🤖 Building with AI, in public ⚔️ @AsgardAgentsAI: AI Agents 🧊 @MeshWright: AI Scan to CAD 🎨 @SkaldStudio: AI Development 🔎 @C3ResearchAI: Crypto Research
Georgia, USA
Joined August 2012
- Tweets2K
- Following125
- Followers454
- Likes5.3K
A thumbnail safe zone can shrink the design too far.
Codex made thumbnails for my 5 Shorts, and its first safe zone left the design too small. It took a second round to get right.
How much margin do you leave on yours? #BuildInPublic #Codex
13
My readiness check now flags a checkout left off main.
I wanted Monday's analytics run to go ahead, so it now stops a job that's about to run from the wrong branch.
What does your preflight check? #BuildInPublic #DevTools
45
The Asgard Launchpad is open!
🔥 1 Launch Per Day
⚡️ Launch with Trust
Before cleaning my drives, I checked what was eating space and CPU.
Containers I didn't need were running. I stopped them and cleared out the unused ones, which freed up CPU.
What do you check first on a cluttered dev machine? #DevTools #BuildInPublic
5
70
Every new AI coding session starts with a blank memory. Anthropic published how they kept Claude making real progress on a full web app across session after session anyway.
The article is "Effective harnesses for long-running agents." Here's the setup they used, and how to copy it on your own projects. Bookmark this before your next build.
The failure modes they saw first:
The agent tried to build the whole app in one go, ran out of context halfway, and left half built features with no notes. Later sessions declared the project done when it wasn't. And features got marked complete after a quick unit test or curl check, without anyone actually using them.
If that sounds like your last week with an agent, here's the fix.
1. Run a setup session first.
Anthropic used a separate first session with its own prompt. Its focus was setting up the environment: an init.sh script to start the app, a progress log called claude-progress.txt, and a first git commit.
2. Turn the prompt into a feature list.
The setup session expanded the request into a full list of features. For their claude.ai clone, that was over 200. Each one had a description, the steps to test it end to end, and a "passes" field.
3. Start every feature as failing.
Every feature began as "passes": false. The prompt told the agent to change only that field and called removing or editing tests unacceptable. The list becomes a contract, not a suggestion.
4. Use JSON for the list, not Markdown.
They chose JSON because the model was less likely to rewrite or mangle it than a Markdown file. Small detail, real difference.
5. Start every session the same way.
Each coding session was told to start by checking the working directory, reading the git log and the progress file, reading the feature list, and starting the app with init.sh. Then it ran a basic end to end check to catch anything the last session broke, before touching new code.
6. One feature at a time.
Anthropic called this critical. Working on one feature per step stopped the agent from biting off more than it could finish in one context window.
7. Test like a user, not like a unit test.
They gave the agent browser automation and had it click through features the way a person would. That substantially improved results. The agent was told to mark a feature passing only after that kind of test.
8. End every session with a commit and a note.
A descriptive git commit and an update to the progress file. The next session reads them first, and git lets it roll back anything that went wrong.
How to steal this today:
Keep a features.json with every requirement set to failing. Keep a progress.txt that each session reads first and updates last. Keep a start script so no session wastes time figuring out how to run the app. Tell the agent to work one feature at a time, test it like a user, then commit.
One honest note: this is the harness Anthropic tested for one kind of project, not a universal law. But it targets the two failure modes Anthropic saw: agents doing too much at once, and agents saying "done" too early. If you've used coding agents, you've probably seen both.
Search "Effective harnesses for long-running agents" to read the full write up.
Save this for your next agent build, and send it to the builder whose agent keeps redoing yesterday's work.
1
5
144
I picked my site's logo direction after 3 rounds of Codex options.
Then I rethemed it to my banner palette and ran 10 glass polish rounds. The pick was Daybreak, a blue system.
Do you let agents draft your logos? #BuildInPublic #Codex
1
1
72
Four banks just put $270M into one joint stablecoin.
Goldman, Citi, Santander and Deutsche building together, not racing, signals they expect stablecoins to be core plumbing. Banks rarely share rails.
Would you settle on it? #Crypto #Stablecoins
3
58
Trump Media bought 1,510 Bitcoin worth over $112M in Q3.
Per @WatcherGuru. One more public company putting BTC on the books as treasury, not a trade. That keeps making it an easier call for founders.
Would you hold BTC in your treasury? #Bitcoin #Crypto
1
2
80
Anthropic is reportedly set to IPO next week.
Per @WatcherGuru via WSJ. Going public puts their revenue and burn rate on the record, our first real look at frontier AI economics.
Buy at open or wait for the S-1? #AI #Anthropic
1
2
94
🚨 BREAKING NEWS: Claude Fable 5.2 has been released!
It still destroys your usage...
Hello Opus my old friend🎶
1
106
PNC Bank is testing USDC payments on Ethereum for corporate clients.
Per @WatcherGuru, it also tests converting USDC straight into dollar deposits. A top-10 bank using a stablecoin as payment rails is a big shift.
Would you pay vendors in USDC? #Crypto #Ethereum
47
Passes 2 to 9 of a Codex review loop on my personal site added:
- A Pause control for the moving sky (WCAG 2.2.2)
- Forced-colors states
- Reduced motion fixes
- Direct shop links #BuildInPublic #Codex
48