For programmers using coding agent skills
Create one workflow by hand. Put repeatable steps, checks, approvals, and saved state in code, then register it with your coding agent.
They combine agent skills with deterministic code, saved state, and verification. The workflow is canonical. Generated adapters are disposable.
See the full anatomy: source files, generated adapters, and trigger path →
One reusable capability.
Skills and code arranged as an executable process.
A generated file that lets an agent discover the workflow.
One request to an agent, command, or person.
One saved execution of a workflow.
A completed, blocked, or refused outcome.
Learn the complete coded-skill path once. The optional helper can accelerate the same path later.
Add the runtime and repository-local CLI.
Scaffold one workflow with init.
Write the program and deterministic fixtures.
Run doctor --test.
Create discovery adapters.
Start a new agent session.
Invoke the registered workflow.
Package installation creates no skill or coding-agent adapter. You create the canonical workflow explicitly, test its real effects with fixtures, and register only the small discovery adapters you need.
npm install --save-exact @operatorstack/yieldnpm exec -- yskill init skills/review --language typescript \
--description "Check and review a branch before it is shipped."
# Edit skills/review/main.ts and fixtures/responses.json.
npm exec -- yskill doctor skills/review --test
npm exec -- yskill register skills/review
# Restart your coding agent, then run the registered skill.
npm exec -- yskill run skills/reviewKeep order, branches, commands, approvals, and finish rules in the program. Put deterministic answers and test-only effects in fixtures/. Keep judgment and reusable guidance in SKILL.md.
yskill register detects verified agents in the project and writes the adapter path each host expects. The workflow and dependencies stay under skills/.
.cursor/skillsVerified with project-local discovery.
.agents/skillsVerified with project-local discovery.
.claude/skillsVerified with project-local discovery.
Run yskill agents with your package launcher to list every supported ID. Other hosts are registry-supported: adapter generation is tested, but Yield has not run every product end to end.
register-all checks every workflow before it writes anything. Use --dry-run first. Add --prune only when you want to remove stale adapters generated from that workflow directory.
npm exec -- yskill register-all skills --agent cursor,codex,claude-code --dry-run
npm exec -- yskill register-all skills --agent cursor,codex,claude-code --pruneThe new session can discover the registered builder and skill workflows. Ask in plain language. You do not need to paste setup instructions into the agent.
Install its language package, register the local adapter, check it, and run it. You do not recreate or move the workflow when you switch agents.
npm exec -- yskill doctor skills/review --test
npm exec -- yskill register skills/review
npm exec -- yskill run skills/reviewThe developer helper teaches and applies the same manual path. Installing a package alone never creates the helper, a skill, or an adapter.
Install the helper explicitly, review its complete plan, approve any mutation, then restart your coding agent so it discovers skills/yield-workflow-builder.
npm exec -- yskill helper install --language typescriptExplain the smallest relevant concept and manual steps. Read-only.
Plan and create one tested workflow in a new destination.
Project an existing skill into code without dropping its guidance.
Inspect one workflow without running its fixture or changing files. Read-only.
Use doctor evidence, apply an approved repair, and verify again.
Update one workflow to the installed Yield version.
Test, register selected agents, and verify their adapters.
Before approval, the installer shows resolved paths, dependency preparation, workflow testing, selected-agent registration, and final adapter verification.
Every changing mode shows the summary, relevant primitives, exact files, and exact commands before asking to apply the plan.
Summary
Primitives
Files
Commands
Apply this plan?learn and check do not mutate files; static checks do not execute fixtures.yskill helper install instead.doctor --test, registration for the selected agents, and adapter verification.yskill bootstrap and npm create @operatorstack/yield@latest remain compatibility aliases only. New documentation and examples use yskill helper install.
Use your language for branches, loops, functions, and local types. Yield only owns the boundaries between the program, the agent, commands, and people.
RunCommandYield executes a command and records its exit code, stdout, stderr, and timeout state.
Read guide →AgentTaskThe coding agent reviews, diagnoses, compares, or proposes. An optional JSON schema keeps the result usable by code.
Read guide →AskUserAsk for missing information or approval through the agent's normal interface. When choices are declared, Yield accepts only one of those values.
Read guide →RequireA failed claim ends the run before later steps. Bind it to the command or result that supports it.
Read guide →CompleteFinish only after the required paths have passed. The structured result becomes part of the run log.
Read guide →Blocked · RefusedBlocked means new evidence or an external change is required. Refused means the workflow deliberately declines.
Read guide →Write the same workflow in Go, TypeScript, Python, or Rust. Pick the language that fits your project.
import { defineSkill } from "@operatorstack/yield"
defineSkill((ctx) => {
const check = ctx.runCommand("check", "npm test", 300)
ctx.require(check.exit_code === 0, "tests pass", check)
const review = ctx.agentTask("review", "Review this branch")
ctx.require(review.critical === 0, "no critical findings", review)
return review
})Each tutorial gives judgment to the agent and moves only the repeatable control flow into code.
Run type checks, inspect the diff, and finish only with zero critical findings.
Ask for approval before your own publish command, then verify the real exit code.
Probe, diagnose, wait while a person applies a fix, then resume and recheck.
Try the cheapest hypotheses first and stop after a declared attempt limit instead of guessing forever.
Dry-run, inspect, approve, apply, and verify in an order code can enforce.
Ten ready-to-adapt skill workflows are available in every language. Open one and replace the example commands with your repository commands.
10 skill workflows · 4 languagesBrowse the whole library ↗
Stop at any point and continue the same workflow later. Answer a user question with --value or send structured agent output through stdin.
npm exec -- yskill run ./my-skill
npm exec -- yskill respond <run-id> --value <answer> --skill ./my-skill
npm exec -- yskill inspect <run-id> --skill ./my-skill
npm exec -- yskill replay <run-id> --skill ./my-skillEach response is bound to the pending step before the run advances. If the workflow changes, Yield stops instead of pairing an old answer with new code.
Yield makes workflow behavior easier to inspect and test. It does not turn every model answer into a fact.