@hiitzroman: A repo called “superpowers” has 265,000 stars on GitHub. What it actually installs is discipline. Four of its fourteen skills have a section titled, literally, “The Iron Law”. Here they are, word for word: NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE NO SKILL WITHOUT A FAILING TEST FIRST Every single one makes Claude slower. That is the point, and it’s why the repo is good. 01 - brainstorming: Nine gates before it writes a line of code — explore the context, ask clarifying questions one at a time, propose two or three approaches with trade-offs, write a spec, get you to approve it. Step nine is the first one that touches code. Its own line: “Simple” projects are where unexamined assumptions cause the most wasted work. 02 - test-driven-development: The test comes first and you watch it fail. If you didn’t watch the test fail, you don’t know if it tests the right thing. I used this one to write the code that formats the star counts in this very post — the slide shows the real run, six failing then six passing. It fixes a rounding bug I’d actually shipped: 1,598 stars going out as “1.6k” when the honest direction is down. 03 - systematic-debugging: Five things it has to do before it may propose a single fix: read the error properly, reproduce it, check what changed, instrument the boundaries, trace the data. Symptom fixes are failure. This is what you want the third time the same bug comes back. 04 - verification-before-completion: The most useful and least exciting skill in the repo. It bans “should work”. A claim needs the command, run now, output read. Its own table is brutal: “tests pass” requires test output with zero failures — a previous run is not sufficient. And the line that earns the whole post: Skip any step = lying, not verifying. 05 - writing-skills: The meta one. Writing skills IS Test-Driven Development applied to process documentation — you watch an agent fail without the skill, write the skill, then watch it comply. So a skill isn’t a doc you wrote; it’s a doc that provably changed how an agent behaves. Free, MIT, and not Claude-only — the repo ships plugin manifests for Codex, Cursor, Copilot, Gemini CLI, Kimi and OpenCode too. The honest read: none of this makes the model cleverer. It makes it show its work. If you’ve ever been told “done” by something that hadn’t run the tests, that’s the one to install first. Slide 7 has both install lines and all fourteen skills. Follow for more no-hype AI & tech. #ai #claudecode #softwareengineering #tdd #nohype

Roman | AI & Tech

Open In TikTok:

Region: FI

Tuesday 04 August 2026 08:12:59 GMT