Teach the tool
how you work.
An LLM can be an extraordinarily capable partner. But it starts with almost no knowledge of your work, your standards, or the way you like to make decisions.

Smart is not the same as informed.
The model may know more general information than any one person. That does not mean it knows what matters in your classroom, your small business, your project, or your life. It does not know your audience, your definition of done, the things you never want it to assume, or the way you want it to speak to you.
That knowledge is not a limitation. It is the work of partnership. You provide the purpose, taste, context, and boundaries. The LLM brings speed, breadth, and detailed execution.
There are two kinds of instructions
The day-one syllabus
Always-on instructions for a project or workspace: what this work is, who it serves, the stack or tools, quality standards, and non-negotiable guardrails.
Use it when the instruction should shape almost every interaction.The focused playbook
A self-contained procedure for a repeatable job: create a lesson plan, review a pull request, make a slide deck, turn a report into a visual, or prepare a client update.
Use it when you want a particular kind of work done well.The exact filename changes between tools. You may encounter names such as CLAUDE.md, AGENTS.md, or another project rule file. Do not get hung up on the label. They are all plain Markdown ways to tell an AI partner how to show up before it starts working.
Build the foundation first
Think of durable context as the syllabus you hand someone on their first day. It should be short enough to stay useful, but clear enough to prevent the same questions and mistakes from happening again.
"Help me write a short project instruction file for this work. Ask me only the questions you need about the audience, the outcome, my working style, privacy boundaries, and what I consider done. Then give me a Markdown draft I can edit before we use it."
Skills are where improvement compounds
A skill is not a giant prompt you paste over and over. It is a reusable capability package. When you notice a task coming up for the third or fourth time, stop and ask: what did I have to explain again? That is the beginning of a skill.
Identity and objective
Define the role, operating baseline, and the result this skill is responsible for.
Inputs and outputs
Say what the skill needs and what a good finished result looks like.
Guardrails
List the no-fly zones, privacy limits, and when the agent must stop to ask.
Workflow
Give the repeatable sequence - including when to use a tool or show work for review.
Examples
Show one or two examples of the kind of answer or artifact you actually want.
A portable skill template
Start simple. You can move the same thinking between AI tools even if each one has a different folder, command, or trigger mechanism.
# Skill: [name / purpose] ## Identity and objective - Role, baseline, and core mission. ## Input and output - What I will provide. - What a finished result must look like. ## Guardrails - Never do ... - Do not guess about ... - Stop and ask when ... ## Workflow 1. First step. 2. Review point. 3. Final check. ## Examples Input: ... Output: ...
Start with a real repeat
Your first skill does not need to be impressive. It needs to be useful. A teacher could make a skill for turning a learning objective into a classroom activity. A small business owner might make one for product descriptions in their voice. A manager might make one for turning rough notes into a clear update.
"We have done this task several times. Review the way I corrected you and the final results I approved. Draft a small reusable skill that captures the goal, inputs, workflow, output format, and things you should never assume. Keep it practical and show me the draft before using it."
You are still in the loop
Context and skills make the LLM more consistent. They do not make it infallible or give it authority you did not intend. Keep the decisions that carry consequences: what is shared, what is changed, what is true, and what represents you. The better your instructions get, the more room you have for that kind of judgment.
Onboard your AI.
Give a new assistant the right personal context so it becomes helpful faster without handing over more than it needs.
Return to the learning path ->