What it will not do

The limits that are built in, the rules I hold mine to, and how to tell a broken connection from a thin brain.

Most of these docs tell you what your assistant can do. This page is the other half, and I would rather you read it in week one than find it out in week six. Some of what follows is built into the software. Some of it is rules I hold my own assistant to, and you will have to give yours the same rules yourself. I will keep the two apart, because confusing them is how people get burned.

It will be confidently wrong sometimes#

The brain underneath is a language model. When it does not know something it does not go quiet; it fills the gap with the most plausible thing, in the same even tone it uses when it is right. The word for this is hallucinate, and it is not lying. It is a guess wearing the clothes of a fact.

So: a number, a name, a date, a quoted line, a "studies show". Anything that would embarrass you if it were wrong. Check it before you act on it, and ask the assistant where it came from. If the source is a page in your brain, read the page. If it cannot name one, treat the claim as a guess. The /verify recipe on starter recipes turns this into a habit rather than a resolution.

The good news is that the fix is mostly not about the assistant. A thin answer is usually a thin brain. If it keeps guessing at what matters to you this quarter, that is because nothing in the brain says so, and grow the brain is the page for that.

What is built in#

These are the software's own limits. You do not have to ask for them, and you cannot talk it out of them in conversation.

• It drafts. It does not send. Email, messages, anything outbound: it writes the draft and stops. You send.
• /inbox asks once before it touches a label. It sorts unread mail into buckets and shows you the lot; nothing is marked read or archived until you give one bulk confirmation.
• /capture shows what it will write before writing it. The end-of-session sweep proposes each change to the brain and waits.
• /explain and /followup change nothing. Both read the mineral and report. Ask /explain what is connected, where your data lives, whether backups are on, and it answers from what is actually there.
• Schedules ship switched off. Nothing runs unattended until you flip a switch on the Skills page and save, and nothing can be scheduled at all until the mineral has its own Claude sign-in.
• It reaches what you connected and nothing else. A skill can only read a service you have added on the Connections page (what you can connect is the catalogue). It does not browse the web on its own. Research means your brain, your connected services, what you paste, and a search service like Exa if you connect one.

That is the list. Notice what is not on it.

Four house rules I paste into the brain#

Nothing in the engine stops the assistant from editing a page without asking, or from deleting a file, or from telling you it read something it only skimmed. Those are behaviours I want from mine, so I wrote them into layer 8 of my brain (the workflow layer, wiki/_layers/8-workflow.md), which the assistant reads at the start of every session. They are rules, not guarantees. A rule in the brain is followed the way a good new hire follows a rule: nearly always, and better the more plainly you wrote it.

Here are the four, in a form you can paste into the Terminal:

Add these four rules to layer 8 of my brain under a heading
"House rules", and show me the page before you save it.

1. Propose before writing. For anything that matters (a
   client page, a draft going out, a change to a page you
   already have), show me the wording or the before-and-after
   and wait for my yes. Do not write first and tell me after.
2. Check before acting. Before asking someone for
   information, check whether I already have it. Before
   recommending who to contact, check whether a plan in the
   brain already names someone. Before drafting to a person,
   read their page for recent contact.
3. Never delete. Move it aside with today's date and a
   reason, and ask. This goes for pages, files, and anything
   in a connected service.
4. Never claim to have read what you have not. Refer to a
   document, an article or a message as read only if it is
   actually in front of you. Present a line as a quote only
   if it is word for word.

Why these four. The first saves you the most: it turns every edit into a before-and-after you glance at, and a glance is enough. The second is about premises. An assistant that asks a supplier for a phone number you already have makes you look disorganised, and the check costs five seconds. The third is because a moved file comes back and a deleted one does not. The fourth is the honesty rule: an assistant that says "your 2023 pricing note sits close to this" when it never opened the note has just made a claim in your voice that you cannot stand behind.

Earn before you act#

There is a ladder and I climb it slowly. Read first: connect a service, let the assistant see it, ask it questions about what it sees. Then draft: let it write the reply, the report, the page, while you stay the one who sends or saves. Only then act: a schedule that writes labels, a skill that updates a page without a look from you.

Each rung up buys you time and costs you predictability. A phrase I picked up from Nate Herk and have not been able to improve on: autonomy equals risk equals cost (where these ideas come from has the rest of what I owe him). So I scope to the lowest rung that solves the problem. A daily report that always runs the same five steps does not need anything deciding its path; it needs five steps. And I scope the tools, not just the words. Telling it "never send email" in a prompt is a wish. Not connecting the thing that can send is a wall. The engine already builds the wall for sending; for everything else, you decide what to connect and when.

The rule of thumb I use: a skill earns a schedule after it has run well by hand three times, and a connection earns a wider grant after I have read its output for a fortnight and not once winced.

When it looks wrong#

Most "it is broken" moments are one of four things, and they have four different fixes.

One more that is not a fault at all: it forgot yesterday's conversation. Sessions do not resume. Whatever mattered had to be captured into the brain, and /capture at the end of a session is how. How your assistant thinks explains the two kinds of memory, and once you have seen them the forgetting stops being mysterious.