Part of Mechanisms of Vibe Coding

Mechanism: Negative Space

One-sentence definition

Negative space is the practice of naming what must stay unchanged — files, behaviors, APIs, and styles — so the agent’s freedom is shaped by explicit exclusions, not by whatever the model assumes is “fair game.”

The problem

You ask for a UI tweak. The agent “helpfully” updates the shared button component used in forty screens. You asked for one page; you got a design system migration. The missing instruction was not what to do — it was what not to do.

Positive prompts (“fix the health check”) leave huge default permission. Negative space fills the gaps: do not change shared components, do not alter public API signatures, do not migrate state management.

Symptoms:

• Shared modules change when the task was local
• Public APIs gain breaking fields “for consistency”
• Generated refactors in files you thought were frozen
• Reviews become debates about unstated assumptions

How it works

Negative space complements constraints. Constraints are the full guardrail set; negative space emphasizes preservation:

1. Frozen paths — src/legacy/ read-only; packages/ui/Button.tsx off limits
2. Frozen behavior — existing REST responses must stay byte-compatible
3. Frozen process — no dependency upgrades, no license header sweeps
4. Frozen style — match existing patterns; no new abstractions

  Positive goal (what to achieve)
       +
  Negative space (what must not change)
       │
       ▼
  Smaller, safer solution space

List negatives in plain language near the goal. The agent sees what is sacred, not just what is assigned.

When to use it

• Legacy modules, shared libraries, and generated code
• Hot paths where regression cost is high
• Any task adjacent to code you cannot afford to churn
• After an agent “helpfully” refactors something frozen once — add the negative line permanently to rules

When not to use it

• Greenfield files with no dependents — negatives add little
• When the task is the migration — then the goal must say migrate X, not smuggle it via silence
• Replacing a goal with only “don’t touch anything” — you need a positive outcome too

Failure modes

Only negatives — Wall of “don’t” with no achievable goal. Fix: Pair with goal symptom + verify.

Vague negative — “Don’t break anything.” Fix: Name files, APIs, or tests that must stay green.

Stale negatives — Frozen folder later becomes active work. Fix: Update rules when ownership changes.

Contradiction — Goal requires API change; negative says API frozen. Fix: Resolve in planning; one must give.

Hidden negatives — Team knows Button.tsx is sacred; agent does not. Fix: Put standing negatives in project rules.

Minimal example

Context: Health endpoint fix must not touch shared middleware or Docker.

Steps:

1. Extend the prompt block:

   Goal: GET /health returns { "ok": true } after startup. [verify line]

   Negative space (do not change):
   - src/middleware/** (shared across routes)
   - Dockerfile, docker-compose*, CI workflows
   - Public API routes other than /health response body
   - Package.json / lockfile

1. If the agent proposes middleware changes, reject: “Violates negative space — find a fix in bootstrap or health only.”

1. Add recurring negatives to .cursor/rules if this keeps happening.

Done when: Verify passes and frozen paths have zero diff hunks.

Tool instances (optional deep-dive)

Portable idea above; this section is tool-specific. Date: June 2026.

Cursor

• Negative space in rules — standing list of frozen paths for the repo.
• .cursorignore — hard exclude for paths that must never be edited by agent indexing.
• Diff review — reject any hunk touching negative-space paths before checkpoint.
• Distinction from constraints: constraints include do-not actions globally; negative space names specific preserved surfaces.

Other tools

CODEOWNERS on frozen paths, linter no-edit zones, Aider CONVENTIONS.md “sacred files” section — portable pattern.

Related mechanisms

• Constraints / guardrails — broader limit set including negatives
• Scope fence — session boundary; negative space defines sacred interior
• Goal — positive target negatives protect while reaching

Try it yourself

Exercise: List three frozen items for your next agent task — one path, one behavior, one process rule. Add them under a Negative space: header in the prompt.

You need: One upcoming task; two minutes

Done when: A teammate reading only your prompt would know what is off limits without oral tradition.