A developer has released Superlite, a Claude Code plugin designed to complete clearly specified software projects with less ongoing supervision. The workflow is aimed at users who already know what they want built and would prefer to approve a plan once instead of answering repeated questions while an agent works.
Superlite begins by sizing the request. Small changes can proceed directly, while changes to an existing flow receive a short design and new projects trigger a batch of up to four questions. It then creates a design and implementation plan for a single approval. After that point, the system is intended to pause only for destructive operations, security-sensitive decisions or actions that affect external systems. The developer says it will not push or merge code without the user.
The plan is structured as a task graph rather than a strictly sequential checklist. Each task identifies the files it owns, its inputs and outputs, and the tests it must pass. Work that has no shared files or unmet dependencies can run in parallel. According to the project's author, one benchmark reached 18 simultaneous agents. Superlite can use either a human orchestrator, who approves the schedule, or an AI orchestrator that records decisions and their tradeoffs for later review.
The plugin also relies on executable controls. Hooks prevent the coordinator from editing project code, move test commands into background runs, and block process termination by name. Plans must pass a stability lint before tasks are dispatched. Usage, elapsed time and token costs are recorded, with a warning when a run reaches 70 percent of its budget. These controls are meant to make autonomy observable and to enforce boundaries outside the prompt itself.
Superlite borrows elements of the Superpowers workflow but is presented as serving a different use case. Superpowers keeps a person close to individual steps; Superlite prioritizes unattended execution after an initial agreement. The author cautions users not to enable both plugins in one session because their lifecycle components and skill names overlap.
The published benchmark used the same specification to build a combined static-site generator and spreadsheet application, then evaluated results against 177 hidden tests. The author reports that Superlite passed all 177 in about half the time of earlier Superpowers runs on the same machine. That is a project-reported comparison from a single current run, not an independent performance study.
Using the plugin requires Claude Code, Git, Bash and Python 3.10 or later. Its source is available in the iha2/superlite GitHub repository. The release adds another approach to agentic coding: reducing conversational checkpoints while retaining explicit approval, logged decisions and code-enforced safety boundaries.
