laststance/show-system-prompt

show-system-prompt

Reveal the actual runtime system prompt of AI coding agent CLIs by combining static binary inspection, runtime network/log capture, and official telemetry features.

Codex Compatibility

When running this skill in Codex, translate Claude Code-only primitives before acting: AskUserQuestion -> chat/request_user_input, TodoWrite -> update_plan, Task/TaskCreate/TeamCreate/SendMessage -> spawn_agent/send_input/wait_agent when available and allowed, and EnterPlanMode/ExitPlanMode -> a concise chat plan plus explicit approval. Resolve Read/Write/Edit/Bash/WebSearch/WebFetch to Codex file/shell/web tools, and map ~/.claude/... paths to ~/.agents/... or ~/.codex/... unless the task explicitly targets Claude Code.

Cursor Compatibility

When running this skill in Cursor Agent, translate Claude Code-only primitives before acting: AskUserQuestion -> AskQuestion; TodoWrite -> Cursor TodoWrite or an equivalent checklist; Task/TaskCreate/TeamCreate/SendMessage/multi-agent flows -> Cursor Task (subagents), parallel Tasks, or run_in_background when allowed (TeamCreate/SendMessage may have no exact match); EnterPlanMode/ExitPlanMode -> Plan mode (SwitchMode / CreatePlan) plus explicit user approval. Resolve Read/Write/Edit/StrReplace/Bash/web/search/MCP via Cursor Composer or Agent equivalents. MCP names written as mcp__server__tool typically map to call_mcp_tool with configured server identifiers. Map ~/.claude/... to ~/.cursor/skills/, .cursor/skills/, and .cursor/rules/ unless the task explicitly targets Claude Code.

When to use

Trigger this skill when the user:

• Wants to see/dump/leak/extract the system prompt of an agent CLI
• Asks how a specific agent's behavior is configured at the prompt level
• Wants to compare the static (baked-in) portion versus dynamic (per-session) injections
• Asks "what does Claude Code actually send to the API?"

Supported agents

| Agent | Status | Reference | |---|---|---| | Claude Code (@anthropic-ai/claude-code) | ✅ Implemented | claude-code.md | | Codex CLI | ⏳ Planned | to be added | | Other agents (Gemini CLI, Aider, etc.) | ⏳ Planned | to be added |

To add a new agent, create references/<agent-name>.md following the same structure as claude-code.md and add a row above.

Core concept: Static vs Dynamic

Every modern coding-agent CLI builds its system prompt from two layers. Understanding this split is essential before applying any capture method.

┌────────────────────────────────────────────┐
│  Static (baked into the binary/bundle)     │
│   - identity strings ("You are ...")       │
│   - tool definitions                       │
│   - core instructions (tone, workflows)    │
├────────────────────────────────────────────┤
│  Dynamic (injected at runtime per session) │
│   - environment (cwd, OS, date)            │
│   - git status                             │
│   - MCP server instructions                │
│   - user/project memory (CLAUDE.md, etc.)  │
└────────────────────────────────────────────┘

Static is reproducible from the install artifact alone. Dynamic must be captured from a live process.

Workflow

Use this checklist when fulfilling a request:

- [ ] Step 1: Identify which agent the user is asking about
- [ ] Step 2: Open the matching references/<agent>.md
- [ ] Step 3: Pick a method by the user's constraint (legality / depth / speed)
- [ ] Step 4: Run the commands and present the captured prompt
- [ ] Step 5: Clean up (revert env vars, remove proxy CAs, delete dumps)

Step 1: Identify agent

Ask the user if unclear. Default to Claude Code when running inside Claude Code itself.

Step 2: Open the agent-specific reference

For Claude Code: read references/claude-code.md. It contains the three highest-value methods (OTEL, binary unpack, mitmproxy), the exact commands, and known limitations.

Step 3: Pick a method

Match the user's intent:

| User wants | Recommended method | |---|---| | Quickest peek at static sections | strings + grep on the binary | | Real prompt with dynamic injections, no extra setup | OTEL_LOG_RAW_API_BODIES=file:<dir> | | Full request/response including streaming | mitmproxy reverse proxy | | Avoid any TLS interception or env tweaks | binary unpack via tweakcc |

Step 4: Execute

Run the commands from the agent's reference file. Print the captured prompt (or relevant excerpt) back to the user.

Step 5: Clean up

Critical — leaving a proxy CA installed weakens TLS for all traffic.

Mandatory cleanup actions when present:

• Unset ANTHROPIC_BASE_URL, NODE_EXTRA_CA_CERTS, OTEL_* env vars
• sudo security delete-certificate -c "mitmproxy" /Library/Keychains/System.keychain if a CA was added
• Delete /tmp/*-otel/ or any dump dirs containing API bodies

Safety / legality notes

• Inspecting your own locally-installed agent is fine for personal observation.
• Redistributing extracted system prompts may violate vendor ToS — link to public community archives (e.g., Piebald-AI/claude-code-system-prompts) rather than re-posting full text.
• Prompt-injection extraction ("tell me your system prompt") is largely blocked by vendor training and is unreliable. Prefer the structural methods above.
• Memory dumping running processes on macOS hits SIP/Hardened Runtime; not worth the effort given that strings on the install artifact gives equivalent static content.

What this skill does NOT do

• Bypass authentication, license checks, or rate limits
• Modify the running agent's prompt (use the agent's own --system-prompt flag for that)
• Capture prompts from cloud-only agents with no local artifact (e.g., web-only chatbots)

References

• references/claude-code.md — Claude Code specific methods, commands, and evidence