AI TOOL SETUP · START SIMPLE
Make your AI
work your way.
Give your AI useful context, clear preferences and a repeatable way to check its work. Start with everyday chat setup; use the coding section only if you are building software.
Official guidance checked 22 October 2026. Settings and features vary by product, account, plan and organisation. This page creates instructions for you to copy; it does not change your AI account.
Your first 10 minutes
- Choose one real task. Pick something you repeat, such as rewriting emails, reviewing notes or planning content.
- Add a short personal profile. Describe your role, experience, preferred language, tone and answer length. Keep temporary tasks out of account-wide instructions.
- Give the task its own context. Supply the audience, goal, relevant files and what a successful result looks like.
- Ask useful questions first. Ask the AI to clarify missing information that would materially change the result. It need not interview you before every small task.
- Test and adjust. Compare the same task before and after setup. Keep changes that improve accuracy, usefulness and consistency.
Put the right information in the right place
| Layer | Use it for | Example |
|---|---|---|
| Personal instructions | Stable preferences across your work | Plain Australian English; concise reports; explain unfamiliar terms. |
| Memory, where available | Selected ongoing facts and preferences | Your usual role or recurring interests. Review and remove outdated entries. |
| Project instructions and sources | One client, subject or ongoing piece of work | Audience, terminology, approved source documents and decisions. |
| Skills or reusable workflows | A repeatable task with a defined method | Review a report against a checklist and return a findings table. |
| Current prompt | The immediate task and exceptions | Draft this week's update in 150 words using the attached notes. |
A project is not a guarantee that every previous conversation or file will be recalled. Keep important decisions in a concise, dated brief and supply the current source when accuracy matters.
Choose your tool
ChatGPT and Codex
Use custom instructions for stable response preferences. Keep project-specific requirements with the project. In Codex, personal instructions use a global AGENTS.md; repositories may also contain their own instructions. Review what applies before changing a project.
Claude chat
Use Settings → Instructions for Claude for general preferences. Use project instructions for a specific body of work, and skills for repeatable behaviours or specialised tasks. Add a short approved writing sample if you want help defining your style. Check the features available in your account.
Official Claude personalisation guide ↗Gemini
A custom Gem can hold instructions for a recurring task. Define its role, context, task and desired format, then preview it on a real example and save it. Start with one focused workflow.
Official Gem setup tips ↗Copilot and other assistants
First identify the exact product and account you use: a personal assistant and an organisation-managed assistant may have different controls. Check its current help and settings for personal instructions, memory and reusable workspaces. If a feature is unavailable, paste a short reusable brief at the start of the chat. Use your approved work account for work material.
Copilot Chat training → · Microsoft 365 Copilot training → · Other AI tools →PERSONALISE · REVIEW · COPY
Build your personal instructions
Enter only information you are comfortable sharing. These fields stay on this page and are not saved when you leave. Review the generated text, then paste it into your AI's supported instruction setting or a new chat.
Give ongoing work a reliable brief
Create one project per distinct subject or client where supported. Add a short purpose statement, audience, source list, terminology, expected output and review criteria. Use dated, approved documents and remove superseded versions. An example of good work is often more useful than a long list of adjectives.
Useful opening request: “Before drafting, ask up to three questions about missing details that would materially change the result. Otherwise state reasonable assumptions and proceed.”
Connected tools: add only what the task needs
Start with a named source or folder. Check the account, access scope and whether the integration can only read or can also change information. Test a low-risk read first. Keep human review for messages, purchases and publication, and remove connections you no longer use. Available integrations and administrator controls differ between products.
Keep reusable instructions short
Give each skill or workflow a purpose, required inputs, steps, expected output and a quality check. Remove contradictory or obsolete rules. Instructions guide behaviour; they do not guarantee correctness or enforce access controls.
Choose coding components with purpose
Before adding a skill, agent or integration, identify a real gap in your existing setup. Review the supplied files and trial one component at a time.
OPTIONAL · FOR SOFTWARE PROJECTS
Set up your coding assistant
- Start with the existing project. Read its README, instructions and scripts. Follow the current official installation guide for your operating system. Claude Code's quickstart offers a native installer; Node.js, Docker and a particular editor are not universal prerequisites. Install the runtime your project actually uses.
- Define a small outcome. Record the user problem, requirements, constraints, exclusions and acceptance criteria in a brief such as
PRD.md. The filename alone does not make an AI load it; reference the brief. - Add project instructions. For Claude Code, use
CLAUDE.md; for Codex, use the supportedAGENTS.mdinstructions. Include verified setup commands, architecture notes, coding conventions and meaningful checks. Keep credentials out of these files. - Verify the starting point. Run the existing build or relevant tests, note failures and establish version control or a recoverable checkpoint.
- Work in small steps. Implement one feature, run relevant checks, inspect the changes and commit a coherent result. Ask the assistant to distinguish tests it ran from tests it only recommends.
- Add integrations and agents when useful. Use bounded tasks and explicit permissions. Extra tools, agents and infrastructure add coordination effort; introduce them to solve a demonstrated need.
Claude Code quickstart ↗ · Claude Code project instructions ↗
Check whether your setup helps
Use three small examples: rewrite an email, summarise supplied notes and explain a topic at your level. Run the same input before and after changing one setting. Score each result from 1–5 for instruction-following, factual accuracy, useful detail and tone. Check facts against the supplied material; do not rely only on the AI rating itself.
Still too generic?
Add the audience, intended decision, a specific example and constraints. Ask the AI what key context is missing.
Too long or too formal?
Specify a word range and show a short example of your preferred writing. Remove conflicting style instructions.
Wrong facts or outdated context?
Supply the current source, date it, and ask the AI to distinguish evidence from assumptions. Correct or remove stale memory. Use browsing for changing facts where available and check the cited sources.
Missing features or inconsistent behaviour?
Check the exact product, account, plan, platform and administrator settings. Test in a new chat. If persistent instructions are unavailable, use a reusable opening brief.
Review your setup when your role, project or tools change. Keep the smallest set of instructions that consistently helps.
Explore prompts.chat resources
Optional community resources for practising prompting and exploring integrations. Review examples before using them and check the current documentation for your tool.
These are external resources; following a link does not install or connect anything.
