. If I paste a complaint, draft the reply and list what to check first.'

Are custom instructions the same in Claude?

Claude's equivalent is profile preferences, which also apply to every chat. The same short shape works in both.

AI skill

Have your AI write the instructions for your AI assistant

Put this skill in AI Skill Pod and everyone on your team gets it in their Claude, Cursor or ChatGPT (Pro, Business, Enterprise): one invite link, free to start. Context brief for an AI assistant has your AI write these instructions for you from a few questions: the role and goal, audience and tone, what to refuse and do instead, the output format and the edge cases, short and in order of importance.

Use this skill in your AI

14 days free · No card needed

Sources

  1. OpenAI Help: ChatGPT custom instructions — applied across your chats; read through search-result text, the page blocks automated reads (read 2026-11-05)
  2. Anthropic: Effective context engineering for AI agents — the smallest set of high-signal information (read 2026-11-05)
  3. Claude Docs: Prompting best practices — clear instructions and examples (read 2026-11-05)
  4. Claude Help Center: personalization features — profile preferences, Claude's equivalent; read through search-result text (read 2026-11-05)