An important detail is stated honestly in the README: installation is done separately for each environment. The skills are portable, but the loading machinery differs between harnesses.
The repository also carries its own tests - sixteen suites, roughly one per supported environment - along with documentation and hooks. This is a maintained project rather than a pile of markdown. The porting guide deserves a separate mention: fifty kilobytes of text, the largest document in the repository. It is what explains how there came to be fourteen environments. Commercial support for companies is advertised too, which usually means the project will not be abandoned in six months.
Cost one: the ceremony does not scale down
The main inconvenience is obvious.
The full cycle - brainstorm, design, sign-off, worktree, plan, subagents, tests, review, finishing the branch - is absurd for fixing a typo. A two-minute task acquires half an hour of procedure.
The project is built for large work and pays off there. But telling apart where the procedure is needed and where it merely burns time and tokens is left to you - and the meta-skill's blunt wording pushes towards switching it on everywhere.
Cost two: dogmatism where flexibility is needed
An absolute requirement to test first is useful as a guard against a hurried agent. But real engineering has exploratory work, generated code, one-off hypothesis checks and changes where a different order of verification makes more sense.
"Wrote code before the test? Delete it" applied to exploratory work means throwing away the thing the exploration was for. Formally you can declare it reconnaissance rather than development - but the methodology offers no such distinction.
Cost three: trusting the autonomy is a separate decision
A couple of hours of autonomous work is the advertised strength, and equally a demand on you.
You approve the plan; after that the subagents carry it out themselves, reviewing one another. If the plan was imprecise, the error replicates across every task before you see it. Isolation in a separate tree saves the repository, but not the hours.
Hence the practical consequence: quality shifts entirely onto the design and plan stage. Skim it and you get the wrong thing built quickly and neatly.
Who it suits
Large pieces of work. Where "just start writing" almost certainly leads to rework, the design and plan stage pays for itself immediately.
Anyone wanting autonomy with boundaries. Approving a plan once and getting a couple of hours of unattended work is exactly the model this project implements, and implements thoughtfully.
Teams. The process stops depending on who phrased the request and how. The same sequence for everyone.
Anyone irritated by the agent's habits. Rushing, skipping diagnosis, writing code before understanding the task, declaring success without proof - Superpowers targets precisely that list.
Who it does not suit
Small edits and one-off tasks. The ceremony costs more than the result.
Exploratory work. A test-first requirement sits badly with reconnaissance, where you do not even know yet what to check.
Anyone unwilling to invest in the design. The whole construction rests on you genuinely reading the specification and the plan. Nodding through turns autonomy into the fast production of the wrong thing.
Conclusion
Superpowers is the most direct answer to the question of why an agent does not follow good instructions. The answer: because instructions are requests, and what is needed are procedures with entry conditions and verifiable outputs.
The value is not in the number of skills - there are only fourteen. The value is that they are wired into a sequence where every step has a trigger and a result you can present. And in a meta-skill that forbids weaselling in plain words.
Whether that strictness suits you is a question of the scale of your work. On a large piece it looks like ordinary engineering discipline finally written down so it cannot be quietly ignored. On a small one, like an excessive ritual.
Sources
The material was verified on 14 August 2026 against the project's repository. The composition was counted in a fresh clone: fourteen skills, four hooks, sixteen test suites and forty-three documents, thirty-four of which are the project's own internal plans and specifications. Metadata - about 272 thousand stars and 24 thousand forks, MIT license. The quotations from the meta-skill and the testing requirement are taken verbatim from the skill files. Supplemented on 15 August 2026 against a fresh clone: both quotations, the set of skills, the number of install environments and the presence in the official plugin catalogue were re-checked and held; the document count was corrected and the porting guide added.
