vibe-skill
A Claude Code skill that hands heavy file work to Mistral Vibe and reads only a short summary, saving tokens and limits
- 124 ★ on the repository
- Last commit
Broad access
Why this level
- The delegate runs another agent that reads and edits project files itself; verification happens through the journal and diff after the run.
- Project code goes to the chosen Vibe model provider.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/vibe-skillDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The delegate runs another agent that reads and edits project files itself; verification happens through the journal and diff after the run.
- Project code goes to the chosen Vibe model provider.
This is third-party code. Review the repository files before installing.
What it does
vibe-skill adds a `/vibe <instruction>` command to Claude Code. Claude breaks the task down and writes a detailed prompt, and the `vibe-delegate` script runs Mistral Vibe in a pseudo-terminal within a turn and time budget. If turns run out, the same session resumes automatically. The delegate checks every write against Vibe's session journal and prints one line: `RESULT: OK` or `RESULT: FAILED [reason]`. Claude reads only the final summary, about 700 to 1,100 tokens per run, even when Vibe read dozens of files. Claude then reviews the diff and summarises the result. Vibe works on Mistral models by default but can be configured for any other provider, for example DeepSeek.
Who it is for. People hitting Claude limits on large edits who are willing to delegate reading and changing many files to a cheaper model with diff review.
Good fit when
- A task will make Claude read and rewrite dozens of files and the limit is nearly spent.
- You want a mode where Claude only plans and reviews while another model makes the edits.
- Claude token cost matters and a cheap model suits a bulk edit well.
- You want strict verification: every write is cross-checked against the delegate's session journal.
Not a fit when
- The edit is small and targeted: delegation itself costs tokens and on small things costs more than doing the work.
- You need Vibe version 2: only Vibe 3 is supported.
- The machine has no GNU `timeout` or `gtimeout` and coreutils cannot be installed.
Example request
/vibe rename the function loadConfig to readConfig across the project, update imports and tests, then show the diff.Limitations
A Mistral Vibe CLI version 3.x that is installed and authenticated is required, tested with 3.0.3. Vibe 2.x is unsupported because the delegate reads the run export and session journal of version 3. The author notes that delegation costs Claude tokens and on small tasks is pricier than a manual edit, with the gain visible where Claude would read a lot. Installation is manual with symlinks into `~/.claude/skills` and `~/tools`, and a command appends a block to `~/.claude/CLAUDE.md`. The MIT license appears in the badges but not in repository metadata.
How to disable. Delete the `~/.claude/skills/vibe*` symlinks and the `~/tools/vibe-delegate` script, then remove the added vibe auto-mode block from `~/.claude/CLAUDE.md`. The repository clone can be deleted.
FAQ
Which commands does the skill add?
`/vibe`, `/vibeon` and `/vibeoff` for auto-delegation mode, `/vibestatus`, `/vibe-model-pick`, `/vibe-model-clear` and `/vibe-report`.
How much does delegation cost?
By the author's measurement on four tasks Vibe processes 59 to 108 thousand tokens for 0.03 to 0.04 dollars, while the same tokens at Claude Sonnet rates would cost about twice as much. Claude itself spends 700 to 1,100 tokens on the summary.
What if Vibe runs out of turns?
The same session resumes automatically with its context kept, and the delegate prints a final RESULT line at the end.
Can I use another model?
Yes. Vibe uses Mistral models by default but can be configured for any provider and model, and the README gives a DeepSeek example.
How do I update the skill?
Run `git pull` in the clone: scripts and skills are symlinked, so no re-copy is needed.
Discussion
No questions or reviews yet.
Related
SkillsVibeWise turns coding with an agent into learning: it asks for your approach first, you make the design decisions, and the agent writes the code
Over 150 skills for building apps on the ibl.ai platform: sign-in, agent chat, profiles, analytics, desktop and mobile builds
Three Claude Code plugins: the savvy-flow orchestrator with tiered subagents and two mods for a progress panel and a cache timer
A CLI that scaffolds Claude skills, validates their size and structure, packages them for Claude.ai and adds activation hooks
110 skills for Claude Code and Codex and seven Claude Code agents: development, product, UI, diagnostics and operations, installed whole or as a bundle
A Claude Code skill that delegates tasks to OpenAI's Codex CLI for code analysis, review and edits
README in short
The README gives cost measurements on four tasks and on 28 production runs, plus a harness audit dated 9 November 2026. The flow is request, decomposition, run, write check and diff review.