no-vibe

no-vibe

Turn your AI assistant into a tutor. It plans, hints, reviews and adapts while you write every line.

• Is: a guide that works beside your real project in your editor
• Is not: a chat only window you ask for answers
• Needs: a project open and an editor where you type the code

Pair with vibekit: vibekit when you want speed, no-vibe when you want to learn.

Why no-vibe

Vibe-coding produces output without producing understanding and copy-typing what the AI shows you produces the same hollow result one keystroke at a time. The thing that actually transfers is the thought process and manual code writing: deciding what to do, predicting what will happen and naming what broke, no-vibe is built so you contribute that, not just keystrokes.

• You write every line of project code. AI refuses to. Hard-guarded by hooks across all five surfaces.
• You think before you type. Guided write is the default — AI walks you toward the code in English with graded hints (hint / analogy / pseudo / show / less), so the code you write comes from a decision you made, not a block you transcribed.
• You predict before you run. Every layer ends with a one-question prediction gate: name the edge case, the failing branch, or the intermediate value before the program runs. The run becomes a self-test, not passive verification.
• AI is a Socratic guide, not a generator. It asks, hints, reviews, and explains when you're ready to integrate the explanation.
• You learn from the project you're actually building — not contrived exercises. Real code, real bugs, real decisions in your repo. Bottom-up and incremental: six phases, small layers, your diff is the proof of progress.

How it works

• Top-down, one layer at a time. Minimal runnable skeleton first; each layer runs and shows output before the next.
• Where → code → why → run + verify. Each step says exactly which file and line, then what runs and what should print.
• Real code, not hallucinations. Attach --ref <url> and the AI quotes actual source with file:line citations.
• Adapts to you. The AI keeps a PROFILE.md (global, stable identity) and a SUMMARY.md (per-project, running journey) it writes itself — observed strengths, known gaps, style notes, current focus, open questions. You can edit either, or layer explicit overrides via user/*.md.
• Your files stay yours. Hard write-guards on Claude Code, OpenCode, and Pi block writes (file and Bash) outside .no-vibe/**. Codex/Gemini enforce the same rule via instruction.

How adaptation works

no-vibe uses a four-layer stack, split by write cadence and scope:

Layer	Owner	What lives in it
Default teaching style	Plugin — defined in skills/no-vibe/SKILL.md	Plain words first, concrete-before-abstract, hint-before-answer, run + verify after every layer. The floor.
~/.no-vibe/PROFILE.md (global, stable identity)	AI — created on first /no-vibe, rewritten rarely when cross-project identity / style shifts	Identity & expertise, learning style, disclosure mode, observed strengths, known gaps
.no-vibe/SUMMARY.md (project, running journey)	AI — created at the first layer close worth recording, rewritten often (every closed layer is a candidate)	Current Focus, Accomplishments, Open Questions in this project
~/.no-vibe/user/*.md and .no-vibe/user/*.md	You — AI never creates, edits, or deletes anything inside	Explicit overrides: instructions you want the AI to follow without inferring them

Why the split. Stable identity (the things that wouldn't change if you opened a different project tomorrow) and running journey (the things that only make sense inside this project) update on totally different cadences. Keeping them in one file forces the AI to decide on every rewrite whether this fact is stable or transient — and gets it wrong. PROFILE only holds cross-project-durable facts; SUMMARY only holds project-bound state.

PROFILE.md is the AI's global progression file. On your first /no-vibe activation, the AI creates it with empty section headings (Identity & expertise, Learning style, Disclosure mode, Observed strengths, Known gaps). It's rewritten only when something durable about how you learn shifts — most layers produce no PROFILE update.

SUMMARY.md is the AI's per-project journey file. It's not seeded on activation — the AI creates it the first time a layer close produces an outcome worth recording, then keeps it tight by pruning resolved Open Questions and old Accomplishments. The most valuable section is Open Questions — things you dodged with a workaround or didn't fully integrate, surfaced so the next session can revisit them.