Course resource
Personal AI OS Template
Your whole AI setup, written down. Not a new system to maintain — a one-page description of what you already have, so you can see the gaps and so someone else could pick it up.
The layers
INPUTS where work reaches you
JUDGEMENT what AI decides or drafts
STORAGE where the durable things live
OUTPUTS what leaves, and who checked it
Inputs
| Source | Volume | Filtered how | Automated? |
|---|---|---|---|
| Forms | |||
| Chat | |||
| Calendar | |||
| Other |
The question to ask: which of these do I check more than three times a day? Those are candidates for a digest instead.
Judgement
| Task | Tool | Model tier | Temperature | Human checks? |
|---|---|---|---|---|
| Classify incoming | cheap | 0–0.2 | sampled | |
| Draft replies | capable | 0.6 | every one | |
| Summarise documents | capable | 0.3 | spot-check | |
| Research | with search | 0.3 | citations verified | |
The rule: cheapest model that passes your bar, per task. Running a frontier model on a classification step is most people's biggest avoidable cost.
Storage — the part that matters most
These are the things you own. If you change tools tomorrow, these come with you.
| Asset | Where it lives | Updated when |
|---|---|---|
| Digital Twin | ||
| Prompt library | ||
| Voice profile | ||
| Knowledge files | ||
| Templates and SOPs | ||
| Automation runbooks |
If this table is empty, you do not have a system — you have a subscription. Everything here is portable; the tools are not.
Outputs
| What leaves | Who checks it | What happens if it is wrong |
|---|---|---|
| Customer emails | ||
| Published content | ||
| Internal documents | ||
| Data written to systems |
Anything reaching a customer needs a named human in the "who checks" column. Not "me" — you, specifically, by name, in the process.
The daily rhythm
| When | What | Minutes |
|---|---|---|
| Morning | Briefing | 5 |
| Morning | Plan the day | 5 |
| Through the day | Draft-first on anything written | — |
| Evening | Three-line close | 3 |
| Friday | Weekly review | 15 |
Total: under thirty minutes a week of overhead. If your system costs more than that in maintenance, it is too elaborate.
The gaps
Having filled the tables above, answer:
- Which input do I still check compulsively?
- Which recurring task has no AI step but obviously could?
- Which AI step has no human check but should?
- Which asset in Storage does not actually exist yet?
- What breaks completely if my main tool is down tomorrow?
- What would someone else need to run this if I were away for a month?
Pick one. Fix it this week. Then re-read this page next month.
The principles
Own the portable parts. Prompts, documents, processes. The model underneath will change several times; those should not have to.
Put the human where the risk is. Not everywhere, or you have automated nothing. Before anything that reaches a customer, moves money, or writes to a record.
Log the judgement. Every AI step should record what went in and what came out. Without it you cannot debug, because the failure may not reproduce.
Retire things. Automations accumulate and keep running long after they stop being useful. Review quarterly and turn things off.
Keep the manual fallback. For anything critical, know how to do it by hand. Tools go down, and skill atrophy is real.
Review
| Date | Added | Retired | Cost/month | Biggest remaining gap |
|---|---|---|---|---|
What this is not
It is not a productivity philosophy and it does not need an app. It is a one-page inventory you update every quarter.
The value is entirely in the gaps column. Everything else is just writing down what you already do.