Skip to main content

Adopt It

Adoption is layered. Each component is independently useful, additive (nothing you already have is overwritten), and skippable if it does not apply. You can evaluate the methodology in two steps and expand to the full setup later without redoing anything.

The adoption path​

The setup follows five rules in order: read before write, show before apply, additive not destructive, skip-friendly, and verifiable (a final self-check confirms what is configured). Working through it with Claude Code open lets the assistant read your existing configuration, propose merges as diffs, and apply them only after you approve.

  1. Install the Superpowers skills. The lifecycle skills enforce the brainstorm, design, plan, implement, verify, review sequence. Without them, the assistant defaults to jumping straight to code; with them, the structured process is the default process.
  2. Add the Stop verification hook. A single prompt hook that blocks completion until verification evidence is present. If you adopt only one thing, adopt this — it directly catches the "it looks done but was never tested" failure mode.
  3. Merge the instruction layer. Merge the methodology templates into your global instructions and your project CLAUDE.md. The project file is the highest-leverage artifact in the repository: it carries the tech stack, build commands, testing rules, architecture, and conventions, turning a generic assistant into a project-aware collaborator. Fill in the template placeholders together with the assistant.
  4. Add reviewer agents. Copy the specialized review agents (for example an API reviewer, a compliance reviewer, a migration reviewer) into your project. These encode domain expertise into repeatable, checklist-driven reviews. Take the ones that match your domain and skip the rest.
  5. Add the quality gate hooks. Post-edit linting, a pre-push test and coverage gate, and the Stop verification hook form a defense in depth: post-edit catches the small mistakes, pre-push catches the medium ones, and Stop verification catches the unverified-completion case.
  6. Scaffold the memory system. Create a memory directory with an index file. Do not pre-populate it — memories accumulate naturally through collaboration so each session starts with richer context than the last.
  7. Configure permissions, then verify. Optionally curate a safe allowlist to reduce friction, then run the self-check that reports each component as configured, partial, skipped, or missing.

Minimal vs. full adoption​

You do not have to adopt everything at once. The minimal path is the instruction layer plus the lifecycle skills — that alone changes how the assistant approaches work, enforcing design before code and verification before completion. Hooks, reviewer agents, and the memory system can be layered on later as you build trust in the setup. A partial setup is a valid setup; the self-check distinguishes intentional skips from oversights.

The methodology is not a one-time install. It is a compounding system: the more you use it, the more accumulated context the assistant has, and the better the collaboration becomes.

Where to next​