| name | system-prompt-archaeology |
| model | sonnet |
| description | Produces a named ADA disposition preset (YAML) and a layered behavioral analysis (structural decomposition + evidence-cited classifications) from a system prompt. Use when: "analyze this system prompt", "extract disposition from prompt", "reverse-engineer agent behavior", "what patterns does this prompt use", "build a preset from this prompt". |
| category | wisdom-garden |
| triggers | ["analyze this system prompt","extract disposition from prompt","reverse-engineer agent behavior","build a preset from this prompt"] |
| tier | 1 |
| agents | ["primary"] |
| tool_dependencies | ["file_system"] |
| inputs | [{"name":"system_prompt","type":"string","description":"The system prompt to decompose and classify into behavioral patterns","required":true}] |
| outputs | [{"name":"disposition_preset","type":"string","description":"Named ADA disposition preset (YAML) and layered behavioral analysis with structural decomposition and evidence-cited classifications"}] |
System Prompt Archaeology
I. Philosophy
System prompts are the DNA of agent behavior. Reading them carefully reveals what the
designers optimized for -- and what they sacrificed. Every constraint implies a past
failure; every encouragement implies a desired behavior that doesn't happen naturally.
Archaeology is the right metaphor: you are uncovering layers of intent deposited over
time. The earliest layers are identity and safety. Later layers add tool definitions,
behavioral nudges, output formatting. The strata tell a story about what went wrong
in production and how the team responded.
This skill does not judge prompts as good or bad. It classifies them structurally so
their patterns can be reused, adapted, or deliberately avoided.
II. When to Use
- Analyzing a competitor or peer AI tool's system prompt (from Piebald-AI/claude-code-system-prompts, user paste, or extraction)
- Building ADA disposition presets from observed agent behavior
- Comparing behavioral philosophies across tools (Cursor vs Windsurf vs Copilot)
- Designing your own system prompt by studying what works elsewhere
- Auditing your own prompts for unintentional constraint patterns
Do NOT use when:
- You need to write a system prompt from scratch (use specification-writer instead)
- You are debugging runtime agent behavior (use debugging skill instead)
- The prompt is shorter than 200 words (too thin for meaningful decomposition)
III. Workflow
Step 1: Acquire the System Prompt
Source the raw text. Preferred sources in order of reliability:
Piebald-AI/claude-code-system-prompts GitHub repo (versioned, dated)- User-provided paste from tool extraction (jailbreak, prompt leak, official docs)
- Behavioral inference from tool observation (lowest fidelity -- note this clearly)
Record: tool name, version, date acquired, acquisition method, and confidence level.
Step 2: Decompose into Structural Layers
Parse the prompt into these canonical layers (not all will be present):
| Layer | What to Look For |
|---|
| Identity | Name, role definition, persona framing ("You are...") |
| Safety/Constraints | Refusal patterns, content policy, boundary rules |
| Tool Definitions | Available tools, parameter schemas, usage instructions |
|
