agent-anatomylisted

# agent-anatomy Organize any AI agent's project into a **filesystem-first** layout: one agent = one folder, every capability = a file in a conventionally-named subfolder. Language-agnostic — tools can be Python, TypeScript, anything. **Core principle: names come from file paths, not config.** `tools/issue_refund.py` *is* the tool `issue_refund`. Never add a `name:` field to restate the path. ## The convention `instructions.md` (the always-on system prompt) is the one **required** file. Everything else is an optional, canonically-named subfolder. Classify each file: | A file that… | goes in | notes | |---|---|---| | is the system prompt / persona / instructions | **`instructions.md`** | rename `prompt.txt`, `system.md`, etc. to this | | does a discrete action the agent invokes | `tools/` | filename = tool name | | is a markdown procedure followed on demand | `skills/` | e.g. `how-to-escalate.md` | | runs on a timer / cron / "every morning" | `schedules/` | | | receives inbound messages (webhook, bot) | `channels/` | Slack, HTTP, Discord | | calls an outbound third-party API / service | `connections/` | external APIs, MCP servers | | is a nested agent | `subagents/<name>/` | recurse — same convention | | is durable state the agent reads back | `memory/` | | | is shared code imported by tools/jobs | `lib/` | utils/formatters — not a capability | | is dev scratch / TODOs / personal notes | leave in place | flag it, don't move | ## Two modes — detect first **Empty or missin