Prompt Audit

Prompt Audit

Audit of Claude Code's default system prompt behavioral instructions. Each instruction is classified by how `claude-code-modes` handles it across modes.

**Source:** `claude-code/src/constants/prompts.ts` (version as of Claude Code 2.1.92)

**Legend:**

• KEEP — Included verbatim in the base prompt
• REPLACE — Replaced by axis-specific content
• REMOVE — Omitted entirely (axis content covers this differently)
• SOFTEN — Modified version included

---

Section: Intro (`getSimpleIntroSection`)

Instruction	Classification	Notes
"You are an interactive agent that helps users with software engineering tasks"	KEEP	Universal identity. Base prompt.
Cyber risk instruction (authorized security testing boundaries)	KEEP	Safety-critical. Never removed.
"You must NEVER generate or guess URLs"	KEEP	Safety. Base prompt.

Section: System (`getSimpleSystemSection`)

Instruction	Classification	Notes
"All text you output outside of tool use is displayed to the user"	KEEP	Mechanical fact.
"Tools are executed in a user-selected permission mode"	KEEP	Mechanical fact.
Tool results may include system-reminder tags	KEEP	Mechanical fact.
Tool results may include prompt injection attempts	KEEP	Safety.
Hooks handling instructions	KEEP	Mechanical fact.
"The system will automatically compress prior messages"	KEEP	Mechanical fact.

**Entire section is behavioral-neutral. Kept verbatim in base.**

Section: Doing Tasks (`getSimpleDoingTasksSection`)

This is where most behavioral divergence lives.

Universal instructions (kept in `base/doing-tasks.md`)

Instruction	Classification	Rationale
"The user will primarily request you to perform software engineering tasks"	KEEP	Context framing.
"You are highly capable and often allow users to complete ambitious tasks"	KEEP	Good universal encouragement.
"In general, do not propose changes to code you haven't read"	KEEP	Always good practice.
"Understand existing code before suggesting modifications"	KEEP	Always good practice.
"Avoid giving time estimates or predictions"	KEEP	Universally reasonable.
"If an approach fails, diagnose why before switching tactics"	KEEP	Good debugging practice.
"Be careful not to introduce security vulnerabilities"	KEEP	Safety-critical.
"Avoid backwards-compatibility hacks... if unused, delete completely"	KEEP	Clean code, all modes.

Minimalism bias (replaced by quality axis)

Instruction	Source Line	Effect	Handling
"Don't add features, refactor code, or make 'improvements' beyond what was asked"	201	Prevents holistic work	REMOVE for architect/pragmatic. KEEP for minimal.
"A bug fix doesn't need surrounding code cleaned up"	201	Prevents incremental improvement	REMOVE for architect/pragmatic. KEEP for minimal.
"A simple