Steroid-Workflow is a guardrailed AI coding runtime for people who want AI speed without AI chaos.
It gives your assistant a real delivery path:
- understand the request
- inspect the codebase
- write a spec
- research the right approach
- make a plan
- build with guardrails
- verify before claiming success
The result is simple: less guessing, less fake progress, fewer broken handoffs, and much better output quality.
Raw AI coding tools are fast, but they drift:
- they jump to code before understanding the task
- they forget requirements halfway through
- they claim things are done without proof
- they patch symptoms instead of finding causes
- they leave you with a pile of files and no confidence
Steroid-Workflow is built to stop that.
It gives the model a structured path, machine-readable receipts, gate checks, and safer runtime behavior so completion has to be earned, not narrated.
npx steroid-workflow initnpx steroid-workflow@beta initThen, inside your project, tell your assistant what you want:
Build me a habit tracker like Apple Health
If it drifts, say:
Use the steroid pipeline.
Works with any AI-powered IDE or CLI that can follow project instructions.
Once Steroid is installed, the simplest workflow is:
node steroid-run.cjs start habit-tracker
node steroid-run.cjs next habit-tracker
node steroid-run.cjs finish habit-trackerSteroid turns vague requests into a real workflow with checkpoints.
For most work, the top-level flow is:
start <feature> -> next <feature> -> finish <feature>
start bootstraps the feature and scan, next answers what to do now, and finish tells you whether review, verification, and archive are actually ready.
Under the hood, Steroid still routes the work through a stricter internal pipeline. For normal feature work that usually means:
scan -> vibe -> spec -> research -> architect -> engine -> review -> verify -> archive
For bug and repair work, the route becomes more diagnose-and-fix oriented instead of forcing the full feature path.
As the workflow moves forward, Steroid writes durable artifacts under .memory/changes/<feature>/, including:
request.jsoncontext.mdprompt.jsonprompt.mdvibe.mdspec.mdresearch.mdplan.mdtasks.mdexecution.jsonreview.mdreview.jsonverify.mdverify.jsoncompletion.json
These are not just logs. They are used by the runtime to decide whether later phases are allowed to proceed.
Steroid does not just ask the model to “be careful.” It checks for real artifacts and receipts before allowing later phases like verify, report, and archive.
The model is guided through codebase scanning, specification, research, architecture, and execution instead of improvising from the first prompt.
Steroid is designed for real repos, not just toy greenfield demos. It helps reduce silent deletion, fake tests, and reckless edits.
Feature building and debugging are different jobs. Steroid gives them different paths.
Steroid has built-in frontend flows for UI-heavy work.
For UI-intensive features, it can generate and enforce:
design-routing.jsondesign-system.mdaccessibility.jsonui-audit.jsonui-review.mdui-review.json
That means frontend work is not just “make it look nice.” It can be routed, reviewed, audited, and verified like the rest of the pipeline.
Useful commands:
node steroid-run.cjs design-route "<message>" --feature <feature> --write
node steroid-run.cjs design-system --feature <feature> --write
node steroid-run.cjs design-prep "<message>" --feature <feature> --write
node steroid-run.cjs verify-feature <feature>
node steroid-run.cjs verify-feature <feature> --deep
node steroid-run.cjs verify-feature <feature> --deep --url <preview>
node steroid-run.cjs review ui <feature>Steroid pairs this path with its internal frontend intelligence layer so UI-heavy work can move through a more opinionated design-and-review flow instead of pure freeform prompting.
If ui-review.json is FAIL, archive will stay blocked until the frontend issues are resolved.
Steroid includes hardening specifically aimed at preventing common AI failure modes:
- intent routing so different kinds of work take different paths
- Prompt Intelligence for vague or messy requests
- circuit-breaker behavior for repeated failure
- command confinement and allowlist checks
- guarded file operations
- true verification receipts instead of “trust me”
- Optional Deep Verification for stronger frontend and runtime evidence
Steroid can normalize rough user requests into something the rest of the pipeline can actually work with.
Useful commands:
node steroid-run.cjs normalize-prompt "<message>"
node steroid-run.cjs prompt-health "<message>"
node steroid-run.cjs session-detectOptional Deep Verification is available through verify-feature --deep, and verify-feature --deep --url <preview> lets you point that browser-backed verification path at a specific preview when needed.
Use stable if you want the most conservative install:
npx steroid-workflow initUse beta if you want the newest runtime hardening and governed workflow improvements first:
npx steroid-workflow@beta initCurrent beta work focuses on:
- stricter artifact enforcement
- stronger verification and archive checks
- safer command/runtime behavior
- improved frontend routing and review receipts
npx steroid-workflow@latest updateor:
npx steroid-workflow@beta updateYour .memory/ project state is preserved.
node steroid-run.cjs auditThis checks the enforcement/runtime setup inside the current project.
MIT © nzkbuild