Lévin™ Harness
Download
Core features

Make expertise repeatable

Skill Creator turns a task you already know into a repeatable way of working.

In Lévin™ Harness, a Skill is a reusable recipe the Agent can follow: what to read, what to do, which tools to use, and what counts as done. Create it once, then use it in Chat, or package it through a Workflow or Plugin.

Four ways to use one method in Levin

Levin keeps methods close to where the work happens. Start in Chat, compile a visual process from Workflows, or package a complete scientific capability through Plugins. A Skill can stay personal or be packaged for a team.

01Chat workspaceType @, choose an enabled Skill, and keep the method attached to the current project task. 02Skill CreatorDescribe the task in Chat, then run the Personal Skill Levin creates for you. 03Workflow compilerBuild a visual process with tools and decisions; Levin compiles the canvas into a callable Skill. 04Plugin packageBundle a domain method with MCP services and a local or remote runtime so it can be enabled where needed.

Use a normal message for a one-off instruction. Use a Skill when the task needs the same context, decisions, tools, and verification more than once.

Find a Skill, enable it, and define its scope

Open Capabilities → Skills in Lévin™ Harness. The catalog shows what is available in the current environment, while the detail view lets you inspect the exact SKILL.md, supporting files, source, and folder before you run it.

SourceWhere it livesWho updates it
Built-inBundled with Lévin™ Harness.The app release; view or enable it, but do not delete it.
PluginThe skills directory declared by an installed plugin.The plugin package; update or uninstall from Plugins.
PersonalYour Levin configuration directory.You or the Skill creator; available across projects.
WorkflowA Skill compiled from a workflow canvas.The source canvas; rebuild instead of editing the generated file.

Before the first run

  1. Search by outcome. Match the task and expected result, not just a familiar tool name.
  2. Open the detail. Read the description, requirements, files, and declared tool or runtime boundaries.
  3. Choose the scope. Use Personal Skills for methods you create. Enable Plugin or Workflow methods only when they belong in the current environment.
  4. Check the switch. Disabled Skills leave the Agent catalog and cannot be selected with @.

The active catalog is part of the Agent's context. Keep descriptions distinct, and treat a connected tool or remote runtime as an execution boundary that still follows the current project and permission profile.

Create a Skill in Skill Creator

Use Skill Creator when you want Levin to repeat a real task. Find it in Chat, describe the work in plain language, and run the Personal Skill it creates.

Follow the creation flow

  1. 01
    Find Skill Creator in Chat

    In the chat composer, type @ and choose Skill Creator from the list.

  2. 02
    Describe the task

    State the goal, input files, tools Levin may use, expected output, and what counts as complete.

  3. 03
    Run it in Chat

    Skill Creator saves the result as a Personal Skill. Use @ to select it, then compare its returned files and report with the expected result.

Start with a concrete request

Create a Skill called validate-protein-design.

Read the selected PDB or mmCIF file and generated sequences
from the current project. Run structure and sequence checks.
Ask before continuing if an input or threshold is missing.
Return a report with the values, failures, and paths to every
output file.

You do not need to write SKILL.md first. Describe the job in Chat and let Creator turn it into a Personal Skill.

Keep one Skill focused on one result. If Levin has to guess what to read, what to change, or what to return, go back to Creator and make that instruction explicit.

How Levin runs a Skill

Every Agent run receives a compact <available_skills> catalog containing enabled names, descriptions, and locations. The full method is loaded only after a match or an explicit @ selection, so the conversation stays focused.

01 / MATCHDescribe the outcome

The Agent matches the task to an enabled Skill, or you select it explicitly.

02 / LOADRead the package

Levin loads SKILL.md and follows relative paths to supporting files.

03 / EXECUTEUse the workspace

Tools work with the active project, structures, MCP services, and selected runtime.

04 / VERIFYReturn evidence

The Agent reports what changed, what passed, and what still needs your decision.

Workflow Skills follow the same loading path. The canvas defines their source; the compiled Skill is the execution guide that Chat invokes.

Keep ownership visible

Choose a source based on who should change the method and where it is safe to run. Levin shows the location and source so a managed capability is easy to identify.

Use this scopeWhen it fitsChange it here
PersonalA method you reuse across your own projects.The Personal Skill creator or its configuration folder.
WorkflowA visual process that should stay connected to a canvas.The Workflow canvas, then rebuild the Skill.
PluginA domain capability that must travel with MCP services or a runtime.The plugin source and version, not the installed copy.

Personal Skills can be backed up with your Levin settings. Workflow and Plugin Skills remain owned by their source. Plugin updates replace the managed package; keep custom changes in the source plugin.

Review the result, then improve the method

A Skill makes a process repeatable; it does not remove your domain judgment. After a run, review the evidence in the conversation and tighten the method where the Agent had to guess.

  1. Check the boundary. Confirm the Agent used the intended project, structure selection, tools, and runtime.
  2. Check the evidence. Read the returned files, measurements, logs, and report instead of relying on a summary alone.
  3. Check the decisions. Confirm that thresholds, branches, and human approvals were handled as the Skill describes.
  4. Improve the method. Update the source Skill or Workflow, rebuild if needed, and run the same example again.

When a method becomes a stable team capability, package it as a Plugin. When it needs a visual sequence, design it in Workflows. For a single task, keep the instruction in Chat.