% open the-build-loop
The Build Loop
Four small tools that decide what to build, pressure-test it, then write it up — so the system around building runs itself.
PROBLEM
I can build a thing in a day. The hard part was never the build, it was everything around it: deciding what's actually worth building, catching myself before I burn a session on the wrong idea, and then the worst job of all — writing up what I just did. The build is the fun bit, so the deciding and the writing kept getting skipped, which meant good builds never became momentum.
WHAT I BUILT
Four small pieces that wrap the build and hand off to each other. Forge gives me three to five scored, one-day ideas every morning and learns my taste over time. Ascend takes the one I pick and pressure-tests it before I commit — attacks it, names the harder move I'm dodging, prices what the session costs against everything else on my plate. Then I build. Then Extract reads the build and writes the portfolio page and a post in my voice. And underneath it all, a quiet hook sets how hard the AI thinks per task, so effort goes where it matters and not where it doesn't.
HOW IT WORKS
All of it lives inside Claude Code as skills and one hook — no separate app. Forge scores each idea on a seven-part rubric (does it compound, does it make content, is it shippable today, is it fun) and logs what I actually ship so it sharpens. Ascend steelmans an idea then tries to kill it, and runs one real search to check it's not been done to death. Extract reads the repo, the project memory and the chat, asks me three quick questions, and drafts. The effort hook reads each prompt I send and sets the reasoning level automatically, set and forget.
WHERE IT'S AT
Live and in daily use. Forge fires at the top of a build block. Extract wrote this very page you're reading. The effort hook ran on the message that asked me to build it. Ascend is built and wired into the loop, though I haven't run it live on a real call yet — that's the next test.
WHAT I LEARNED
The honest lesson came from Ascend. The first version claimed it could make an idea 'world-class' by having the AI critique itself against a fixed standard. I pressure-tested that claim and it fell over — a model grading its own idea with no outside truth doesn't elevate the idea, it just flatters it. So I cut it down to what's actually real: friction, not prophecy. The win across the whole loop isn't a smarter model, it's a better-equipped one. Same lever that took my restaurant site from a flat v1 to a v2 that looked expensive: upgrade the tools and the rubric, not the effort.
STACK
Claude Code (skills + hooks) · Python · Markdown
% cd ..