# Do we still need skills?

What a few months of our own agent transcripts say about skills and memory

Canonical: https://agentrun.ai/blog/do-we-need-skills-anymore
Published: 2026-11-02T19:40:43.000Z
Author: Miguel Rios Berrios

_What a few months of our own agent transcripts say about skills and memory_

For weeks I thought I was using our /ship-pr [skill](https://agentskills.io/), the one we use to take a PR from open to merged, but it turns out, this skill was never installed on my machine. The first time Claude went looking for it and came up empty, it found the procedure in other files in our repo and wrote itself a [memory](https://code.claude.com/docs/en/memory) note, and from then on it shipped my PRs off that note.

![Workflow diagram: an unknown ship-pr skill leads the agent to read its notes, ship the pull request and update the note for next time.](/blog/imported-writing/do-we-need-skills-anymore/97d6a901e083dd1d.png)

That made me realize I hadn't seen one of my agents load a skill in a while, so I went through every Claude Code transcript on our main development machine since June, about 14,000 of them and a bit over a million tool calls, and counted every skill load and every time an agent read or wrote its memory.

## Skill issue?

All summer, the share of sessions that read memory and the share that loaded a skill moved together, and in October memory pulled away. Half of our sessions now read memory, a quarter load a skill, and skill loads went from two or three per thousand tool calls to about one.

![Chart comparing agent sessions that read memory with those that call a skill from July through October.](/blog/imported-writing/do-we-need-skills-anymore/2a716d6812d40c39.png)

I think skills were a byproduct of models that couldn't work out a procedure on their own and harnesses that didn't notice what you do over and over, and that both of those have changed. In my experience, Claude Code's memory has gotten pretty good at remembering the things you do repeatedly. Someone on our team had written the ship-pr procedure down once, and the agent did the rest. If the harness can tell that something was valuable and repeatable, remember it and save it as a markdown file, I'm not sure we need to sit down and write it as a skill first.

Most of the skill loads we still have are ones we forced. We use 67 skills for engineering across the machine and our repos, and only half of them were loaded at all in October. Of the loads, 40% were for skills our [CLAUDE.md](https://code.claude.com/docs/en/memory) files tell the agent to load, 26% were named in the prompt and 12% were typed as a command. The agent picked a skill on its own about one time in five.

![Reasons agents loaded skills in October: required by CLAUDE.md, named in a prompt, typed as a command or chosen by the agent.](/blog/imported-writing/do-we-need-skills-anymore/91dc7d8cf84d148d.png)

Looking back, we were also using skills for things that belonged in the prompt. In [grep.ai](https://grep.ai/), research jobs loaded their effort level, their expert's procedure or both as skills, so the first thing each job did was spend turns finding and loading them. Across 445 production jobs, that was a median of 9.2 seconds before the first useful tool call. An agent going through our telemetry spotted it, and we moved those skills into the system prompt. My rule of thumb now is that a skill that's always loaded probably belongs in the system prompt, or in [CLAUDE.md](https://code.claude.com/docs/en/memory) as a pointer to a file.

## Are skills still useful, tho?

I think they are for two things: organizing knowledge and sharing it. To me, a library of skills in one standard format beats pointing people at a blog post here, a tweet there, a gist, an internal note and a doc buried in someone's repo. It's also an easy way for person A to hand a method that works to person B, or from one company to another, and Lauren Tan's [pstack](https://github.com/cursor/plugins/tree/main/pstack) and Garry Tan's [gstack](https://github.com/garrytan/gstack) are good examples.

Every time I make that argument, though, I can make the counter-argument too. A skill is a piece of text, and a piece of text can live in a blog post or in a tweet. The scripts that come with it are things an agent can often find with a web search, on GitHub or in your codebase. More and more of what I see online is "tell your agent to do X" or "point your agent at this post", and installing software works the same way: I'm not sure you need an install skill when the agent can go read how the thing works. I expect that to get easier as models, harnesses and search get better.

## And there goes a useful standard

I don't think it's a secret that models are being trained to use their own harness more efficiently: Claude models in Claude Code, OpenAI's models in [Codex](https://developers.openai.com/codex). Part of what made skills appealing was [a standard](https://agentskills.io/) that worked across all of them. Our numbers make me think Anthropic is betting that the model knows better than you when to write a note, update it and use it, without you stepping in, and I wouldn't be surprised if Codex is doing the same.

The two use different memory systems, so if what used to be a skill now lives in each harness's own memory, it doesn't travel with you. Those of us who switch between harnesses and models, which is most of the people I know, would feel that lack of portability, and I think open-source harnesses will need their own answer for memory. It would be a shame to lose a standard that works across harnesses if people stop writing skills.

## So, what's next?

Our numbers come from one team on one machine, mostly in Claude Code, so they show how we work and not how everyone does. We also have [recall](https://github.com/Parcha-ai/parcha-skills/tree/main/recall), which lets anyone at Parcha search what every agent in the company has done (ironically, it's a skill), and that probably makes our agents more aware of how much past work they can learn from.

And we still write procedures down, because text helps until the model can work something out on its own; we call that documentation. I think skills will stay useful for organizing and sharing knowledge. For everything else, my bet is that the harnesses eat them and skills end up a victim of intelligence getting better. If my own usage is any indication, I don't think we'll be talking much about skills a year from now.