Save a finished chat as a skill or agent in one click

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

When I work on something complex with Claude, the first try is always generic. It gets good only after I correct it a bunch of times: cut this, format it like that, sound like this, add more depth and domain knowledge here. By the end it's exactly right. But all of that lives in one chat.

And I can't just keep working in that one chat. The longer it gets, the worse it runs, it slows down, drifts, and starts forgetting things I said earlier, so the quality I worked to build up starts breaking down. So I open a new chat, and now it is gone. I have to do the same iterations and corrections over again, re-feeding context every time, and it just wastes my tokens.

Skills and agents are meant to fix this. But building one from a chat is slow and manual. You have to reread the whole thing, figure out which notes were one-offs and which are better rules, then hand-write the file. Most people won't bother. So a good session just gets thrown away.

Proposed Solution

A button at the end of a chat that turns it into a skill or an agent for me. Something that maps out the whole architecture of what made it click, pulls out my corrections, keeps the rules and drops the one-offs, and hands me a draft I can check and save.

Let me pick which one it becomes. A skill if I just want the saved rules and style to come back in future chats. An agent if I want a worker I can hand the whole task to and have it run the same way every time. Same capture step, I choose the output.

Alternative Solutions

Right now the only option is to do it by hand: reread the chat, work out which corrections are durable rules versus one-offs, and write the SKILL.md myself. Even then it's not reliable, the hand-written version doesn't reproduce the same output, so I end up correcting it again anyway. And it's slow enough that most sessions just get dropped instead. Project instructions and CLAUDE.md help a bit, but they still have to be written and maintained manually, and they have the same problem. None of them reliably reproduce what the original chat was doing.

Priority

High - Significant impact on productivity

Feature Category

Configuration and settings

Use Case Example

I edited a long essay over four rounds. The first draft was bloated and generic. The last one was good: in my voice, properly formatted, with the depth it was missing. Huge difference, and all of it came from corrections that are now stuck in that one chat. If I could've saved it as a skill or an agent, my next doc would start where that one ended.

This isn't just writing either. The same thing happens when I design a website or architect a system: I spend a whole chat getting the design language, the structure, the spec exactly how I want it, and then the next project starts from zero. Any task where you tune Claude to a specific way of working, it hits this same wall.

Additional Context

_No response_