Layer 1 of 7 · Makes AI smarter about you
Identity
Who are you, and what rules are always enforced?
The foundation. This layer defines who you are: your role, communication style, values, and the non-negotiable rules your AI must always follow. Your tool reads it first, before anything you type.
Good identity files state hard rules as always/never statements: "Never flatter me." "Always tell me what I'm not seeing."
- Where it lives
- System instructions (project instructions in Claude, custom instructions in ChatGPT) plus a personal context portfolio.
- In practice
- Don't write it from scratch. Have the AI interview you, draft it, then edit and ship the 70% version.
Layer 2 of 7 · Makes AI smarter about you
Context
What do you know?
This layer represents what you know: a library of focused, dated text files the agent reads on demand, rather than text you paste into every prompt. The single biggest predictor of whether AI output is generic or genuinely useful.
The trap is one 40-page "everything" document that goes stale immediately. What works: 3 to 5 one-page files (stakeholders, strategy, operating principles). Dated, fresh, single-topic.
- Where it lives
- A personal context portfolio, plus reference files within skills and role-specific plugins.
- In practice
- Every time you catch yourself re-explaining something to an AI, that thing belongs in a context file.
Layer 3 of 7 · Makes AI smarter about you
Skills
How do you work?
Skills are SOPs: reusable instruction sets for workflows you do repeatedly: weekly updates, meeting prep, decision memos. The pattern every skill follows: when I say [trigger], do [process], using [sources], and produce output in [format].
Groups of related skills package together as plugins. Most knowledge workers have 20 to 30 skill-worthy patterns; start with 1 to 3, ship the rough version, and patch it as you use it.
- Where it lives
- Skill files in your AI tool; bundled into plugins when a role needs a set of them.
- In practice
- Like employees: v1 usually disappoints, but v5 can be like magic.
Layer 4 of 7 · Makes AI smarter about you
Memory
What does it remember between sessions?
Memory layers nuance onto Context and lets Skills improve over time. It is what makes the other layers stick across conversations instead of resetting every session.
Three options, each with limits: native harness memory, skills that write to files, and markdown files you maintain yourself. There is no magic pill; good memory is deliberate work, like taking notes instead of trusting your recollection.
- Where it lives
- Native tool memory, memory-writing skills, or structured markdown files (decision logs, relationship context).
- In practice
- Pen and paper versus trusting your recollection: this is hard work, and magic pills do not exist.
Layer 5 of 7 · Lets it act in the world
Connections
What can it reach?
Connections let the agent touch real systems: email, calendar, chat, project management, CRM. The mechanism can be MCP or direct API, but it's usually MCP, and increasingly plug-and-play.
The safety rule for this layer: start read-only, and be deliberate about who has access to what. Add write access only after observed trust, and even then prefer drafts-for-approval.
- Where it lives
- MCP connectors in your tool's directory; CLI tools or APIs for the rare deeper case.
- In practice
- Be careful about permissions early on, and who has access to what.
Layer 6 of 7 · Makes it trustworthy
Verification
How do you know it's right?
The biggest failure of an AIOS is not an error message; it's hallucinations and AI slop: confident, wrong output that ships before you notice. Verification is knowing what to check for each job; it is the prerequisite for Automations.
It can live inside prompts or skills, or as separate validation agents, depending on stakes. Define 3 to 5 quick checks per skill; verify high-stakes output closely and learn to trust the low-stakes.
- Where it lives
- Checks written into prompts and skills; standalone validation agents for high-stakes work.
- In practice
- The more you trust these systems, the more important this becomes, just like with humans.
Layer 7 of 7 · Runs without you
Automations
What runs when you're not watching?
Scheduled tasks, managed agents triggered by events, a morning brief delivered before you wake. Powerful, and optional. An AIOS can include this layer, but it is not essential.
As with all automations, and all new hires: do it manually first to work out the kinks, then automate if the ROI justifies it. Start with drafts for your review, and always keep logs.
- Where it lives
- Scheduled tasks and event-triggered agents inside your tool; always with logs of what ran and what it did.
- In practice
- First you do it manually to work out the kinks, then you automate if the ROI justifies it.
