Upskill routes AI agents to the right playbook first

Upskill is a new open-source tool that helps AI agents choose a proven workflow before they start a task. The project describes itself as a free, MIT-licensed routing layer for skills: the agent states what it needs to do, upskill finds the best matching playbook, and the agent follows that instead of improvising from memory.

The tool is aimed at assistants used in Claude Code, Cursor, Codex, Cline, Windsurf, and other shell-capable environments. According to the project description, it is meant for serious work across code, docs, slides, email, research, spreadsheets, browser tasks, design, data, auth, cloud, CRM, support, and automation.

In upskill’s model, a “skill” is a reusable playbook made up of instructions, examples, constraints, tools, and patterns for a specific kind of work. The project says this matters because many agent failures come from guessing: a model may create a generic deck outline, a fragile browser flow, or an incomplete auth implementation even though a better workflow already exists.

The example the project gives is a pitch deck. Without a skill layer, the agent may fall back to a familiar but thin structure such as title, problem, solution, market, product, team, and ask. With upskill, the assistant can be pointed toward a deck-writing skill that includes the right narrative arc, slide order, visual rules, and review checklist before it generates an editable PPTX.

Upskill’s workflow is straightforward. An agent runs upskill find with a task description, inspects the best match, and then follows that skill before execution. The project says this can reduce retries, token waste, and generic output because the model starts from a specific workflow instead of a broad prompt.

The registry behind the tool can surface skills from vendors and the community, including workflows from Anthropic, OpenAI, Stripe, Vercel, Microsoft, Garry Tan’s gstack, obra/superpowers, and other sources cited by the project. Upskill says it uses trust tiers such as verified, reviewed, and community, and it can show why a skill ranked where it did.

The project also emphasizes control and privacy. According to its description, telemetry, local context sharing, and submissions are off by default. It says only names are used for matching environment fit, not secret values, and that secret-looking payloads are scanned before feedback or uploads leave the machine.

Upskill is not trying to be another chat interface. The project frames it as a selection layer for agents, closer to package management for workflows than to a prompt library. It also says the registry can be self-hosted or pointed at a private server for teams that want internal skills behind a firewall.

The release includes install commands for agents and a CLI path for humans who want to run searches directly. The project says users can install the helper, set their privacy choices, and then search for tasks such as inbox triage, spreadsheet cleanup, or browser automation. The core claim is simple: the expertise already exists, and upskill is built to route agents to it before they begin work.