[FEATURE]

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

Authoring a Skill today means writing SKILL.md by hand, or going through skill-creator‘s guided Q&A. That works well for tasks that are easy to describe, but it breaks down for workflows that are easier to show than to describe — multi-step UI/CLI/browser procedures that carry a lot of tacit detail (naming conventions, field defaults, the exact order I always do things). Those implicit steps usually only surface while I’m actually performing the task, so hand-authored skills tend to be slow to write and tend to miss them.
OpenAI’s Codex recently shipped “Record & Replay”: you demonstrate a workflow once and it drafts an editable, reusable skill (a SKILL.md-style file) that can later be replayed with new inputs. Claude Code already has the Skills system and supports the open Agent Skills standard, but there is no equivalent capture-by-demonstration path — the human still has to translate the demonstration into a SKILL.md manually.
Reference: Codex Record & Replay — https://developers.openai.com/codex/record-and-replay

Proposed Solution

Add a “record a skill from a demonstration” capability to Claude Code that:

1. Captures a demonstrated workflow — terminal commands + file edits at minimum, and optionally browser actions via the existing Chrome DevTools / browser MCP (and desktop/computer-use where available).
2. Drafts a SKILL.md from the trace — Claude inspects what was captured and writes a skill describing when to use it, the variable inputs, the steps, and how to verify success. This should be a natural-language skill, not a literal click/coordinate macro.
3. Writes it as a normal project skill at .claude/skills/<name>/, so it is readable, editable, diffable, reviewable, and shareable via git like any other skill.
4. Replays via /<skill-name> (or auto-invocation) with new inputs, using whatever tools are available in the session.

In effect this generalizes the existing /run-skill-generator idea (which already captures build/launch recipes into .claude/skills/run-<name>/) from “capture a build recipe” to “capture an arbitrary demonstrated workflow.”
Why this fits Claude Code specifically
• Portable, open-standard output. The artifact is a plain SKILL.md in the open Agent Skills standard, so a skill captured this way works across Claude Code, Claude.ai, and the API, and stays interoperable with other tools that read the standard.
• Git-native, not an opaque recording. Because it’s a plain file under .claude/skills/, code review, PRs, and version history all apply — unlike binary/opaque screen recordings.
• Cross-platform where possible. Terminal + file edits + browser MCP already work on Linux/Windows, so capture should not be limited to a single OS.

Alternative Solutions

Add a “record a skill from a demonstration” capability to Claude Code that:

1. Captures a demonstrated workflow — terminal commands + file edits at minimum, and optionally browser actions via the existing Chrome DevTools / browser MCP (and desktop/computer-use where available).
2. Drafts a SKILL.md from the trace — Claude inspects what was captured and writes a skill describing when to use it, the variable inputs, the steps, and how to verify success. This should be a natural-language skill, not a literal click/coordinate macro.
3. Writes it as a normal project skill at .claude/skills/<name>/, so it is readable, editable, diffable, reviewable, and shareable via git like any other skill.
4. Replays via /<skill-name> (or auto-invocation) with new inputs, using whatever tools are available in the session.

In effect this generalizes the existing /run-skill-generator idea (which already captures build/launch recipes into .claude/skills/run-<name>/) from “capture a build recipe” to “capture an arbitrary demonstrated workflow.”
Why this fits Claude Code specifically
• Portable, open-standard output. The artifact is a plain SKILL.md in the open Agent Skills standard, so a skill captured this way works across Claude Code, Claude.ai, and the API, and stays interoperable with other tools that read the standard.
• Git-native, not an opaque recording. Because it’s a plain file under .claude/skills/, code review, PRs, and version history all apply — unlike binary/opaque screen recordings.
• Cross-platform where possible. Terminal + file edits + browser MCP already work on Linux/Windows, so capture should not be limited to a single OS.
Alternatives Considered
• Hand-authoring SKILL.md (current state) — high friction, and tends to miss tacit steps.
• skill-creator — great for guided authoring, but it’s interview-driven, not demonstration-driven.
• External record/replay tools (e.g. OpenAdapt-style desktop record/replay, Playwright codegen) — these can capture and replay, but they don’t produce a Claude Code SKILL.md, so there is still a manual translation gap between “recording” and “reusable Claude skill.”

Priority

High - Significant impact on productivity

Feature Category

MCP server integration

Use Case Example

_No response_

Additional Context

The key differentiator I’d like to see preserved vs. a generic macro recorder: the output should be an inspectable, editable SKILL.md (intent + steps + verification), so it adapts to new inputs and a changing UI rather than replaying a brittle fixed sequence.