It started because one of our maintainers pointed out that the newest Claude Code version just… did the thing. Our "deep research" skill was a 14-step chain: search the web, extract content, summarize, cross-reference, repeat. That exact workflow is now a single instruction in the base model. Same for "browse a page and pull key facts"—the model just calls a tool itself without me telling it how.
What we removed:
Multi-step workflow templates(superpowers-style). We had five variations for "agentic research", "plan-act-review", etc. Gone.** Web wrapper skills**(Agent Browse-like). The native tool loop is faster and doesn't blow up context.** Prompt chains for code review**. We realized the raw model caught the same issues we scripted step-by-step, with less fuss.
The speedup is real. Our context window usage dropped about a third, and response time feels snappier because we're not injecting long instructions for every step. One of our devs said it best: "We were basically doing prompt engineering for a model that already graduated."
The pushback we got #
Not everyone was on board. Two teammates pushed back hard—they'd invested weeks refining those workflows and thought removing them was a regression. One even reverted the change for a day. But after a quick side-by-side test on the same set of tasks, the base model matched or beat the skill chains every time. That ended the argument.
My practical takeaway for anyone doing** LLM agent cleanup**: go through your skill library and ask, "Can the current model do this from a natural-language instruction alone?" If yes, deprecate it. You'll simplify your AI workflow, cut context overhead, and improve real-world performance on actual tasks. This isn't a judgment call—it's just where the models are now.
We kept a few genuinely useful skills: anything requiring company-specific schemas, custom API auth, or interface with legacy internal tools. Those still add real value. The generic agent skills didn't. It's worth doing a pruning pass every couple of months because the baseline keeps moving.
Next How persistent memory solved our agent rollout problems →