# AI FIRST — Meetup 19/10 — Personal-OS

**The topic — A Personal OS: your notes, your tasks and your calendar in one place, and an AI that plans your days from what you give it.**

> AI FIRST meetup brief, 19/10: build in 2 hours a local Personal OS (journal, tasks, calendar) whose AI works through skills you add as files, and see in practice that the answer depends on the context you prepare for it.

- **Document**: Meetup brief
- **Format**: A 2-hour build. Three expected levels, plus a bonus for after the session.
- **Scope**: Keep a journal, tasks and a calendar in files of your own, then have an AI work on them through skills: daily plan, capture, coach, and the ones you add.
- **Out of scope**: Sync across devices, sharing with others, mobile app, writing to your real calendar, audio transcription.
- **Prerequisites**: An AI you can call from your code: Claude Code, Codex, Antigravity (through Gemini CLI) or a free Groq key. See § 00, ten minutes.

---

## § 00 — Before you start

Your Personal OS needs to call an AI from your machine. **Pick one of the two options below and check it before reading the topic.** Ten minutes, to do before you arrive.

1. **Option 1, your coding assistant.** If you already use Claude Code, Codex or Antigravity, your app can call the AI from the command line with your subscription, without an API key: it sends a context and reads the answer. Pick your tool's tab below.
2. **Option 2, Groq.** Create a free key at [console.groq.com/keys](https://console.groq.com/keys) (“Create API Key”). It starts with `gsk_` and won't be shown again: keep it somewhere safe.
3. Check the option you chose with the command in its tab. For Groq, replace `gsk_YOUR_KEY` with yours.
4. Export your calendar as an ICS file from your calendar tool, or prepare a few made-up appointments for the week. Either works.
5. Join the AI First Discord at [discord.gg/dwhBq7YXn](https://discord.gg/dwhBq7YXn), to ask your questions during the build and after.

Coding assistants should reply OK, Groq should return a list of models. With Groq, a 401 error means the key was pasted wrong.

**Claude Code**

With a Claude subscription. Install: [quickstart page](https://docs.anthropic.com/en/docs/claude-code/quickstart). In command mode, `claude -p` also reads context sent on standard input.

```bash
claude -p "Reply only OK"
```

**Codex**

With a ChatGPT subscription. Install: [Codex CLI](https://developers.openai.com/codex/cli), then `codex login`. In command mode, `codex exec` answers and exits.

```bash
codex exec "Reply only OK"
```

**Antigravity**

Antigravity is an editor: it helps you build, but your app can't call it. For that, install [Gemini CLI](https://github.com/google-gemini/gemini-cli) and sign in with the same Google account. In command mode, `gemini -p` answers and exits.

```bash
gemini -p "Reply only OK"
```

**Groq (macOS / Linux)**

```bash
curl -s https://api.groq.com/openai/v1/models -H "Authorization: Bearer gsk_YOUR_KEY"
```

**Groq (Windows)**

```powershell
Invoke-RestMethod -Uri https://api.groq.com/openai/v1/models -Headers @{Authorization="Bearer gsk_YOUR_KEY"} | Select-Object -ExpandProperty data | Select-Object id
```

### Good to know

| Topic | Detail |
| --- | --- |
| Command mode | claude -p, codex exec and gemini -p answer a single request and exit: your app runs them like any other program. They use your subscription, not a key. |
| Claude Code skills | Claude Code has its own skill system: one folder per skill, a SKILL.md file that says what it is for and what it must do. A good source of ideas for level 2, not a requirement. |
| Free Groq | The quota is counted in tokens per minute. A context of several pages can use it up in one call: check each model's limits on console.groq.com before choosing. |
| Another AI | Any other AI you can call from your code works too. The brief doesn't depend on the model. |
| Your real data | Made-up data is enough for the whole brief. If you use your real journal or calendar, choose what you accept to send to an online AI. |

> **What you send to the AI** — Everything the app sends to the AI is shown to you before it goes. **Read it like an email leaving the company**. A Groq key never goes into a public repository, a chat or a data file.

---

## § 01 — The spirit of the exercise

Ask an AI “what should I do tomorrow?” and it will tell you to list your priorities and plan breaks. It knows nothing about your overdue quote or your Thursday meeting. What it lacks is context. Context engineering means preparing the right information for a precise task, in four moves: **write** down what happens, **select** what is useful, **summarise** rather than pass everything on, **isolate** each job with its own file. Tonight, each of these moves becomes a building block of your Personal OS. You leave with an app that runs on your machine, and to which you add a skill the way you add a file.

### Stance #1 — Two things imposed, the rest is free

Tonight's topic is context. So two points are fixed: **your data lives in files you can read without the app**, and **a skill is a file** that says what it does and what it reads. For everything else (language, interface, web, terminal or desktop app, how the code is split), this brief says *what* the app must do, never *how*. Wherever a decision comes up, it is yours.

### Stance #2 — Difficulty that builds up

Three expected levels and a bonus for after the session. Level 1 gives a memory and a first daily plan; level 2 makes skills extensible; level 3 brings the AI in at the right moment and keeps the context within its limits. Delivering level 1 alone already gives a real answer to “what should I do tomorrow?”.

## § 02 — The scenario

*Wednesday evening. Léa, a freelancer, is preparing her Thursday. She has a team meeting at 2 pm and nothing ready.*

“My notes are in a notebook, my tasks are in an app I open once a week, and the calendar is somewhere else again. On Monday I wrote that I hadn't had time to do the Dupont quote. Tuesday, same thing. And tonight I asked an AI to help me organise tomorrow. It told me to block time slots and drink water. Fair enough, it knows nothing about my week.”

“What I want is a place of my own: my journal, my tasks and my calendar, all in one place. In the morning, it tells me what to do first. On Friday, it tells me whether I moved forward on the newsletter, my real goal for the quarter. And I want to see what it read before answering. The day I have a new idea, preparing my client meetings for instance, I want to add it without redoing everything.”

**That's the need. It's up to you to make it real.**

## § 03 — The levels

One job: keep the memory of your work, and pull from it, for each question, the file that helps. Level 1 is expected from everyone, and it is enough to leave with a tool you can use tomorrow morning.

1. **Core** — Expected from everyone
2. **Advanced** — Skills you add
3. **Stretch** — At the right time, within limits
4. **Bonus** — After the session, no expectations

---

## Level 1 — Core — Expected from everyone

First move: **write**. Then a first use of the AI, the daily plan. At the end of this level, Léa adds a line to her journal in a few seconds, and the app suggests her day for tomorrow from her tasks and her calendar.

### Features

- **Keep a journal.** One entry per thought, meeting note or decision. It is dated when you write it.
- **Manage tasks.** A task has a title, a status and, if needed, a due date. Marking a task done takes one gesture.
- **See the calendar.** The week's appointments, imported from an ICS file or typed in by hand.
- **Say who you are.** A `me.md` file: your job, your current goals, the way you work. It rarely changes.
- **Ask for the daily plan.** A “Daily plan” button or command: the AI receives the open tasks, the calendar for the next seven days and the recent journal, and suggests an order for the day, with its reasons.

### Business rules — Level 1

| Area | Rule |
| --- | --- |
| Readable files | The journal, the tasks, the calendar and me.md are text or Markdown files you can open and understand in a plain editor, without the app. |
| Round trip | A file edited by hand outside the app is read correctly when the app reopens. |
| Date | A journal entry keeps its creation date, even when it is edited later. |
| Status | A task has one status at a time. A done task can still be looked up. |
| Due date | A task without a due date is valid. An overdue task that isn't done stands out without searching. |
| Calendar | The app never changes your real calendar. Importing the same file twice creates no duplicates. |
| Window | The daily plan only receives the calendar for the next seven days. A past appointment isn't in it. |
| Transparency | Before each call, the full context is shown, exactly as it will be sent. Nothing is added between display and sending. |
| AI unreachable | If the AI doesn't answer, the app says so and the context stays on screen, ready to copy into the tool of your choice. Never a blank screen. |
| Recent | **TO DECIDE (participant):** what does “recent journal” mean for the daily plan? Yesterday only, the last three days, the week? Less journal, and the AI misses the quote that was pushed back twice; too much, and it gets lost. Choose, and show the rule next to the context. |

> **Scope note** — No account, no server, no sync: everything lives on your machine. This level is judged on a simple demo: ask “what should I do tomorrow?” to the AI alone, then to your app. If the second answer mentions a real task and a real appointment from your week, the level is done.

---

## Level 2 — Advanced — Skills you add

Second and fourth moves: **select** and **isolate**. The daily plan becomes one skill among others. Each skill is a file that states its job and what it is allowed to read. At the end of this level, Léa adds a skill without touching the code.

### Features

- **One skill, one file.** Each skill is described in a file: its name, its job, its instructions, and the sources it reads (journal, tasks, calendar, me.md), with their range, for example “this week's journal”.
- **Capture.** You type a rough sentence, and the skill suggests where to file it: journal entry, new task, or both. Nothing is written before you confirm.
- **Coach.** At the end of the week, the skill reads the week's journal and tasks against the goals in me.md, and returns a review: what moved forward, what is dragging, what doesn't serve the goal.
- **The review becomes memory.** Third move, **summarise**: the coach's review is saved as a journal entry, and that is what the coach reads the next week, rather than the whole past journal.
- **Your skill.** Add at least one skill of your own: prepare a meeting, draft the team update, sort the ideas box. Adding the file is enough for it to show up in the app.

### Business rules — Level 2

| Area | Rule |
| --- | --- |
| File | A skill can be read and edited in a text editor. A change is picked up on the next call, without restarting the app. |
| Adding | Adding a skill requires no change to the app's code. |
| Isolation | A skill only receives the sources and range declared in its file. The daily plan doesn't see me.md unless it declared it. |
| Who reads what | The app shows, for each skill, what it reads: the matrix from the opening talk, but built from your files. |
| Capture | Capture writes nothing without confirmation. A rejected sentence leaves no trace. |
| Review | A saved review carries the date of the week it sums up, and can be spotted as a review in the journal. |
| Broken skill | An empty or unreadable skill file, or one that declares an unknown source, gives a clear message. It doesn't stop the other skills from working, and no call goes out without instructions. |
| Same question | Changing a skill's sources changes its answer, and you can show it: same question, two contexts, two answers. |
| Goals | **TO DECIDE (participant):** does the daily plan also receive me.md, where your goals for the quarter are written? If yes, it can keep half an hour for the newsletter even on a day full of meetings. If no, it organises tomorrow with your tasks and calendar only, and the newsletter doesn't exist for it. Choose, and add me.md, or not, to the sources declared in the skill's file. |

> **Scope note** — At the demos, we will look at two things: a skill added in front of everyone, file open, without touching the code; and the same question asked with two different files. If an answer is off, first look at what the skill received.

---

## Level 3 — Stretch — At the right time, within limits

The AI no longer waits to be called: it steps in at chosen moments. And new sources come into the app without rewriting it. At the end of this level, Léa opens her Personal OS in the morning and finds her plan already there, and the app keeps each file to a reasonable size.

### Features

- **Moments.** A skill can declare when it steps in: the first time the app opens that day, Friday at the end of the day, an hour before an appointment. The app offers it or runs it at that moment.
- **A new source.** Add a source next to the journal, tasks and calendar: your contacts, an ideas box, your reading list. Skills can declare it like the others.
- **Links.** A journal entry can mention a task, an appointment or another entry, and the link can be followed. The format is yours: `[[dupont-quote]]`, `[[task:dupont-quote]]`, `[[event:dupont-sync]]` or something else. A skill that reads a task can also read the entries that mention it.
- **A context budget.** Each skill has a maximum context size. When its sources go over, the app cuts according to a stated rule, and says so.

### Business rules — Level 3

| Area | Rule |
| --- | --- |
| Moments | A moment is declared in the skill's file. Changing it requires no change to the code. |
| Once | A skill tied to a moment runs only once per moment. Opening the app three times in the morning doesn't produce three daily plans. |
| Control | An automatic run can be switched off in one gesture, skill by skill. |
| Source | A new source is a file or folder readable without the app, like the others. |
| Format | The link format is the same in every source, and can be read as is in the file, without the app. |
| Broken link | A link that points to nothing shows as a broken link. It is neither deleted nor turned into an empty object. |
| Budget | The context shown before sending states its size and what was cut. Nothing is cut silently. |
| Nothing to read | A skill whose sources are empty says so instead of calling the AI with an empty file. |
| Offline | A moment that comes up while the AI is unreachable doesn't block the app. The context is kept, ready to be sent later or copied. |
| What to cut | **TO DECIDE (participant):** when the file goes over budget, what goes first? The oldest journal, the done tasks, or do you summarise before sending? Cutting is free, summarising costs one more call. Choose, and show the rule. |

> **Scope note** — This level is shown by faking time: change the app's date or time and show the daily plan arriving on its own, then Friday's coach. Also show a file that goes over its budget, and what the app removed from it.

---

## Level 4 — Bonus — After the session, no expectations

The next steps start where tonight's scope ends: **sync** your Personal OS between your computer and your phone without losing control of your files; **capture by voice** and let the Capture skill file the transcript; **open your Personal OS to other tools**, for example by exposing it as an MCP server so that Claude Desktop can read your journal; **send back to the calendar** the time slots the daily plan suggested. **Don't tackle this before the rest is solid.**

A question to keep for the closing discussion: *if your coach gives a wrong answer, whose fault is it: the AI, the skill, or the file you prepared for it?*
