- Verified user-pasted tool prose (video_gen full 'product demos, animated illustrations, looping ambient...' block + concise web_search/web_fetch/open_page/open_page_with_find + x_keyword_search etc. + scheduler_create with fireImmediately/interval 5m-1d/max-50/7-day-expiry) was already present verbatim in the prompt text for those tools.
- Fixed critical gap: replaced generic Skills section with ONLY the exact verbatim SKILL.md content from the 4 user-mandated paths (/Users/asgeirtj/.grok/skills/help, /check, /best-of-n and /Users/asgeirtj/.grok/bundled/skills design/implement/pr-babysit/review). No other skills, no annotations, no expansions, no parentheticals.
- File grew from 501 to 3761 lines due to full skill descriptions (as required for self-contained raw dump).
- All prior behavioral sections (task_completion_discipline with todo/reseed/end-of-turn gates, making_code_changes, tone_and_style, output_efficiency, formatting with startLine:endLine:filepath rule, inline_line_numbers, project_instructions_spec, memory system) and full long-prose tools (run_terminal_command with background/timeout rules, read_file, search_replace 'MUST read first + unique string' rules, todo_write, spawn_*, scheduler_*, monitor, image/video_gen, MCP search_tool/use_tool etc.) remain intact and pure.
Local zero-state maintained. (git push may 500 transiently.)
Per CLAUDE.md + repeated user purity rules: only verbatim org raw system prompt content from the agent's actual prompt + the 4 specified skill paths.