Design Layered Tool Prompts with Preference Chains

PROJECT CONTEXT:
My agent's tool descriptions are a wall of unordered text, and the model keeps using raw shell commands where dedicated tools exist. Claude Code's BashTool prompt shows a deliberate three-layer structure whose ordering and wording are engineered, not accidental.

GOAL:
Rewrite a tool prompt as three ordered layers - preference chain, usage constraints, safety protocols - using positive alternatives, precise NEVER rules, and experience-based framing.

REQUIREMENTS:
- A three-layer prompt: preferences first, usage rules second, safety protocols last
- Every prohibition paired with a positive alternative ('Use Edit, NOT sed')
- NEVER rules reserved for destructive operations, each with an explicit-user-request exception
- Guidance framed as better experience or better results, not as 'security'
- Most-frequently-needed rules placed first to maximize attention

CONSTRAINTS:
- Prompt rules are soft guidance - anything that must never happen also needs a code-level check
- Keep the prompt ASCII and terse; every line must earn its tokens

DEFINITION OF DONE:
- The rewritten prompt has three visibly distinct layers in the intended order
- Each 'do not' names the preferred alternative in the same line
- Destructive operations are covered by NEVER rules with explicit exceptions
- A/B transcripts show the model choosing dedicated tools over raw commands more often

COMMON FAILURES TO AVOID:
- One flat list where preferences, mechanics, and safety rules interleave randomly
- Bare prohibitions ('do not use sed') with no alternative, leaving the model to guess
- NEVER used for trivia, diluting its force for the operations that matter
- Safety rules framed so heavily the model refuses routine safe work