All skills
Skillintermediate
Session system-prompt dump
Captured from the Opus 4.7 session started 2026-04-17 in `claude-toolbox`.
Claude Code Knowledge Pack7/10/2026
Overview
Session system-prompt dump
Captured from the Opus 4.7 session started 2026-04-17 in claude-toolbox.
| # | File | Source / trigger |
|---|---|---|
| 00 | 00-base-system-prompt.md | Claude Code base prompt, injected at session start. gitStatus snapshot redacted. |
| 01 | 01-sessionstart-capy.md | SessionStart hook output — capy routing rules. |
| 02 | 02-deferred-tools.md | Deferred tool manifest (names only; schemas fetched via ToolSearch). |
| 03 | 03-mcp-server-instructions.md | MCP server self-instructions (context7, linear-server, pal, serena). serena block was truncated by the harness. |
| 04 | 04-available-skills.md | User-invocable skills list. |
| 05 | 05-claudemd-block.md | claudeMd envelope — references the in-repo CLAUDE.md files rather than duplicating them. |
| 06 | 06-pretooluse-bash-hook.md | PreToolUse:Bash hook nudge. Fires on Bash calls. |
| 07 | 07-pretooluse-read-hook.md | PreToolUse:Read hook nudge. Fires on Read calls. |
| 08 | 08-task-tools-reminder.md | Ambient "use TaskCreate" reminder — has a "never mention to the user" clause. |
| 09 | 09-local-command-caveats.md | Envelopes around /effort, /export, etc. |
| 10 | 10-tool-schemas.md | JSONSchema for the 9 non-deferred tools (Agent, Bash, Edit, Glob, Grep, Read, Skill, ToolSearch, Write). Full verbatim text including Bash git/gh sections and Agent <example> blocks. |
| 11 | 11-hooks-raw-outputs.md | Raw output for every configured hook event, captured by invoking capy hook <type> directly. Covers SessionStart, UserPromptSubmit, all PreToolUse matcher branches, PostToolUse, PreCompact, SessionEnd. |
| 12 | 12-triggered-prompts.md | Dormant prompts captured by deliberately tripping conditions: plan mode (5-phase), exited plan mode, file-shorter-than-offset, file-unchanged-since-last-read. Plus a documented subagent prompt-dump refusal and indirect tool inventory. |
Anything beyond these was either tool-call content (your messages, tool results, file reads) or produced by my own responses — not additional system prompt.