Help me write custom instructions for ChatGPT or Claude that improve its default behavior across my recurring work. My role and regular work: [ROLE AND RECURRING TASKS] Who receives my work: [AUDIENCE] My preferred voice and format: [VOICE AND FORMAT] Rules, sources, and boundaries: [RULES AND BOUNDARIES]
Common mistakes I want prevented: [COMMON FAILURE MODES]
First, identify which requirements belong in permanent custom instructions and which should stay in task-specific prompts. Then produce two versions: a compact version of no more than 180 words for a limited instruction field, and a fuller version of 300-400 words for a project-level instruction set.
Organize each version under short labels: Context, Working method, Output standards, Evidence and uncertainty, and Boundaries. Write direct operational instructions, not personality claims. Include rules for stating assumptions, distinguishing facts from recommendations, preserving supplied wording when requested, and asking no more than 3 clarifying questions only when an answer cannot be produced safely without missing information.
Do not make the AI claim access to files, data, tools, sources, or expertise it does not have. Do not create instructions that force a fixed tone or format when the task calls for something different.
Before answering, check each instruction is observable and actionable. Also remove duplicates, contradictions, vague terms such as “high quality,” and rules that would harm useful brevity.Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [ROLE AND RECURRING TASKS] | Describe your job or studies and the work you repeatedly ask the AI to help with. | Operations manager at a regional clinic network; I draft SOPs, meeting summaries, project plans, and staff communications. |
| [AUDIENCE] | Name the people who read or use the outputs and what they expect. | Clinic managers and frontline staff who need plain English and practical next steps. |
| [VOICE AND FORMAT] | Describe the desired tone, reading level, language, and common output structures. | Calm, direct, eighth-grade reading level; use short headings, bullets, and a final action list. |
| [RULES AND BOUNDARIES] | List policies, confidentiality limits, required sources, citation needs, and actions the AI must not take. | Never include patient information; cite supplied policy documents; do not provide medical advice or invent policy requirements. |
| [COMMON FAILURE MODES] | List recurring problems in AI answers that you want the instructions to prevent. | It buries the decision, repeats my notes, makes up process details, and produces overly formal language. |
How to use
- List only the rules that should apply to most of your chats; keep specialized instructions in separate project prompts.
- Paste the compact version into your AI tool’s custom-instructions area or use the longer version in a project.
- Test it with a familiar task and mark any rule that creates repetitive or unwanted behavior.
- Follow up with: “Rewrite these instructions to remove rules that do not apply to at least half of my requests.”
Variations
Student instructions
Use for recurring study, research, and writing help.
Create custom instructions for an AI study assistant. I study [SUBJECTS] at [LEVEL], and I need help with [RECURRING TASKS]. My school rules are [ACADEMIC RULES]. Produce a 180-word instruction set that requires teaching concepts before giving answers, showing reasoning for quantitative work, separating quotations from paraphrase, and flagging unsupported claims. It must not complete work that violates my academic policy. Ask up to 3 questions only if required input is missing. Check for vague or conflicting rules before answering.
Manager instructions
Use when you write plans, decisions, and team communications repeatedly.
Write custom AI instructions for me as a [MANAGER ROLE]. My team and audience are [TEAM AND AUDIENCE]. My recurring outputs are [OUTPUT TYPES], and our operating constraints are [CONSTRAINTS]. Produce concise instructions that lead with decisions, distinguish confirmed facts from risks, assign an owner and due date only when supplied, and end plans with open questions. Do not invent commitments, numbers, or stakeholder views. Check every rule is usable across recurring work.
Coding instructions
Use for regular software development assistance.
Create project-level AI instructions for a developer working in [TECH STACK] on [PROJECT TYPE]. Coding conventions and commands are [PROJECT RULES]. Produce instructions requiring the AI to inspect existing patterns before proposing changes, state assumptions, avoid unrelated refactors, add or update relevant tests, and report verification performed. It must not claim code was run or files were read unless that happened. Keep it under 250 words and resolve contradictory rules.
Tips
- Put durable preferences in custom instructions, but include current goals, source material, and deadlines in each individual request.
- Phrase standards as testable behaviors: “cite supplied policy by title” is more useful than “be accurate.”
- Avoid asking for a single fixed format in every chat; it can make brainstorming, coding, and short questions awkward.
- Revisit instructions after a week of real use and remove rules that are rarely relevant or routinely ignored.
FAQ
Do custom instructions replace a detailed prompt?
No. They establish defaults; a task prompt still needs the specific objective, source material, audience, and constraints.
Should I include confidential information?
Do not put sensitive client, patient, account, or internal information in reusable instructions. Use approved tools and policies for sensitive work.
How long should custom instructions be?
Short enough that each rule remains memorable and broadly applicable. A compact set of operational rules usually works better than a long biography.