solo-to-team-skills agent skills: rules for solo AI teams
A week-old MIT repo packages four agent skills for dispatching work, demanding evidence before "done" and keeping a register of what was actually verified.
A week-old repo for one person running a team of agents
On 2026-11-14 the GitHub repository dabingfy-dot/solo-to-team-skills showed 853 stars and 90 forks, seven days after it was created on 2026-11-07. It ships four solo-to-team-skills agent skills, one task register template, install notes for each assistant tool and two anonymised case notes, under the MIT licence. The README marks the current state as v0.9-draft and says it will change before a formal public release. The project is written in Chinese with an English README, credits 大冰 as its initiator, and includes a 17-page PDF report in its docs folder.
What the four skills cover
The bundle targets a narrow problem: one person coordinating several long-running AI assistants, where work gets handed out and then quietly forgotten. The README describes one line of flow: set the goal, split the work, hand it over, check it, decide.
dispatch-accept-verifyregisters a task before it goes out (who is doing it, who collects the result, who checks at the deadline), then logs receipt and acceptance as separate events.evidence-before-donerequires evidence that matches the current version of the work before anyone says it is finished. Parts that did not change can reuse an earlier check.beginner-ops-writingis original to this project and covers how to write steps a newcomer can follow once, one step at a time, with a clear statement of what they should see when done.ops-case-notesis also original: two desensitised operations cases written up as problem, fix, result and what remains unsolved.
The register line is the unit of work. A task is dispatched as pending receipt, moves to result received with a pointer to where the output lives, and only leaves the open list once it is accepted or cancelled with authorisation. The first two skills are described as adapted from public writing on closed-loop communication and from obra/superpowers (MIT), with source notices kept in the skill files.
Where it runs
Eight tools carry a supported label: Codex, WorkBuddy, DeepSeek Harness, Cursor, CodeBuddy, Kimi Work and QwenWork, plus Grok Bot, which is supported only with manual configuration, saving a SKILL.md as a private skill inside a conversation. Claude Code has plugin structure prepared but is not in the first-release support list. For most hosts installation is a directory copy into a per-tool skills path, and the README warns that WorkBuddy is not CodeBuddy, QwenWork is not Tongyi Lingma, and Kimi Work is not the Kimi Code CLI.
One timing figure is given. On Cursor CLI, non-interactive with the prompt prepared in advance, the run took about 0.5 minutes (29.6 seconds) from copying the skills into ~/.cursor/skills/ to printing a readable register line. That excludes installing and registering the host, and typing time, and the README explicitly refuses to present "about ten minutes" as a measured guarantee.
Limits and open questions
The plugin manifest sets hooks to an empty object, so the package has no scheduling or wake-up of its own. Deadline checks, reminders and start signals come from the host tool or from the person. The verification record is thin in the ways the README itself admits: eight hosts each have one limited-task acceptance record, several passed only after blocked reruns or revisions, not every host exercised all four skills, and some kept no raw trace of skill loading. There is no controlled experiment, so the README makes no claim about lower error rates or fewer follow-ups. The walkthrough example is a teaching sample, not a test log. Risky actions such as spending, sending, deleting, login and payment confirmation are left to the human, and copying a SKILL.md into some other tool's skills folder does not count as verification there. Tongyi Lingma, TRAE and Comate are listed as untested.
What it means for building apps with AI
The transferable part is the bookkeeping, not the install paths. Anyone building an app with an AI assistant meets the same gap this repo targets: a model can report a feature as finished without evidence. Draftship, the AI app builder this site covers, writes and hosts apps from a plain-language description, with a built-in database, user sign-in and hosting, and paid plans adding private projects, a choice of model, source code access, custom domains and GitHub push. Nothing here suggests Draftship supports these skills, and the project has no connection to it. What a builder can borrow is the habit of defining what done means before work starts, which is also the point of writing effective prompts.
Primary sources
- dabingfy-dot/solo-to-team-skills: Skills for human-led AI teams: delegate tasks, verify evidence, explain steps clearly, and learn from real cases. 一个人统筹 AI 团队:派活、核证、清楚说明、经验复用。 · GitHub trending AI repositories, Nov 14, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.