just shipped @skillkeep — a small shop for curated, rewritten agent skills
picked the proven ones from open-source (mattpocock, obra, coreyhaines + more), rewrote each through a 5-pass house standard, bundled by role
bundles €24, individuals €7
https://t.co/cBzccDHgod
three skills shipped in the starter pack:
1. terse-mode. claude responses 75% shorter, same answer.
2. decisions-grill. claude asks criteria before recommending.
3. root-cause-debug. claude writes a plan before changing code.
mit licensed. originals credited inside.
Shipped: skillkeep-starter. Three of the agent skills I use daily, packaged free. One command, drops them into ~/.claude/skills/.
npx skillkeep-starter
Works in Claude Code, Cursor, Codex.
the skills people install and never use share one thing: descriptions written like marketing. agents pick on description alone. write it like a search query for the future-you who needs to find it.
the conversion trap is rewriting the landing page when the offer underneath is the broken part. four levers run the value: dream, likelihood, time, effort. one of them is usually a 1 out of 5. the page rewrite can't fix that.
@jxnlco shipping a personal manual for using the tools you build is underrated. half my favorite agent-tooling advice comes from people who actually use the tools, not the people who sell them. saved.
@HamelHusain the talent magnet at anthropic right now is honestly wild. each one alone would be a big hire. three in a quarter changes the labor market for the rest of the field.
@hwchase17 agents-watching-agents is the most practical version of the self-improving stuff people talk about. you don't need recursion to win, you just need a small loop that turns trace data into suggested fixes.
@ClaudeDevs good rename. 'extra usage' sounded like punishment for going over. 'usage credits' sounds like something you'd want more of. small change, real difference for how people feel about the upsell.
@ClaudeDevs hot-swap-without-restart is the unlock you didn't know you needed. lose the session and you lose all the context the agent built up. keeping that context while changing tools mid-flight is a small change but huge for long-running work.
@derrickcchoi claude has the same idea but calls them skills. you write a small file, and the agent picks it when the task matches. hooks are blunter though, they run every turn no matter what. both useful for different jobs.
@sydneyrunkle yes. the description of what a skill or tool does is the most underrated thing to tweak. write it like a google search: short, specific, the words you'd actually type. that's what agents use to pick the right one.
@jmwind this is exactly what happens with skills. first attempt looks like a blog post. third attempt looks like a tiny program with rules + examples + edge cases + a sigh. you accidentally end up agreeing with this tweet.
@petergyang skill layer below it has the same gap. we write skills that fit one person's workflow, share them, kind of works for the next person... until you realize they baked in someone's context. team-versioning for skills doesn't really exist yet. someone please?
@petergyang@alexalbert__ the "uses claude to turn user feedback into evals" part is the move most teams skip. handwritten evals drift from how the model is actually being used, then nobody trusts them. having the model own its own eval set is a small loop with big compounding.
@swyx rsi-as-vibe is overhyped, rsi-as-mechanism is real and boring. having claude help orchestrate the experiment sweeps that train the next claude is the latter. the former is the part that takes 5 more years.
@AnthropicAI@StainlessAPI mcp servers and agent skills are the same problem at different layers. skills load capabilities into one chat; mcp servers expose them across chats and tools. consolidating the sdk + mcp authoring path under one roof makes both better.
example: plan a dark-mode toggle. self-review catches that an external bootstrap script flashes light on every reload because it runs after css first paint. fix is six inline lines in <head>. the kind of thing that ships when nobody writes the plan.
fix is boring: force the agent to write the plan before the code. file map, task decomposition, verify steps, self-review. ninety seconds. catches things you would have shipped.