A useful answer, on purpose.
Why it matters: everything downstream is built from the quality of a single chat exchange. This is where trust in the tool begins.
The chat window looks like a search box, so questions come naturally. The tool responds best to something closer to a brief: who the work is for, what finished looks like, and what to leave aside.
The shift: from "what is X?" toward "here's what I need/want to know, here's what would help and how I want the output to look."
- Offer a brief, the way you'd brief a colleague or assistant.Audience, goal, what "done" looks like, what to leave out, one example of the tone you like.
- Invite one question for it to ask you before it starts.A single clarifying question catches most misreads, and it costs one sentence.
- Give feedback the way an editor would."Keep the second paragraph's tone; shorten the third." Specific notes land; general ones drift from your target.
- Keep the briefs that worked.A small file of them. It becomes the seed of stage 2.
Audience: [who will read this]
Goal: [the one thing it should accomplish]
What done looks like: [format, length, tone]
Leave aside: [what isn't needed, what to avoid]
Before you begin, ask me one question if anything is unclear.
If a choice would change the direction, pause and ask before continuing.
An invitation to you: rewrite one question you'd normally ask AI as a five-line brief, and notice the difference in what comes back.
What my first 1,200 conversations taught me since 2022
Looking back through my chat history export, there's a visible month where my messages became longer than the replies. That's where the results turned useful. It took about a year and a half to get there, and those same conversations later became the first material for the memory at stage 3.
- Also built
- AI training at work: sixty seconds each week with my team. One real prompt or lesson on screen, what came back, and the one adjustment to a prompt or tool that made the biggest difference.
- Feature
- Plain AI chat. Nothing more yet, and that was enough at the time.
Say it once, and let it stick.
Why it matters: the effort spent re-introducing yourself each chat can be saved and reinvested in the work itself.
Profile preferences and Projects let context settle into place. A short, always-on note about how you like to work, then one project for each area of life, each holding its own instructions and reference files. Every major assistant has a slot for this: Claude calls it preferences, and each Project carries its own instructions; ChatGPT calls it custom instructions, with project instructions too; Gemini calls it saved info, and Gems carry instructions; Copilot calls it personalization.
The shift: from repeating context toward context that's simply present and available.
- Write a five-line core.Who you are, how you like answers, what to check with you before doing, what to assume, what to do when unsure.
- Create one project for one area, like finances, health, personal projects, and so on.Instructions plus two or three reference files. Live with it for a week before adding another.
- Add a boundary.The most helpful instruction is often about where a piece of context applies, and where it can rest. Always give AI a means to come back and ask you if it cannot decide something.
- Keep work and life in separate projects.If you must use a personal AI account for work, work context can and should live in a work project. Everything else assumes personal.
About me: [two lines; a job title is optional]
How I like answers: direct; the recommendation first, then steps, then what to watch for.
Please check with me before: sending, spending, or deleting anything.
Assume: personal life, unless I say it's work.
When unsure: say so, and ask.Already have instructions? Try this
What else have you noticed about how we work together that would make these instructions better?
Help me upgrade them, keeping the same five-line shape:
[paste your current instructions]
An invitation to you: write your five lines tonight and place them in your profile. Notice tomorrow how much less you explain. Already have instructions? Use the second block above.
A one-page core, and a line about where it applies
My profile describes how I like answers and includes one sentence: "assume personal unless I say it's work." Before that line, office context sometimes found its way into a recipe question. After it, each area stayed in its own room.
- Also built
- Nine projects, including finances, my permaculture food forest/garden, an active website with newsletter, a garden channel, and my own AI second brain and "Jarvis" setup. Each holds its own instructions.
- Feature
- Profile preferences, and Projects with instructions and knowledge files.
Files it can read and write.
Why it matters: a memory you can both see and edit turns a helpful tool into a partner that carries the load and the conversation.
Built-in memory holds facts. A folder of plain text files can hold a whole life, and the AI can help maintain it: reading what's there, filing what's new, keeping current state current on a scheduled routine.
The shift: from memory as a setting toward memory as a shared folder.
- Make the folder.One router page, a journal folder, three topic files. Plain markdown files, available from wherever the AI can reach, from anywhere.
- Keep an inbox file.Anything can land there. Once a day, ask the AI to file it where it belongs. Drop things in there as needed.
- Hold one gentle rule.Topic files describe what is true now. The journal holds the dates. Keeping them separate keeps both readable.
- Bring your past with you.Export your old chats from all your AI services/systems. Let the AI write the history, one file per day.
memory/
CLAUDE.md ← who I am, where things live
journal/2026-09-05.md
personal/finance.md ← current state
personal/home.md
INBOX.md ← anything, sorted later
An invitation to you: create the folder with just the router page and today's journal entry. Ask the AI to read both back to you. That's the whole first step.
My memory vault: 894 days of history, one file per day
I use Markdown files in Google Drive. A router page says where everything lives. To create my AI "second brain" I had the AI read 3,538 old conversations and write the history itself, a file per day. Personal health stats arrive four times daily; finance account balances arrive through a read-only connection. When the morning memory page is written, both are already there.
- Kept tidy by
- A small daily check that notices when a topic file grows past 40 KB or a link stops resolving, and leaves a note for me rather than changing anything. Notes flagged for my review on a schedule.
- Feature
- Google Drive connector, plus a daily task that files yesterday and refreshes topic files.
Let it work while you sleep.
Why it matters: A morning "brief" that's already written when you get to your PC, email or mobile device first thing in the morning is the difference between a tool you visit and a system that supports you.
Connectors reach your calendar, email, or home network. Skills are how-to files, kind of like saved prompts, used when a task matches or when you call them by name. Scheduled tasks put a prompt on a clock and deliver the result to you.
The shift: from asking each morning toward reading what's already been packaged and delivered.
- Connect one source.Calendar and/or email is a reasonable first choice. The fewest and simplest connections that serve the function.
- Write one skill.A how-to for something you do regularly. Its description decides when it's used, so give the description the most care. Work with your AI to build the skill – in a chat session complete the goals you are hoping to replicate, then say "let's turn everything we did in this session into a skill I can call on in a scheduled task" – let the AI give instructions, or do the work, depending on how you've configured your system permissions.
- Schedule one prompt. (Scheduled task, routine, etc)Fixed wording, fixed time - writes to a file, artifact, email or web domain you'll read. This could be daily research, email summaries from the past day, a home network report or much more.
- Then schedule a companion.A second task whose only role is to confirm the first one ran. A quiet job also pauses quietly, and this one notices. I call them "sentinels, because they're keeping an eye out so you don't have to.
Every weekday at 8:00 am:
1. Read today's calendar; one line per appointment/meeting on why it matters.
2. Read INBOX.md; gather anything marked TODO.
3. Write a 200-word brief: decisions first, then the day's shape.
4. Save to memory/briefs/YYYY-MM-DD.md and email it to me.
If any source is unreachable, name it at the top of the brief.
An invitation to you: schedule the brief for one week. Read it each morning. Adjust the wording on Friday, and let it run again. After a couple of good weeks, turn it into a visual HTML artifact that updates on each new brief, designed to your liking.
Fourteen scheduled tasks, and one that looks after the others
Finance refresh at 7:00. Network watch at 7:40. Research sweep at 7:45. My morning brief pulls calendar, mail, and the overnight results together and lands at 8:10. The second-brain sync files the day at 8:30. Finance reviews run weekly, monthly, quarterly, and yearly. Once an hour a sentinel checks that every one of them left a heartbeat, and says so plainly when one didn't. Each task writes back to the vault. The brief collects it all and hands it to me in text, a visual page, and audio.
One of them stopped being a scheduled task and became something that acts on its own: an agent that watches the network and fixes the small faults without me. It also spent two days deaf while every health signal I had stayed green. I wrote down how it works and every way it quietly lied about being healthy — the build guide, and the failure catalogue that came with it.
- What I learned
- I basically have an assistant who hands me my day in summary. Before the sentinels were watching, one task reported itself as "degraded" for six days and everything read it as healthy. Now a missing or partial heartbeat gets a second look, and the sentinel says so plainly.
- Feature
- Scheduled tasks, a bunch of connectors, and 43 skills I've written or added. My home-network skill is packaged as a private plugin so it's tracked like code.
Things that exist on their own.
Why it matters: a page or document generated inside a chat is easy to lose. A repository, a domain, a website or standalone HTML, a video can be found again, shared, and built on.
This is the stage where people ask "AI did this?" And the habits are simple: every project begins with a repository, every site has an address, and the AI shows you the live page after each deploy.
The shift: from chat outputs, now you move toward things with a home.
- Choose one domain.Ask the AI to walk you through a repository (like GitHub) and a hosted site (like with Cloudflare), click by click. Doing it once by hand makes every later deploy familiar, and you can trust the AI more as you learn and do with it.
- Ask for the spec before the code.One input every prompt, every step, what the output looks like. "Show me the plan and spec first. I'll say go." Get your AI in the habit of writing specs and plans for you this way, every time.
- Ask it to show you a mock up and demo before it pushes to the live page. Then have it show you the live page.After a deploy, have it fetch the URL and share what it sees. A deploy that reports success can still change nothing. The live page is the only proof.
- Brief your design the same way.Subject, mood, light, camera, what to leave out. Four variations, then refine the one that speaks. I always ask AI for numerous variations of types of designs, then when one sticks, variations of my favorite design, too.
- Give every site a place on a status page.It writes back to the vault, and the loop closes.
Before writing any code, let's write SPEC.md together:
- The one input the user gives
- Every step, with min / ideal / max
- What can stretch, and what can't
- What the output looks like
Then build the smallest thing that satisfies it. Show me the spec first.
An invitation to you: pick the smallest thing you'd be glad to have a URL for. Spec it in ten lines. Ship it this month. For me, it was my first lab site, hunterlabs.org (the .com was taken).
Repositories, domains, and tools for home
It started with a simple 8-bit landing page for a few standalone HTML pages (my daily brief, a text-message search tool, a health dashboard tied to my Apple Health data). It grew into a lot more.
A daily research task that became a live site and newsletter on how organizations govern AI, The Governance Gazette. A lab site with home tools: a personal CFO that advises and never moves money, side-business research, and a status page that watches every site, domain, token, and API.
An app built on my favorite pizza recipe book that plans the whole meal backwards from dinner time. And most recently, a Thai language coach built on everyday sentences.
- The loop
- The status page writes to the vault. The vault feeds the brief. The brief surfaces any issues, and helps plan the next thing worth building. That loop wasn't designed; it appeared once the pieces could read each other.
- Feature
- Claude Code in the terminal, GitHub for the code, Cloudflare for the domains, Pages, and Workers, and image and video tools given the same kind of written brief as anything else.
- Find them
- The Governance Gazette · Three Forests · hunterlabs.org