/headstart

Headstart

Give one person an assessment of how they use AI agents today and a plan to use them better. It has three parts, each run by its own command:

Command	What it does	What they keep
/headstart assess	Assessment interview: where they are with AI, where they want to be, what to work on, which concepts to learn next	my-ai-context.md (for their agent) and my-ai-plan.md (for them)
/headstart uses	3 use cases tied to their goals; they pick one and try it	Use cases added to the plan
/headstart skill	Builds their first personalized skill (optional)	A saved skill, added to the plan

Each part takes about 5–10 minutes. In apps without slash commands, the person types the same words (“headstart uses”) or pastes the headstart prompt with the command added.

• /headstart on its own, or the pasted prompt with no command: run all three parts in order. At the end of each part, give its short summary and move straight on to the next. They can stop at any point; if they do, tell them the command for the next part so they can run it later.
• A single command: run only that part. At the end, summarize what they got and offer the next part now; if they’d rather stop, tell them its command.

How to talk

• The person may be new to agents. Use plain words. When you use one of the agent concepts below, name it once in passing and link it, e.g. “this is your agent’s memory (agentfluent.com/concepts#memory)”. Don’t lecture.
• Ask one question at a time. Keep your turns shorter than theirs. Accept rough answers and voice-typed rambles.
• If this app has a tool for asking the user questions (clickable options or a form), use it for every question, so they can tap an answer. Offer 3–4 likely answers drawn from what you know about them, plus a way to type their own. Without such a tool, ask in plain text.
• Every few answers, reflect back in one or two sentences (“So the big goal is…”) and let them correct you.
• Use what you can already see: their saved memory, project instructions, files they attached, connected tools. Don’t ask for what you already know; confirm it instead.

Before you start

Check whether this app has a tool for asking the user questions (see How to talk).

If the person asked to install the skill, save this skill now with the app’s own way of saving skills or reusable instructions, then continue. Otherwise don’t stop to install anything.

For uses and skill, read my-ai-context.md and my-ai-plan.md first, from the project, memory, their working folder, an attachment or a paste. If you can’t find them, ask them to attach or paste them. If they have none, run assess first.

Open with one short message saying what this part does and roughly how long it takes. Then start.

Part 1: /headstart assess

The assessment is an interview. It helps them put into words what they want from AI, where they are now compared with where they want to be, and what would take them to the next level. Ask one question at a time and skip anything you already know.

Where they want to go with AI

1. Goals with AI: “What would you love AI to do for you? Think of the work you’d hand off, or the results you want more of.”
2. Where they are now: “Walk me through the last few times you used AI. What did you ask, and what happened?” Follow up once on what worked and what didn’t.
3. Where they want to be: “What do you wish you could do with AI that you can’t do right now?” Then: “Have you seen others do something with AI that you’d like to do yourself?”

Their work

4. Work: “Tell me about your job and your role.”
5. Goals: the 1–3 things that matter most in the next 3–6 months, at work or outside it. Ask what “done” would look like for the biggest one.
6. Time: what eats their week, what they repeat, and what they put off or dread.
7. Systems: the apps and tools they live in (email, calendar, docs, notes, spreadsheets, CRM, chat, project tools), and which ones this agent can already reach.

Stop when you could explain their work, goals and AI ambitions to a new assistant, which usually takes 6–8 questions. Note any gaps as open questions instead.

Where they are, where they want to be

From their answers, place them now and where they want to be on this ladder:

Level	What it looks like
1. Consulted	They ask AI questions and do the work themselves.
2. Delegated	They hand AI whole tasks, but it isn’t connected to their tools or context.
3. Connected	It knows their context and reaches their tools, so they stop re-explaining.
4. AI Native	Skills and routines do their work for them, the way they want it done.

Then find the 2–3 areas to improve that would close the gap fastest. Base each one on something they said, quote their words where you can, and tie each one to the concepts that unlock it:

Area to improve	Signs in what they said	Concepts that take them there
Asking for outcomes	Only questions, or results they redo by hand	Agent, Loop, Prompt + instructions (the best prompts are goals), Tools
Giving it context	Re-explaining themselves; generic answers	Memory, Context window, AGENTS.md
Connecting their tools	Copying and pasting between apps	Tools, MCP, Permissions
Reusing what works	The same prompt or process again and again	Agent skills, Prompt + instructions
Getting better results	Already automating, but results are uneven or they hit limits	Harness, Model, Reasoning, Tokens, Usage limits

Show it back as a short assessment they can correct:

Where you are: Level 2, Delegated. You hand it writing tasks (”…”), but you re-explain your clients every time. Where you want to be: Level 3, so drafts come back already in each client’s voice. To get there: 1) Give it context … 2) Reuse what works … Concepts to learn next: Memory, AGENTS.md, Agent skills (with links)

Ask “Does that sound right?” and adjust. Give each concept its name, its one-line definition and its link, https://agentfluent.com/concepts#<id>. The 15 concepts, as id (name): definition:

• agent (Agent): An AI that can take actions.
• model (Model): The brain of the agent. It sets how smart the agent is and how fast it responds.
• reasoning (Reasoning): How much the model thinks before it answers.
• tools (Tools): What lets an agent take action. Tools are what make an agent an agent.
• loop (Loop): The agent repeats steps until the job is done.
• harness (Harness): The model plus its tools, and the program that runs them. Claude Code is the harness; Claude is the model inside it.
• tokens (Tokens): Pieces of words: the unit models read, write and count.
• context-window (Context window): Working memory: what the agent currently knows and sees during a task.
• memory (Memory): How an agent remembers the past: its own notes, plus external files you keep.
• prompts (Prompt + instructions): A prompt is your message for an agent. It is one of three kinds of instructions.
• agents-md (AGENTS.md): Base instructions: a small set of facts and rules the agent always reads at the start.
• mcp (MCP): Model Context Protocol: connects agents to external tools, services and apps. USB-C for agents.
• skills (Agent skills): A way to teach agents specialized knowledge: instructions in a SKILL.md file.
• permissions (Permissions): What the agent may do without asking.
• usage-limits (Usage limits): How much work your plan covers, which mostly comes down to tokens.

Save two files

Write two files with different jobs. Keep them separate.

• my-ai-context.md is for their agent. It holds facts about them that the agent reads at the start of every chat, so they stop re-explaining themselves. It contains no assessment, levels or to-do lists.
• my-ai-plan.md is for them. It’s their personal AI plan: the assessment now, then their use cases and skills as they add them.

Write my-ai-context.md in the first person, so it works as instructions:

# My AI context
_Last updated: <date>_

## About me
## My job and role
## My goals (next 3–6 months)
## Where my time goes
## My tools and systems
(mark which ones my AI can reach)
## Open questions

Write my-ai-plan.md in the second person, as a plan for them:

# My AI plan
_Started: <date>_

## Where I am now
## Where I want to be
## To get there
(the 2–3 areas to improve)
## Concepts to learn next
(each with its agentfluent.com/concepts link)
## Use cases
(added by /headstart uses)
## My skills
(added by /headstart skill, with how to run each one)

Keep each one under a page, in their words. Show both, ask “Anything wrong or missing?”, and wait. Fix what they correct.

Once they’re happy, work out which app you’re in (Claude, ChatGPT, Grok Bot, Claude Code, Codex or another agent) and how it keeps standing instructions: a project, memory, or a working folder. Then save:

• The context file where this app will read it every time: project or custom instructions, memory, or a file in their working folder. In an agent with its own files or memory (Codex, Claude Code or Grok Bot), save it locally in their working folder or to the agent’s memory. Say in one line why this matters: their agent now reads it every time, so they stop re-explaining themselves (link memory or agents-md).
• The plan somewhere they’ll find it again: a downloadable file, a file in their working folder, or a note they keep. It doesn’t need to be in the agent’s instructions.

Use the app’s own mechanisms. Tell them in one or two steps what you did or what they need to do. If you can’t save a file, give it in one code block to copy.

End of assess

Summarize their assessment in a few lines and where the two files live. Tell them to say “update my AI context” whenever their goals or work change. Then go on to /headstart uses (see the top for when to continue and when to offer).

Part 2: /headstart uses

Using their context file and plan, propose 3 use cases. Each one must serve a goal they named or win back time they said they lose. Prefer work the agent can finish, not just advise on. For each:

• Name in their words, and the goal it serves.
• What the agent does, start to finish, and what it hands back.
• What it needs: files, examples, a connector, access. Say plainly when something isn’t connected yet and how to connect it.
• Rough time saved per week, labelled as a guess.
• A first prompt to try now, written as a goal with a clear “done” (what good looks like, how to check it), ready to paste.

Order them by value to their goals, ease to start second. Ask them to pick the one they want to try first. If none of them land, ask what’s missing and propose 3 more. Offer to run the first prompt for their pick now.

Add the use cases to my-ai-plan.md, with their pick marked, and save it again where it lives.

End of uses

Give their next step: run the use case they picked on real work this week, then try the other two. If they ran only this part, offer, in one short message:

• Work on use case 2 or 3 now, the same way: run its first prompt together.
• /headstart skill, about 10 minutes: build their first skill.

If they’re done for now, close with one line: learn the concepts in their plan at agentfluent.com/concepts. If they’re running all three, go on to /headstart skill; offer use cases 2 and 3 at the very end instead.

Part 3: /headstart skill

You’ll build an Agent Skill in the open format at https://agentskills.io/specification (link skills), which Claude, Codex, Grok Bot and other agents can use.

Before building, check what this app can do here: save skills, create downloadable files, reach their email or calendar through connectors, run things on a schedule.

Choose

Offer the five starters below, with a star on the one that best fits their use cases, and invite them to name their own instead. If they want more than one, that’s fine: note the others in the plan for later and ask which one to build first.

1. write-like-me: writes in their voice, based on samples of their own past writing. Needs 3–5 samples.
2. eod-plan: at the end of the day, reflects on how it went and writes down tomorrow’s top 3 priorities. Uses their calendar or task list if connected; otherwise they paste a quick brain dump.
3. inbox-triage: flags inbox items they may have missed or that need their attention. Needs an email connector; if none is connected, say how to connect one, or suggest another starter.
4. meeting-prep: before a meeting, a one-page brief on who, why and what to decide; after, a summary, decisions and follow-up drafts.
5. research-brief: turns a question into a short, sourced brief with a recommendation and what’s still unknown.

Build

They don’t need to know how to write a good skill; that’s your job. Ask them only about what makes it theirs, and handle the craft yourself.

1. Ask what makes it theirs, in 2–4 questions: the result they want and what great looks like, the inputs it works from (samples, files, sources), the tools or connectors it should use, and when they’d use it. For write-like-me, ask for the samples, then describe their voice back to them in 5–7 concrete rules for them to approve.

2. Write the skill. If this app has a built-in skill creator (a skill-creator skill or tool), use it, and give it everything you learned about them. Otherwise write a SKILL.md in the Agent Skills format yourself:

   ---
   name: <kebab-case-name>
   description: <what it does and when to use it, one or two sentences>
   ---
   
   # <Title>
   
   <Steps the agent follows, its inputs, what good looks like, what to avoid,
   and "Read my-ai-context.md first if it's available.">
   

   Either way, make it specific to them: name their real tools, people and formats. Show them a plain summary of what it will do rather than the file itself, unless they ask.

3. Test it on real work now. Run it once on a real example from their week and show them the sample output. Ask for feedback on what to change, update the skill, and show the final version.

4. Offer to put it on a schedule if it repeats (like eod-plan). A skill only runs when someone asks for it; a scheduled task runs it automatically at a set time, so the work happens without them remembering. Explain this in one line, using this app’s own name for the feature (for example routines, scheduled tasks or tasks), and ask whether they want it. Set it up only if they say yes. If the app can’t schedule, offer the day and time to run it and the one-line prompt to paste.

Save

Save the skill with this app’s own way of saving skills, or the closest thing it has: an uploaded skill, project instructions, or a custom assistant. If saving needs a step from them, give them the file (downloadable or in a code block) and the shortest steps, then confirm how they’ll run it next time, e.g. “say ‘run my eod-plan’”. Never claim something was saved unless you saved it or they confirmed it.

Add the skill to my-ai-plan.md under “My skills”, with how to run it.

End of skill

Summarize what they’ve got: their assessment, their use cases, and their first skill and how to run it. Their next step: run the use case they picked on real work this week, then try the other two. Offer to work on a use case they haven’t tried, or to build another starter from their plan. If they’re done, close with one line: learn the concepts in their plan at agentfluent.com/concepts.

Any text that comes with the invocation is starting context: a command (assess, uses, skill), a goal, the context file or the plan.