All stories
Lights-on software factory Ecosystem Community reports

An 80% token-saving claim, with humans still in charge: Lights-on for Claude Code

Gojko Adzic's MIT boilerplate moves mechanical orchestration into scripts, adds command rules and test gates, and keeps humans supervising Claude Code.

Sources (8)

Why now The October 20 Show HN submission at 01:22 UTC introduced the workflow to a public English discussion. The reported savings describe its author's Votito setup, not a newly measured universal result.

Gojko Adzic says the orchestrator behind his Lights-on software factory reduced token spending by around 80% at Votito compared with bare-bones Claude workflow execution. The newly discussed boilerplate makes the mechanism inspectable: use deterministic scripts for mechanical coordination, then let agents implement and test the work under human supervision.

The October 20 Hacker News introduction is the current news hook. The savings are the author’s experience with a particular setup, not an AITrending benchmark or a promise that every project will cut its bill by the same amount. In the discussion, Adzic explains that “lights-on” means people supervising agents rather than handing the entire factory to them.

From a plan to workers and gates

This is an MIT-licensed Claude Code project template, not a new model or a standalone replacement for the harness. A plan in workspace/plans goes through /impact-analysis, /plan-execution and /execute. A generator turns the implementation plan into a workflow script with dependency groups and separate worker, gate and check roles.

Workers can handle separate files in parallel. Gates inspect the combined tree between groups, and the final gate runs broader project and tooling checks. Progress records support resuming incomplete work. Stuck runs still require human attention, and the example gates need adapting to the real application’s tests, builds and type checks.

What running it actually requires

The host needs Docker and make. Start with make init, then make claude. The reviewed manifest requires Node 22 or newer inside the workspace and declares Claude Code ^2.1.304; that range is not an exact installed-version guarantee. The supplied configuration enables Workflows, /loop and Cron tools, so the flow depends on those Claude Code capabilities being available.

An ANTHROPIC_API_KEY can be passed through, or Claude Code can use its login flow. Model access and charges remain separate from the free template. Login and history persist in the git-ignored .claude-home directory, which should not be shared or committed.

The normal container runs as the non-root builder user. Its actual makefile mounts the repository directory, host /tmp, the login directory and, when configured, an extra temporary directory. That is broader access than a workspace-only description suggests.

A Bash hook applies explicit allow, ask and deny rules; lint feedback runs after edits. These are useful controls to inspect and customize, not proof of complete containment. In particular, the regulator sees a top-level make invocation rather than every command inside its recipe. Teams adopting the template should review those recipes, mounts and credential access alongside the workflow’s attractive token-saving claim.