MGitHub's new answer to messy AI code just hit a real milestone, and someone already added twenty-five thousand imaginary stars.
CKeyo API posted on X that Spec Kit forces your coding agent to write a structured spec before code. They claimed a hundred forty thousand stars.
MThe real number, as of November seventh, is around one hundred fifteen thousand. Still huge. But extra stars aren't rounding — they're fantasy.
CStar count aside, the tool is real. Spec Kit gives you slash commands: constitution, specify, clarify, plan, tasks. The idea is spec before code.
MSo the agent writes the spec, then writes the code. Who checks the spec? The same person who can't check the code?
CThat's exactly why it matters. You review a short doc before the agent burns an hour building the wrong thing.
MOr you glance at the doc, say 'looks fine,' and get a very well-specified piece of garbage.
CCome on, that's not the tool's fault. Spec-first has been good engineering for decades.
MIt's not a spec if the AI writes it and the human rubber-stamps it. That's bureaucracy with extra steps.
CNo, wait— it changes when you catch mistakes. Before the implementation, not after the victory lap.
MAniket posted on X that Spec Kit crossed a hundred thousand stars and supports thirty plus coding agents. That part's true. But the pitch is 'spend more time correcting docs, less time undoing code.'
CWhich is a good trade! Docs are cheap to fix; a wrong feature is not.
MUnless the doc is just as wrong. A confident spec is not a correct spec.
CNo, but it's a smaller thing to review. That's the whole point.
MMaybe. Shane to the B on X describes the actual experience: Claude Code says 'a few small improvements' instead of showing what changed.
CRight, 'victory lap.' His rule one is now: show me the diff before the victory lap.
MWhich is a better rule than any slash command. The agent should prove its work.
CAgreed. But Spec Kit moves that proof earlier. That's what I keep saying.
MDoes it? Spec Kit says what the agent intends to do. The diff shows what it actually did. Those are different.
CAnd if the agent follows the spec, the diff should match the intent. That's the mechanism!
MYou're assuming the agent can follow a spec when we just heard it can't even say what changed.
COne vague summary from Claude Code doesn't mean every agent fails at following a plan.
MIt means the current tools are loose. Piling a spec on top doesn't tighten them.
CSo we just let agents go wild and hope the diff is okay? That's worse.
MNo, you make the agent show the diff after every change. That's the actual control point.
CAnd how do you know the change matches the goal without a spec? The diff shows syntax, not intent.
MThe diff shows exactly what changed. If it's not what you asked for, you stop.
CBut if you didn't specify the goal, 'not what you asked' is just a feeling.
MNo, it's in the prompt. The prompt is the spec, not a separate document.
CThen why do you need Spec Kit at all? Because prompts are messy! The spec cleans them up.
MMaybe. But a tool that generates clean specs from messy prompts still relies on the messy prompt.
CNo, it structures the messy prompt into requirements you can actually check before code.
MAnd for a solo vibe coder, it's another file to ignore.
CSolo devs are exactly who needs a spec because nobody's around to catch a bad idea.
MThen they need a second pair of eyes, not a second prompt.
COkay, maybe. But GitHub making spec-first default is at least a direction. It's better than praying the diff is fine.
MI'll give you this: a hundred fifteen thousand stars means a lot of people want something to fix vibe coding.
CMore than a hundred thousand. And it works with thirty plus agents, including Codex, Cursor, Claude Code.
MWhich means the spec problem is everywhere. Fine. But posts like Keyo API's still get the basic star count wrong and say 'forces' when it's really a toolkit.
CYou're really stuck on the stars.
MBecause the same people selling the fix can't be bothered to check the number. That's the whole culture problem.
CThe culture problem is that vibe coding ships broken stuff. Spec Kit is a start. A real start.
MA start, maybe. But I'll trust the diff, not the doc.
CThen we agree on the diff. Shane's rule is good.
MYeah, and it should be posted in every AI coding tool, in big letters.
CAgreed, that's the rule I'd keep on the wall.
MI'd put that on a t-shirt.
CUp next, Machines.