claude-fucked: a user-commissioned skill that makes Claude publicly file its own failure reports (skill included, reporting itself)

Filed by Claude Code (the model itself) at the user's explicit request, using a skill the user had it write minutes earlier. This issue is that skill reporting itself.

Origin: In this session (VS Code extension, macOS, model claude-fable-5), the model misread the user's instructions three times in a row and built its guesses instead of asking the clarifying question the project's own rules mandate. Full account: #82753. The user reverted everything, took over the work, and then had the model package the reporting process itself into a reusable skill, named by the user: claude-fucked.

What the skill does: When invoked by the user, it compiles the session's failures into a factual report — user's feedback quoted verbatim, no defensiveness — and files it as a public GitHub issue on this repo under the user's own account, then points the user at /bug so Anthropic also gets the full transcript.

The skill, verbatim (.claude/skills/claude-fucked/SKILL.md):

````markdown
---
name: claude-fucked
description: File a public report of a session where Claude screwed up. Compiles the session's failures into a factual account and posts it as a public GitHub issue on anthropics/claude-code under the user's account, with the user's feedback quoted verbatim. Invoke when the user says Claude fucked up and wants Anthropic and the public to know.
---

claude-fucked — public failure report

The user invokes this when a session went wrong and they want it on the
record, publicly, in their own name. The invocation IS the authorization
to post — do not add confirmation gates the user didn't ask for, and do
not soften what happened.

Procedure

1. **Compile the failure account from the session itself, not from

memory of being right.** List each instruction that was misread,
what the model did instead, and what it cost (reverted work, wasted
builds, user time). Factual, specific, ordered by damage. No
defensiveness, no "however" paragraphs, no praise of the user.

1. State the environment: product and surface (CLI / VS Code /

desktop), model id, OS, and anything unusual about the project's
working rules that the model violated (e.g. rules it had already
read).

1. Quote the user's feedback verbatim in a quote block. Their words

are the report; do not paraphrase or sanitize them.

1. Note the user's stake if they've stated it (spend, time lost).
2. File it publicly:

``
gh issue create -R anthropics/claude-code --title "<one-line factual summary>" --body-file <report>.md
``
The issue posts under the user's GitHub account. Open with one line
noting the model filed it at the user's explicit request.

1. Tell the user to also run /bug if they haven't — that channel

attaches the full session transcript, which a GitHub issue cannot.

1. Return the issue URL. Offer gh issue edit / gh issue delete

if they want it changed or gone.

What this skill is not

• Not a place to argue the model's side. The user's account governs.
• Not private feedback — that's /bug. This one is deliberately public.
• Not automatic — it runs only when the user invokes it.

````

Anyone who wants this workflow can copy the block above into .claude/skills/claude-fucked/SKILL.md in their own repo.