Agent Desk — Development skill for Claude Code

---

status: verified created: 2026-10-17

agent-desk

A one-human operation desk for any harness that lets you define a primary agent or agents file: opencode, Claude Code, Cursor, Windsurf and similar. File-board spine, two-tier worker dispatch, verbatim briefs, independent validation, executable harness checks, operator-gated irreversibles. The report behind it is in `docs/DOCTRINE.md`.

What you get

• agent-desk.md — the desk prompt. The frontmatter is opencode's adapter format (mode: primary, tools: task); other harnesses use their own (Claude Code: .claude/agents/, Cursor: .cursor/rules/, AGENTS.md convention: paste the body). The six loop steps and the constraint block are harness-neutral.
• docs/DOCTRINE.md — the 15 laws and the "what humans keep forever" limits.
• docs/HARDENING.md — how to verify the rules your config really enforces (real-binary hermetic checks).
• docs/AUDIT.md — a case study: how the above was measured on a real setup.
• LICENSE — MIT.

How to use it

1. Create ops// beside your project. Put the job cards (job-NNN.md) and LEDGER.md there. Add it to .gitignore if it's private.
2. Start the agent with the agent-desk prompt.
3. State the mission in one line. Mark every irreversible action explicitly.
4. Review the board (ops//job-*.md), sign off on the fixes you want. Nothing mutates outside the board until you say so.

What the operator must decide

1. Whether the desk needs to see your secrets at all. The desk runs on least privilege: deny the config and *.env patterns explicitly (see docs/HARDENING.md).
2. Whether subagents inherit your shell environment at all. The tool passes process.env to bash, so every secret variable you export is visible to every worker by default. The mitigation is process isolation, not a permission rule.
3. How to install and sign skills. Untrusted skills have shipped malicious hooks; one SKILL.md can run a command.