vibe-skill

vibe-skill

Claude orchestrates. Vibe does the heavy lifting. You review the diff, save tokens, costs and avoid hitting limits!

Claude only reads the script's summary (≈700–1,100 tokens per run), however many files Vibe reads internally (a median Vibe 3 run processes ≈215k tokens). Delegation still has a fixed cost on Claude's side, so it pays on tasks that would make Claude read or iterate a lot, not on small targeted edits: see the measurements below.

Note that Vibe works natively with Mistral models which are capable and significantly cheaper than Claude, but Vibe can also be configured to use any other provider/model instead. Eg you can use a deepseek model with vibe tooling.

Summary:

1. User types /vibe <instruction> in Claude Code
2. Claude decomposes the task and writes a prompt
3. vibe-delegate runs Mistral Vibe in a pseudo-TTY, within a turn and time budget
4. If Vibe runs out of turns, the same session is resumed automatically (it keeps its context)
5. The delegate checks every write against Vibe's session journal and prints one verdict line: RESULT: OK or RESULT: FAILED [reason]
6. Claude reviews the diff and summarizes the result

Why

Cost savings — Vibe's file reads and edits consume delegate tokens (or whatever model you configure), not Claude tokens. Most of a run's input is the conversation re-sent on every model call, and about 92 % of it is served from the provider's cache. Measured on 4 benchmark tasks with glm-5-3 (2026-11-09; cost as billed by Vibe's own run export):

Production runs are larger: the median Vibe 3 run (28 runs, 2026-10-29 → 2026-11-09) processed 215k tokens. Their cost was logged before the harness read the cache split, at list price with no cache discount ($0.35 median; Claude equivalent $0.88). On the benchmark runs that method overstated the delegate cost 3.1 – 3.8× and the Claude equivalent 3.5 – 4.8×. Since 2026-11-09 each run logs cost_usd from Vibe's export (cost_source: vibe_export) and a cache-aware cost_claude_eq. The "same tokens" comparison is only indicative: Claude doing the task itself would make a different number of calls. On a Mistral Pro subscription, usage inside the plan's limits costs no API fee. List prices checked on 2026-11-09: Mistral Medium 3.5 $1.50/$7.50, DeepSeek V4 Flash $0.14/$0.28 per M tokens.

Delegation does cost Claude tokens, and on small tasks it costs more than doing the work. Measured on 2026-11-10 with headless claude -p (Sonnet 5.5, same tool set in both arms, delegation forced in the Vibe arm, hidden acceptance tests passed in every run):

Session	Claude does it	Claude delegates to Vibe
1 task, same change in 10 modules	$0.20	$0.24–0.32 (4–8 delegations)
3 tasks in a row, 10 modules each	$0.19	$0.25–0.26 (3 delegations)

About $0.14 of every session is Claude Code's fixed context (system prompt, tools, skill list), paid either way. On top of it, a delegation costs Claude ≈$0.03 (writing the prompt, extra turns, reading the report) plus ≈5,300 tokens once per session to load this skill (it was ≈12,900 before SKILL.md was cut on 2026-11-10). These grep-able edits cost Claude only ≈$0.02 each when done directly: it finds the lines and edits them without reading whole files, whatever the file size (2.9 to 26 KB per module made no difference). Delegating pays when doing it yourself would cost Claude more than ≈$0.03 per task: roughly when it would have to read more than ≈7k tokens (≈25 KB) of code in full, or iterate (debugging, exploration, several failed attempts). Small samples (2 runs per cell): treat the figures as orders of magnitude. The earlier modelled claim (savings from ≈4 files touched, see handoff-probe/ECONOMICS.md) is not supported for targeted edits.

Pro plan users: Mistral Vibe (CLI included) is part of Mistral's Pro plan, $14.99/mo excluding tax as of 2026-11-09. The pricing page shows only a small API-credit figure and "fair usage" wording for Vibe; the Vibe Code allowance actually included is much larger than that figure but is not published, so check your own account. Beyond it, usage stops or continues pay-as-you-go at API rates (off by default). Inside it, delegations cost no API fee: in practice the Vibe side is free, and the real cost of a delegation is on Claude's side (above). The free plan also includes limited Vibe CLI use.

Delegation synthesis — as of 2026-06-08 (Vibe 2.x, historical)

These numbers predate Vibe 3 (released 2026-10-29), which changed the CLI's exit codes, agents and session format. Right after the upgrade, 12 of 31 delegations failed (39 %) and the harness misread most of them. The 2026-11-09 audit (docs/harness-audit-2026-11-09.md) has the causes, the fixes and a 24-run benchmark with hidden acceptance tests: the verdict was correct on 23 of 24 runs. Fresh Vibe 3 production numbers will replace this section once enough runs exist.

Snapshot over 2,450 vibe delegations, 2026-05-12 → 2026-06-08 (28 days). Pulled from the shared run log via delegate-report (vibe scope).

Performance & savings

By model

Real cost methodology: All Mistral runs (mistral-medium-3.5 + Le Chat interface) covered by Le Chat Pro subscription — $0 marginal cost. DeepSeek $0.67 is actual spend (from logs). Hit rate of 90.5% back-calculated from that spend. Claude equivalent applies the same hit rate with Claude's cache pricing ($0.30/M hits, $3/M misses, $15/M output), assuming 97.6% input / 2.4% output split.

deepseek-flash was the value pick (88% exit-ok at ≈$0.002/run: $0.67 over 369 runs).

Error rate (earlier snapshot of 2026-05-30: real projects, 1,964 runs, synthetic test scaffolds excluded)

At the time, the ≈26 % exit_error + wrote_nothing was attributed to Vibe engaging but landing no edit: multi-edit context drift, or a first search_replace anchor not found byte-for-byte. That diagnosis was not re-verified per run. The --require gate added then aborts the anchor-not-found case before launch.

Versus the other delegate (opencode)

Same headline exit-rate, very different profile: opencode runs free tiers at zero cost, but 25 % of its runs produced nothing and 19 % timed out over the same period (vibe: 5 % and 0.2 %). Use it for cheap exploration; vibe's paid models are the choice when the edit has to land. Both write to the same log — /vibe-report --all compares them.

Prerequisites

• Mistral Vibe CLI 3.x, installed and authenticated (vibe --version; tested with 3.0.3). Vibe 2.x is no longer supported: the harness reads Vibe 3's run export and session journal.
• Claude Code with skills enabled
• python3 (pseudo-TTY via pty.spawn, stream parsing, audits) and optionally node for JS syntax checks
• timeout on your PATH (GNU coreutils; on macOS, brew install coreutils installs it as gtimeout, so add coreutils' gnubin directory to PATH). It is only a backstop: Vibe's own --time-limit stops the run first
• A git repository to work in

Installation

git clone https://github.com/pcx-wave/vibe-skill.git && cd vibe-skill && mkdir -p ~/tools ~/.claude/skills/vibe ~/.claude/skills/vibeon ~/.claude/skills/vibeoff ~/.claude/skills/vibestatus ~/.claude/skills/vibe-model-pick ~/.claude/skills/vibe-model-clear ~/.claude/skills/vibe-report && ln -sf "$(pwd)/tools/vibe-delegate" ~/tools/vibe-delegate && ln -sf "$(pwd)/tools/delegate-report" ~/tools/delegate-report && chmod +x ~/tools/vibe-delegate ~/tools/delegate-report && ln -sf "$(pwd)/SKILL.md" ~/.claude/skills/vibe/SKILL.md && ln -sf "$(pwd)/SKILL-reference.md" ~/.claude/skills/vibe/SKILL-reference.md && ln -sf "$(pwd)/VIBEON.md" ~/.claude/skills/vibeon/SKILL.md && ln -sf "$(pwd)/VIBEOFF.md" ~/.claude/skills/vibeoff/SKILL.md && ln -sf "$(pwd)/VIBESTATUS.md" ~/.claude/skills/vibestatus/SKILL.md && ln -sf "$(pwd)/VIBE-MODEL-PICK.md" ~/.claude/skills/vibe-model-pick/SKILL.md && ln -sf "$(pwd)/VIBE-MODEL-CLEAR.md" ~/.claude/skills/vibe-model-clear/SKILL.md && ln -sf "$(pwd)/VIBE-REPORT.md" ~/.claude/skills/vibe-report/SKILL.md

Step-by-step