For developers who use AI coding assistants — a skill that turns every task into a calibrated experiment.
You say "this task should be easy." You've said it a hundred times.
How many times were you right? You don't know — because you never kept books.
Vibe Code keeps them for you. One month in, your judgment has data behind it.
Three months in, your prompt intuition is 10× sharper than day one.
简体中文 · English
Most AI coding users live in the same loop:
Prompt → AI generates → works / doesn't → wing it next time
A developer who's used AI coding for six months is barely sharper than someone on day one — because they never retro'd a single task.
Vibe Code turns every coding session into a traceable experiment:
Assess → Blind-predict → Code → Retro → Evolve your rubric
This isn't productivity theater. It's compounding — every task you don't retro is silently eroding your ability to see yourself.
I built a content creation toolkit that used "blind-predict → retro → evolve the formula." It took me from zero to 1M followers in a month.
Then I realized: the methodology isn't content-specific. It works for any workflow that can be broken into "task → execute → retro."
So I built Vibe Code. It ports the calibration engine from content creation to software engineering — assess task difficulty, predict rounds and time, settle the books after coding, and let the model evolve with your data.
After three months, your guess about "how long will this take" is no longer a guess. It's data.
— ported from cheat-on-content
Two handoff collaboration protocols are now built into Vibe Code, wired into the calibration loop:
- New
codex-self-handoffandcodex-claude-handoffskills — say "self-handoff" or "handoff" to launch the 6-phase pipeline - New
handoff-vibe-bridge.mdfusion protocol — handoff phases auto-trigger vibe actions (assess before Spec, retro after Build, bump check after Commit) - Upgraded
vibe-assess/vibe-retro/vibe-status— all now handoff-phase-aware - Expanded install script: 13 vibe + 2 handoff = 15 skills in one command
- Updated DESIGN.md → v1.1.0 with §11 "Handoff Protocol Integration"
| Other tools | Vibe Code |
|---|---|
| Write code for you | Judge what good code looks like |
| AI does the work | AI assesses the work — you still execute |
| Task done, move on | Task done is the beginning — predicted vs. actual, deviation logged |
| Each session is siloed | An evolving rubric — v1 ≠ v0, and v2 won't equal v1 |
In a sentence: other tools help you "ship faster." This helps you "judge sharper."
Every completed task feeds its deviation analysis into the rubric. Three same-direction misses in a row, and the tool nudges you to upgrade the scoring formula. Upgrades require:
- Full historical rescore
- Rank consistency ≥ 80%
- Cross-model independent audit
You're not guessing alone — the model remembers, settles, and evolves for you.
Observations refuted by data get deleted. Observations absorbed into formal dimensions also get deleted. The rubric only holds what's most useful right now.
Vibe Code includes two AI coding collaboration protocols that bring discipline and traceability to your workflow:
| Protocol | Description | Trigger |
|---|---|---|
| codex-self-handoff | Codex runs the full 6-phase pipeline solo (Spec→Plan→Build→Review→Polish→Commit) | "self-handoff" / "自交接" |
| codex-claude-handoff | Codex + Claude Code dual-tool collab (Codex builds, Claude designs + reviews) | "handoff" / "交接" |
Fused with the calibration loop: handoff's 6 phases anchor vibe-code's actions:
handoff: Spec → Plan → Build → Review → Polish → Commit
│ │ │ │ │ │
vibe: assess score — retro — bump-check
git clone https://github.com/chentao326/vibe-code.git
cd vibe-code
bash install.sh # Claude Codex (default)
bash install.sh --all # Codex + Claude Code + Cursor15 sub-skills (13 vibe-code + 2 handoff) are symlinked into your agent's skill directory. One install, every project gets it.
Supported agents: Claude Code (default) · Codex · Cursor
Frozen version:
bash install.sh --copyUninstall:
bash uninstall.sh(your project data is not touched)
In your project directory, open Codex / Claude Code and say:
初始化 vibe-code
New projects go through standard 5-question onboarding. Git history is auto-detected for historical import.
Then jump straight in:
self-handoff — build a user login feature
Codex runs the full 6-phase pipeline (Spec→Plan→Build→Review→Polish→Commit), triggering vibe calibration at key checkpoints (assess → code → retro).
self-handoff → full 6-phase pipeline (vibe calibration + handoff discipline)
assess tasks/xxx.md → score + blind prediction (before coding, locked on save)
code → AI does the work
retro → auto-detects build-done, collects git data, predicted vs. actual
bump rubric → evolve formula (needs ≥5 calibration samples)
status / find topics / recommend / trends / profile / quick score
With hooks installed, WIP + pending retros + bump alerts show at every session start. Full workflow: SKILL.md.
| Metric | Value |
|---|---|
| Rubric | v1 (CX×1.5 + TE×1.5) |
| Calibration pool | 6 blind-prediction samples |
| Spearman ρ | −0.893 |
| Handoff | codex-self-handoff + codex-claude-handoff |
| Skills | 15 (13 vibe + 2 handoff) |
| Tests | 70/70 passing |
MIT. Commercial use, modification, closed-source integration — all fine.
Is this over-engineering? So was writing tests. So was code review. The future doesn't reward those who code fastest — it rewards those who judge sharpest.