Skip to main content

System Prompt (SOP)

The SOP is the highest-leverage AI artifact in this setup. It decides how every harness interprets a request before a skill, subagent, model, or workflow gets involved.

Read these pages as human documentation about the system's contract: where the prompt lives, what it requires, and how the surrounding skills and workflows layer on top.

The moving parts

SliceOwns
Source of truthsingle file, symlink fan-out, and update workflow
Truth and verificationcompatibility, external truth, runtime truth, completion, compact-output recovery, self-report skepticism
Execution workflowreverse interview, persistent specs, verification loops, state-machine harnesses
Side-effect gatesgit push safety, ownership, publication, bot/human split
Code-quality and dotfiles policystyle matching, semantic dedupe, docs hygiene, shell/helper rules

Using it

Start with Source of truth when changing the SOP itself. Use the other slices to understand the rule family a harness, skill, or review workflow is enforcing.

The mental model is layered:

  1. SOP sets global invariants.
  2. Skills add intent-specific procedures.
  3. Subagents isolate heavy or parallel work.
  4. Workflow pages explain the common composed flows.

Source map

SurfaceSourceTarget
Single SOP sourcehome/readonly_AGENTS.md~/AGENTS.md
Claudehome/symlink_CLAUDE.md~/CLAUDE.md
Geminihome/dot_gemini/symlink_GEMINI.md~/.gemini/GEMINI.md
Cursorhome/dot_cursor/symlink_AGENTS.md~/.cursor/AGENTS.md
Codexhome/dot_codex/symlink_AGENTS.md~/.codex/AGENTS.md
OpenCodehome/dot_config/opencode/symlink_AGENTS.md~/.config/opencode/AGENTS.md
Copilothome/private_dot_copilot/symlink_copilot-instructions.md~/.copilot/copilot-instructions.md