A one-person AI team, run from a single line of text

What it does

solo-to-team-skills is a set of four agent skills plus templates and sanitized case notes, written in Chinese, for managing multiple long-lived AI assistants as a team. The core loop is deliberately boring: dispatch a task (log a “pending” line naming who does it, who receives it, who checks on it), receive the result, verify evidence, and only then mark it accepted. The four skills cover dispatch/handoff (dispatch-accept-verify), evidence-before-declaring-done (evidence-before-done), writing beginner-proof step-by-step instructions (beginner-ops-writing), and reusable lessons from real incidents (ops-case-notes).

The interesting bit

The pack is host-agnostic by design: the skill bodies are shared across eight agent tools (Codex, Cursor, WorkBuddy, DeepSeek Harness, CodeBuddy, Kimi Work, QwenWork, Grok Bot), so you learn the method once and swap hosts without relearning it. It also ships zero automation — hooks: {} — meaning timers, reminders, and cross-assistant messaging are explicitly your job or the host’s, never the skill’s. The README is unusually honest about this, and about what its verification does and does not prove.

Key highlights

• A task register line (one line per task) is the whole tracking system — visible in a notepad, a spreadsheet, or a kanban.
• “Done” requires evidence matching the current version; changed or previously-failed parts get re-verified, unchanged parts can reuse prior checks.
• Two sanitized real-world case notes document what went wrong, the fix, the result, and what’s still unsolved.
• Verified on eight hosts with a consistent standard: obtain → install → invoke by name → produce a checkable register line; one limited-task acceptance record per host.
• The README explicitly forbids citing the “about ten minutes” walkthrough as a tested guarantee (the actual CLI timing was 29.6 seconds, under pre-prepared conditions).

Caveats

• v0.9-draft, explicitly not final; the author says it will change before official release.
• Verification is one limited task per host — the README itself states this doesn’t demonstrate stability, doesn’t mean all four skills were invoked everywhere, and some hosts didn’t retain skill-loading traces.
• Claude Code support is structurally prepared but not yet walked through; Grok Bot needs manual configuration.
• Chinese-language content throughout (an English README exists); the value is the methodology, not code — this is process glue, and proudly so.

Verdict

Worth a look if you’re already juggling multiple AI assistants and want a lightweight, human-in-the-loop accountability method that survives tool switches. Skip it if you want turnkey automation — the pack deliberately refuses to be that.