Prime Radiant

Superpowers 6.4.2

Superpowers 6.4.2 makes planning faster and plans more concise.

Share

Superpowers 6.4.2 is out, with an update to writing-plans that produces leaner plans and keeps agents focused on planning.

In the 6.4 announcement, we mentioned that we were testing changes to make planning faster and cheaper. This release includes some of that work.

Some frontier models, including Opus 5.5, were getting carried away while writing plans. With certain prompts, they would start building the entire project just to work out what belonged in the plan. That meant spending time and tokens on implementation before the planning stage was finished.

The updated skill focuses on the decisions an implementing agent needs: which files to change, the exact function names and signatures, the values specified in the design, and the tests that prove the work is correct. Once those details are settled, the implementing agent can write the code. The plan still includes an algorithm when the interface and tests leave an important choice unresolved.

We've also added a check for the plan's length during self-review. If code blocks dominate the document, the agent replaces unnecessary function bodies with signatures and test assertions. The goal is to keep the plan specific enough to follow without writing the implementation twice.

When we reproduced the original report with the updated skill, the premature builds stopped. Writing the plans took a quarter of the time and about a third of the tokens. Those figures come from that reproduction. In separate tests using Sonnet 5, every new plan scored 9/9 against checks with deliberately introduced defects, matching the results from plans that included the full code.

Thanks to Harper Reed for the report and session bundle that helped us track this down.

You can read the full 6.4.2 release notes and compare the changes since 6.4.1.

Installation instructions for your coding agent are in the Superpowers README.