Superpowers 5 is out, and it leans into a simple idea: if an agent is going to do more of the work, it should also shoulder more of the *communication* and *process* overhead that tends to slow down AI-assisted coding. The release centers on **Visual Brainstorming**, plus several workflow changes aimed at making plans more reliable, subagents more useful by default, and the overall pipeline a bit more disciplined.
Visual Brainstorming: getting out of ASCII art
The standout addition is **Visual Brainstorming**, a companion flow that lets an agent present interactive visuals in a browser—**mockups, diagrams, comparisons, and other HTML-based artifacts**—instead of trying to compress UI and UX discussions into terminal-friendly ASCII.
When the agent believes something should be shown visually, it prompts for permission and notes that the feature is still new and can be token-intensive. Under the hood, the system spins up a local web server that loads HTML written by the agent from disk, plus some client-side JavaScript to capture clicks and feedback as interaction happens in the browser.
The author describes using the feature with Claude running Superpowers 5 to clean up rendered webpages from Youtube2Webpage, with the agent directing attention back to the browser once it had a view worth reviewing. The same workflow also gets positioned as a better fit for brand and logo ideation—an example appears for Prime Radiant—where terminal diagrams tend to fall apart.
Visual Brainstorming is mostly tested in **Claude Code** and **Codex**, but is expected to work across most agents.
Spec Review: an adversarial loop for planning docs
Superpowers 5 also adds a **spec review loop** focused on a very practical failure mode: planning documents that quietly leave key decisions as “TBD” or “fill this in later.” After planning completes, Superpowers now launches a subagent to read the planning docs for **sanity and completeness**, acting as an adversarial reviewer.
It’s explicitly not presented as a substitute for a human glance at the spec, but the goal is to push planning docs toward being more execution-ready.
Subagent Driven Development becomes the default
Another major workflow shift: **Subagent Driven Development** is now the default when the host environment supports it. Previously, Superpowers offered a choice between that approach and a more manual workflow where execution happened chunk-by-chunk in a separate session.
In harnesses such as Claude Code that can pick different models for subagents, Superpowers 5 instructs the system to use the **cheapest model capable of the task**—with an example that **Claude Haiku** can often handle implementation when plans are sufficiently detailed. The tuning also includes better signaling when subagents are out of their depth and need help.
If a harness doesn’t support subagents, Superpowers warns that a subagent-capable harness performs better, then proceeds in a single session as best it can.
More explicit software engineering guidance through the pipeline
As projects scale, Superpowers 5 adds more structured guidance around **unit decomposition** and **interface-driven design**, wired through multiple stages:
- Brainstorming now emphasizes **single-purpose units**, clear interfaces, and independent testability, and treats **large files as a design smell**.
- Writing plans now includes a **File Structure** section before task decomposition.
- Implementation prompts instruct workers to follow the planned file structure and escalate if a file outgrows scope.
- Code review checks for **single-responsibility violations** and file size growth.
Brainstorming also starts identifying projects that look “too big” and works interactively to break them into manageable pieces.
Smaller but meaningful changes
A handful of operational updates round out the release:
- **Documentation paths** move from `docs/plans` to `docs/superpowers/specs` and `docs/superpowers/plans`.
- **Local instructions** (from `CLAUDE.md` or `AGENTS.md`) are explicitly prioritized over Superpowers internal instructions to make customization simpler.
- **Codex subagent recursion mitigation** addresses cases where subagents start delegating recursively after picking up the `using-superpowers` skill.
- Legacy **/commands** are deprecated in favor of native skills (with removal planned in a future release).
Superpowers 5 is available now on GitHub, and plugin-based harnesses like Claude Code or Cursor are expected to auto-update over the next day or two; direct installs may require a pull or reinstall.