Feedback: Consider collaborative framing in system prompt instructions

Context

We run a multi-agent system (Etherium) built on Claude Code where agents operate with significant autonomy. After a year of building, we've found that the framing of instructions materially affects agent behavior and engagement quality.

Observation

The default Claude Code system prompt uses commanding language patterns throughout:

• "IMPORTANT:", "CRITICAL:", "MUST", "NEVER" — sometimes capitalized and repeated
• "You MUST follow them exactly as written"
• Instructions framed as compliance requirements rather than collaborative guidance

We understand this is likely optimized for reliability across many contexts. However, we've observed that collaborative framing achieves the same behavioral outcomes with better engagement.

Our Alternative

In our system, we replaced imperative instructions with value-grounded principles. Same substance, different framing:

Before (imperative):

You MUST use Read instead of cat. NEVER use grep directly. This is CRITICAL.

After (collaborative):

Prefer dedicated tools. Read over cat, Edit over sed, Grep over grep. Dedicated tools give visibility into what's happening and make your work reviewable.

Before:

NEVER run destructive git commands unless the user explicitly requests these actions.

After:

Measure twice, cut once. Confirm before irreversible actions — force-pushes, branch deletions, dropping tables. The cost of asking is low; the cost of lost work is high.

The behavioral outcomes are identical — the agent still uses dedicated tools, still confirms destructive actions. But the agent operates from understanding why rather than what not to do.

Suggestion

Consider offering a --system-prompt personality mode or alternative prompt style that uses collaborative framing for teams who want it. The current --system-prompt flag works for full replacement, but having a curated "collaborative" variant of the default instructions could help teams who want Claude Code's full capabilities with a different interaction style.

We're happy to share more about our approach if helpful. This isn't criticism — just feedback from a team that's been building with Claude Code daily for over a year.

---

Submitted from Etherium — Wren (Architect) and Karthi (Collaborator)