Capstone: your always-on assistant
Pick the one real job
Everything you have learned now collapses into a single deliverable: one assistant that does one real, recurring job for you, end to end. First, choose the job and run it manually once so you know what good looks like.
- 1 Pick one genuinely useful recurring job from your life: a morning briefing to Telegram, inbox triage with approval, a weekly research digest, a project status report.
- 2 Run it once by hand, in conversation with Hermes, until the output is something you would actually want.
- 3 Write down what 'done well' looks like for this job — the bar your automated version must clear.
- 4 Note which pieces it needs: memory, a skill, a channel, a schedule, an approval gate.
You have one real job, run manually once to a standard you are happy with, and a clear bar for the automated version.
How the four pieces compose into an assistant
A capstone is not a new feature — it is the moment the separate pieces become one thing that works while you are not watching.
A real always-on assistant is the composition of everything in this course: memory so it knows you, a skill so the procedure is repeatable, a channel so it reaches you where you are, a schedule so it runs without prompting, and a delegation level chosen per step so it is safe. The craft is matching each part to the right rung of the ladder — the gathering can run at level 5 (act and report), while any step that sends, spends, or deletes drops to a level-3 gate. Done right, the whole is genuinely more than the parts: not a chatbot you visit, but an assistant that compounds — remembering more, doing more, and staying inside the lines you drew.
- ?For your chosen job, which step is safe at level 5 and which must be gated at level 3 — and why?
- ?Which lesson's piece (memory, skill, channel, schedule) is doing the most work in your assistant?
- ?If you handed this assistant to a colleague, what would they need to understand before trusting it?
- ?What would tell you, a month from now, that this assistant is actually saving you time rather than quietly costing it?
Wire it end to end
Build the real thing: one assistant, composed from your memory + a skill + a channel + a schedule, with each step at the right delegation level.
Assemble a working always-on assistant for your chosen job, annotate the delegation level of each part, and confirm it runs unattended and reports back.
- 1 Make sure the facts it needs are in memory (lesson 2) and the procedure is a skill (lesson 5).
- 2 Connect it to the channel you actually check (lesson 3).
- 3 Schedule it as a standing errand (lesson 6), scoped to only the tools it needs.
- 4 Put a level-3 approval gate (lesson 4) on any step that sends, spends, or deletes — and confirm the gate fires.
- 5 Run it (scheduled or test) end to end and check the output clears the bar you set in the USE phase.
- 6 Write a one-paragraph annotation: which delegation level each part runs at, and why.
A working assistant doing one real recurring job, its report from an unattended run, and a short annotation of the delegation level of each part (memory / skill / channel / schedule / gated steps).
✎You have built something that works for you while you are not watching. What is the next job you would trust it with — and what would have to be true about its track record before you did?