Student founder. Building Chiefly, an AI chief of staff that runs on your own computer. I also help Carolinas gyms get found on Google. Support the build ๐
Most boutique gyms donโt have a marketing problem.
They have a โnobody nearby can find you this weekโ problem.
Google Business Profile half-done.
Reviews stuck at 40.
Instagram posts that donโt turn into trial calls.
Supporters: what should I build into Chiefly next?
Vote below. If you're on the tip jar, you can also request a feature directly, and I'll tell you straight if it fits.
@niclasvestlund nice, easiest version I've found: only trigger the full suite on pull_request with types: [ready_for_review] (plus a label to force it). drafts get the quick checks only, saves a ton of minutes
@swombat makes sense. for execute I've had better luck making it opt-in per command instead of per agent. agent asks, you see the exact command, approve once. slower but you never wake up to a surprise
Build log, day 3.
The first time you open Chiefly, it runs a 2-minute interview card to learn about you.
So it isn't starting from zero, and you aren't re-explaining yourself every time you ask for something.
@chevonphillip 45 visitors is too small to judge the product on. i'd dm 10 people who look like your buyer and just ask what would get them to pay. they'll tell you what to fix
@aslambeg84 honestly the fix is it should never have been able to drop the db at all. give it creds that can't delete anything, keep read tools and write tools separate, and make anything destructive wait for a human yes
@dark_coderz don't trust what the agent says it did, trust what the tool returned. have the booking tool hand back the airline's confirmation number and check for it in code. no number = failed task, doesn't matter what the model claims
Charlotte and Belmont gym owners, a 5-minute Google Business Profile check:
- Hours match what's on your door
- Classes listed as services
- Recent photos of the actual space
- A phone number someone answers
- A reply on every review
Free 15-min call: https://t.co/QH1ncxXoA8
This weekend a few friends get to try Chiefly before anyone else. Private, small, on purpose. I want them to break it so I can fix it. Local-first, no account, and it asks before it acts.
No CS degree. No credentials. Just a stress-test log full of my own app's failures, and I'm fixing them one at a time.
Tomorrow I freeze the eval run for a hackathon entry. Win or lose, I'll post the real numbers.
https://t.co/xtfbpL3fN4
Chiefly can now read photos and scanned PDFs, right on your computer.
Nothing gets uploaded just to read a receipt.
Waitlist is open at https://t.co/xtfbpL3fN4
New rule for Chiefly: every approval card asks once and runs once.
If you didn't say yes, it doesn't happen. If you said yes, it happens one time, not twice.
Sounds basic. Took a full stress test to get right.
Last week my AI chief of staff invented a client named Tony Blue and said strawberry has 2 r's.
So I spent the weekend trying to break it on purpose.
Monday: 6 merges, 1140 of 1145 tests passing.
Building in public at https://t.co/xtfbpL3fN4
Build log, day 2.
You can now send Chiefly several messages back to back. Each one gets its own answer, in order. No waiting for it to finish before you type the next thing.
Changed your mind? There's a Stop button.
Small thing. Makes it feel a lot less like a form.
Build log, day 1.
Jarvis was the codename. Chiefly is the product.
Today: Chiefly reads screenshots, receipts and scanned PDFs on your own computer and saves the text into your notes. A test receipt read "TOTAL $11.47". Correct.
Local by default. Cloud only if you opt in.
@andrew_jennings Honestly I think a lot of it is habit plus each one being slightly better at one thing. Your setup works because the context is yours and lives in one place. That's the part most people are missing.
@mouad_builds Tool calling is where small local models fall apart. Tighter tool schemas and fewer tools per call often help more than swapping models. Also worth checking if a bigger quant fits in RAM.