Superlite is a Claude Code plugin designed to let a user hand an agent a specification and return to finished, tested, committed work without needing to approve every step. The author of the post on contextfirst.blog explains that the motivation came from frustration with the existing Superpowers workflow, which required frequent human gates—estimated at 8 to 12 per feature—and executed tasks strictly sequentially, with review loops and repeated test suite runs that slowed progress.
According to the post, Superlite changes this by collapsing the approval process into a single batch of questions and one design‑and‑plan document that the user approves once. After that approval, the plugin only stops for actions that are destructive, outward‑facing, or security‑sensitive; otherwise it proceeds autonomously. The plan is expressed as a task graph where each task lists the files it owns, the interfaces it consumes and produces, and the tests it must pass. Tasks that do not share files or depend on each other can run in parallel, allowing the build to use many agents at once. In the benchmark described, the graph peaked at 18 agents working simultaneously, which the author likens to the size of a small startup minus a ping‑pong table.
Agent roles are divided by capability: Opus acts as the coordinator, planning and dispatching work but is barred from editing code by enforceable hooks; Sonnet builds most tasks; Haiku handles small, fully specified “mechanical” tasks that are not on the critical path; Sonnet also reviews packages; a fresh Opus final reviewer examines a review packet built from run logs and only points out findings; fix‑only agents in three tiers address those findings based on how well the cause is understood.
To prevent the coordinator from overriding its constraints via prompts, Superlite uses code‑level hooks. One hook blocks the coordinator from editing project code; another converts raw test commands into background runs so tests never block the coordinator; a third refuses to let anyone kill processes by name. Before work begins, the plan must pass lint --stable, and stable tasks are dispatched immediately after passing.
Every run records each agent’s token usage, cost, and time. This data feeds predictions for future runs and triggers a soft budget warning at 70 % of a predefined cap.
To use Superlite, a user needs Claude Code, git, Python 3.10 or newer, and bash (Git Bash on Windows). The source lives at the GitHub repository iha2/superlite. Installation can be done via the plugin marketplace with /plugin marketplace add iha2/superlite followed by /plugin install superlite@superlite-dev, or by cloning the repository and loading it locally with claude --plugin-dir <path-to-superlite>. A quick sanity check is to start a fresh session and prompt “Let’s make a react todo list,” which should trigger the plugin to brainstorm before writing any code.
When starting a task, the user provides a paragraph or a spec file; no lengthy requirements document is required, though the plugin will read one if supplied. The user then chooses whether a human or an AI orchestrator will manage the plan. The post notes that Superlite shares skill names with Superpowers, so the two should not be enabled in the same session.
Key facts
- Superlite is a Claude Code plugin for autonomous build workflows.
- It reduces approvals to a single batch of questions and one design‑and‑plan document.
- Tasks are represented as a graph enabling parallel execution; benchmarks showed up to 18 agents running simultaneously.
- Agent roles: Opus coordinates (no code edits), Sonnet builds most tasks, Haiku handles small mechanical tasks, Sonnet reviews packages, a fresh Opus reviewer only points findings, fix‑only agents address those findings.
- Code‑level hooks enforce constraints such as preventing the coordinator from editing code and turning test runs into background processes.
- Installation requires Claude Code, git, Python 3.10+, bash; source is iha2/superlite on GitHub.