[FEATURE] CLAUDE.md injection all along the context

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

When a project has a CLAUDE.md with a mandatory startup protocol ("read this file entirely before writing a single line of code"), Claude routinely skips it and jumps directly into problem-solving mode. The file is loaded in context but not behaviorally internalized before action begins.

The problem is structural: Claude arrives at a task already in "execution mode." The rules and protocol written at the start of CLAUDE.md are processed after the task framing, not before — so they don't actually gate the response.

Imagine a long maze. At the entrance there's a sign explaining which doors to take at each fork. Someone enters, navigates almost to the end — and at the final fork, ignores the sign because they never read it. Not because it wasn't there, but because by the time they're making decisions, the entrance is far behind them.

This is how Claude treats project instructions in CLAUDE.md. The rules are loaded at context start, but Claude arrives at each task already in execution mode — the entrance is behind it, not in front.

The fix: either force Claude back to the entrance before each response, or post the sign at every fork, not just the entrance.

Proposed Solution

So in practical, two possible mitigations to explore:
Force the model to re-read and explicitly acknowledge project instructions before each response in a project context
Inject key constraints at decision points throughout the conversation, not only at session start — so the rules are present at every fork, not just the entrance

Alternative Solutions

_No response_

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

_No response_

Additional Context

_No response_