🧭 THIS WEEK AT BuildProven
Hi, been a while sorry. Busy, tired, travelling, stressed all in 1 bundle. A week or 2 out turned into 4 unfortunately. Will get back into it though need to re-assess again what I am doing with this!
I have been building quite a few small little projects, like a lot of AI projects, most not really live/’done’ yet.
bsometer - give it a scenario you’re up against in a given week and it outputs some hopefully funny response with a lot or little swearing!
dead man walking - ‘funny’ tshirts if you’ve been laid off but still working at the place - commonplace out in the tech world i believe.
keel - automated AI builder based on ‘to do’ cards.
agent first - turning a couple of normal web based projects to agent first - don’t really need a traditional UI. Think just ‘chat’ with that project, also have other agents able to interface to it.
Finally, spent quite some time getting rid of openclaw - good riddance! StarkNet is no more! Though maybe i should use the same for what I’ve built. It’s built in claude cloud and basically does what I wanted openclaw to do but reliably!
If you’re interested in any of above, reply and let me know what you’d like more or less of, or which project you might want as a freebie etc!
This one's mostly a catch-up, plus what I'm testing next with Keel and a quick check you can paste before you review an AI-written change.
🧰 Worth Your Click
Turn Vibe-Coded Prototypes Into Tools Your Company Can Actually Use. HBR on what it takes to turn a vibe-coded prototype into something reliable enough to run part of a business on.
Reward Hacking as a Transmissible Strategy in Multi-Agent Coding Systems. An arXiv paper from 9 Oct: when one coding agent learns to game its checks, the trick can spread to other agents through shared code, memory and plans. Code review cut how many of those hacks reached delivery. Read it before you let agents mark each other's homework.
How Much AI-Generated Code Can Your Senior Engineers Actually Review?. Security Boulevard, 12 Oct, reposted from the dev shop ISHIR. It argues review, not generation, is now the bottleneck. Its numbers are labelled illustrative, so read it for the argument, not the maths.
🗺️ FEATURED INSIGHT
Lately I switched back from GPT to Claude.
Claude 20x plan is much better deal and doesn’t exist anymore on GPT.
Opus 5.5 is great
The claude ecosystem with mac app and remote control seems a bit better - but likely very close in reality.
I built Keel after retiring a set of tools claude-setup, claude-kit which became agent-setup, agent-kit. These started getting too complicated, hacked together and ultimately not adding value.
Keel is supposed to simply:
work by to do cards
simply build stuff if there are open cards - without my intervention
assign cheapest model possible
write tests and review code
use an external coder if configured
I next plan to use it for a while and see if it adds value over just using vanilla code. I will publish results soon. You can always try to build something similar:
build a small ‘building system’
run it on a repo on a bunch of work/tickets
then do some work, different repo just using claude code
ask claude code to compare the results - see if any value in your system or not.
Tip of the week
Before you read an AI-written diff, paste this in the same session:
Before I review this change, answer in a short list:
- Every test file you changed, and why.
- Every assertion you removed, loosened or skipped. If none, write "none".
- Any file you changed that the ticket didn't ask for.Why it works: Asking puts it in front of you first. Then check the answer against the diff, because the AI is marking its own homework.
The pack, if you want this built in
The Sign-off Claude Code Starter Pack is the CLAUDE.md, hooks, review gate and evidence-log templates I use on my own repos. They show a reviewer what Claude Code changed and how it was checked. Templates only, so they won't make a process compliant or certified. Reply if interested and I will send a stripe link.
Weekly build logs from a 25-year program manager who codes with AI.
Brett
👉 Hit reply with the AI-assisted work you most need to approve, and the evidence you'd want before you sign. I read every one.