Practical vibe coding — the middle way

Neither over-plan nor yolo. Iterate feature-by-feature with an anchoring AGENTS.md, focused ICC prompts, behavior-constraints, and verify each before starting next. The umbrella framework.

Practical vibe coding

The middle way

Two failure modes at the extremes:

• Pure vibe coding — yolo the prompts, don’t read the code. Features 1-3 fly; feature 5 breaks something and you can’t tell what. Speed without direction.
• Over-engineering — 2 days of planning, folder conventions, and architecture docs before a single screen exists. Feels productive; ships nothing.

Practical vibe coding sits in the middle. Keep the speed of AI, add just enough structure so:

1. AI knows what it’s working on (AGENTS.md at repo root, always read first).
2. You don’t lose track (feature-by-feature, verify each).
3. Failures are localized (one change per prompt).

The four anchors

1. AGENTS.md at repo root (living doc)

Written once, updated as recurring issues surface. Every prompt starts with “read AGENTS.md first and follow it strictly.”

See agents-md-living-doc for the update discipline.

2. 4-part prompt structure

Every non-trivial prompt has these four parts, in this order:

1. Read AGENTS.md — “Read the AGENTS.md first and follow it strictly.”
2. One task — exactly one feature, one screen, one integration. Not three merged.
3. Constraints that protect what works — behavior constraints, not code prescriptions. See icc-prompt-formula §“Constraints describe behavior.”
4. Design references — screenshots, images, docs pasted inline. AI reads layouts better than descriptions of layouts.

3. Feature isolation

One prompt = one feature. Never merge tasks like “set up NativeWind AND build onboarding AND add auth.” If something breaks across 3 tasks, you can’t tell which caused it.

4. Verify-then-move-on

After each feature: test on real device/simulator, review the diff, then move on. Don’t stack unfinished features.

See session-hygiene for when to reset chat between features.

When to escalate to over-planning

Reach for full architecture docs ONLY when:

• ≥5-file cross-cutting refactor
• Security/auth boundary redesign
• Cost-sensitive infra decision (paid API, hosting migration)

Everything else = practical vibe coding.

When to escalate to pure yolo

Never. Even a 5-line change gets an AGENTS.md read.

Anti-patterns

• ❌ “Let me plan the whole app first” → 2 days spent, nothing built
• ❌ “Just ship it” → 5 features later, spaghetti code you can’t explain
• ❌ Merging 3 features into one prompt to “save time” → save 2 min, lose 30 min on debugging
• ❌ Skipping the AGENTS.md read at prompt start → stale conventions leak through
• ❌ Perfectionist AGENTS.md upfront → block on writing; ship, update as you learn

Composition with existing rules

Cross-refs

• Source: JavaScript Mastery YouTube “Practical Vibe Coding” walkthrough 2026-07
• Andrej Karpathy 2026-01 “vibe coding” tweet — the term this rule refines

Edit on GitHub · Back to index