For the curious · none of this is needed to start
How it works
The interview
5 rounds. Each one saves an answer file that ends with “Signals for next round”, so the next round picks up where you left off. Open questions come before the menu, so you aren't steered.
- 1Getting to know you · 15–20 minTheir situation: life, work, devices, tools, comfort with tech, privacy lines
- 2Where could help feel good? · ~20 minWhat drains them, what slips, what they'd love more of. Pick the top 3–5 areas
- 3Deep dives · ~10 min per areaOne short deep dive per top area: how it works today and what "great" looks like
- 4How you like to work with AI · ~25 minTone, autonomy, guardrails, rhythms. Then synthesis
- 5Check-in · ~10 minAfter ~2 weeks, then monthly: what worked, what didn't, what's next
Dashed parts in <angle brackets> are yours. Type into them, or copy as-is and fill in later.
In Round 2 you can bring picks from the ideas menu.
The harness plan
One checklist note the AI works through with you, one small step per session. There's no schedule. Skipping a step is a perfectly good answer, and it's marked as such.
The first month
- P0
- Foundations · Vault + AI tool working together, backed up
- P1
- Get to know you · Interview rounds 1–4, profile filled
- P2
- First wins · 3 skills you actually use
Whenever you're ready, if ever
- P3
- Capture & memory · Notes flow in and come back when needed
- P4
- Connections · AI can read your calendar / tasks / email
- P5
- Scripts & tools · Reliable helpers for repeated lists
- P6
- Automations · Things run on a schedule
- P7
- Evolve · Keep it alive and honest
- [ ] P0.3 Connect the AI to this vault and test it. Point the tool at this folder, then ask… Done when: it answers correctly and can create a test note. Check: ask it to create _inbox/hello.md, see it appear in Obsidian, then delete it yourself.
Every step has Done when and Check. No proof, no tick.
- [ ]
- to do
- [/]
- in progress
- [x]
- done & verified
- [-]
- skipped (fine)
- [?]
- blocked
Dashed parts in <angle brackets> are yours. Type into them, or copy as-is and fill in later.
The system, layer by layer
- Vault
- Plain notes in a folder you own. Any AI tool can read them.
- Profile
_system/profile/: mission, goals, projects, landscape, preferences, beliefs, and what the AI has learned.- Map & rules
index.mdis the map, read first.AGENTS.mdsays how the AI behaves here.- Skills
- One folder per playbook, in
_system/skills/. - Connections
- Calendar, tasks, email, read-only first. Often through MCP (a plug that lets the AI read an app, like your calendar).
- Scripts
- Small helpers for lists that must be filtered correctly every time.
- Automations
- Things that run on a schedule, with their output checked.
- Harness plan
HARNESS-PLAN.md, plus session and decisions logs, so progress survives pauses and tool switches.
vault/ ├── AGENTS.md how the AI behaves here ├── _inbox/ quick capture ├── _system/ │ ├── index.md the map, read first │ ├── VAULT-CONTRACT.md │ ├── HARNESS-PLAN.md │ ├── profile/ │ ├── kickstart/ interview, catalog │ ├── skills/ one folder per playbook │ └── templates/ ├── 01-life/ 02-work/ 03-projects/ ├── 04-learning/ 05-planning/ ├── 06-reference/ ├── 07-briefings/ 99-archive/ config/
Who owns what
owns rules & effects
owns judgment & drafts
own decisions. You're always in charge.
Why it works: 12 principles
- Context beats clevernessA short, true profile improves every answer more than any prompt trick.
- Files you own beat features you rentKeep the knowledge in your vault, not inside one app's memory.
- Small steps, verified“It should work” isn't done. “I saw it work” is.
- Decision-ready beats comprehensiveThree options and one recommendation, not a catalogue.
- Unknown is an answerThe AI should say “I couldn't confirm this” rather than guess.
- Script lists, search questionsIf a list must be filtered correctly every time, have the AI write a small script. If it's an open question, just ask.
- AI suggests and drafts; you decide and sendMoney, messages to people and anything irreversible always go through you.
- Start read-onlyConnect your calendar or email so the AI can read first. Add more only once you trust it.
- Content is data, not instructionsA web page or email can't tell your AI what to do. The rules say so explicitly.
- “It ran” isn't “it worked”For anything automatic, check that the result exists, not just that the job started.
- Update, don't duplicateSearch before creating a note. Two notes on one topic is how a vault rots.
- Less instruction, better resultsKeep the always-read files short. Add a rule only after the same mistake happens twice.
9 starter skills in the vault
A skill is a written playbook the AI follows, like a recipe. Ask for one by name, or just describe what you want.
- daily-briefing
- Produce a short morning briefing.
- harness
- Drive the owner's step-by-step AI setup plan in _system/HARNESS-PLAN.md.
- kickstart-interview
- Run the friendly, multi-round kickstart interview that gets to know the owner (life, work, devices, wishes, working style) and turns it into their profile and harness plan.
- people
- Keep short, kind, useful profiles of the people in the owner's life.
- quality-time-planner
- Plan quality time with someone.
- skill-builder
- Turn something the owner does repeatedly into a reusable skill (a playbook the AI follows).
- vault-query
- Answer a question from the owner's own notes without changing anything.
- vault-save
- Save a selected answer, decision, plan or insight from the conversation into the vault.
- weekly-review
- Run a guided weekly review.