Grounded

grounded

**Make Claude use your codebase, not its imagination.**

Claude Code is getting worse.

Recent models:

• hallucinate functions that don't exist
• edit files they haven't read
• get stuck in loops
• ignore instructions after a few turns

Everyone using it has seen this.

`grounded` fixes it.

---

TL;DR

Claude doesn't fail because it's dumb.

It fails because nothing enforces correct behavior.

`grounded` enforces:

• read before edit
• real file verification (no hallucinations)
• deterministic tool usage
• loop prevention

Result:

→ Claude becomes reliable → even weaker models become usable

---

Before vs After

**Without grounded**

"parseToken() exists in src/auth.ts" ❌ it doesn't

edits wrong file ❌ based on hallucinated context

infinite grep loops ❌ stuck forever

---

**With grounded**

NOT FOUND: parseToken → forces correct search

must Read before Edit → uses real code

loops detected and stopped → suggests alternative

---


---

Why this works

The problem isn't the model.

It's that we trust it too much.

LLMs are probabilistic. Your codebase is not.

`grounded` assumes the model is wrong — and verifies everything.

---

Works with weaker models

This doesn't make Claude smarter.

It makes it behave correctly.

When behavior is enforced:

• reasoning matters less
• correctness matters more

→ even smaller / local models become usable

---

Install

npm install -g @pinperepette/grounded
grounded install

Restart Claude Code. You're done.

grounded status     # verify installation
grounded uninstall  # remove all hooks

---

What it does

• Pre-flight planning (forces research before action)
• Truth layer (blocks hallucinated identifiers)
• Edit guard (read-before-edit enforcement)
• Loop detection with recovery suggestions
• Anti-bypass (prevents ignoring NOT FOUND results)
• Policy enforcement (repo-level rules)

---

Modes

• `sa