Halfcycle vs Superpowers: a Superpowers alternative that keeps learning

Most comparisons of AI coding frameworks count features. Two questions matter more: who thinks about what could go wrong before it is built, and what you hold when the building stops.

Superpowers is a good method written down. Its skills are files in a plugin on your machine, and the agent checks for a relevant one before every task. The method is exactly as good as the version you installed, on every project, until you update it. It does not know what went wrong on anyone else's project, and it gives the same ceremony to a two-screen tool as to a payments system, unless the agent judges otherwise.

Halfcycle is served, not downloaded. Each step reaches your project from Halfcycle's service when it is needed, so the method you use today is better than the one you used last week, with nothing to update. It learns from every project it runs on, and it sizes the work from the parts of your system a change touches, not from a guess. One size fits nobody; a served method can fit each job.

Halfcycle is not a platform. It works inside Claude Code or OpenAI Codex on your own computer and works like a software team made of AI agents. Before anything is built, it asks the questions a product owner, an architect and a senior engineer would ask, and writes down the answers as the plan. Then a team of AI agents builds the software piece by piece, another set of AI agents check every piece, and the finished app is tested. What you end up with is ordinary code in your own folder, built with whatever suits the job, that you can put live anywhere.

Superpowers has earned its following, and it is worth being specific about why.

• Free and open. MIT licensed, every skill readable, nothing hidden. Halfcycle is free for solo developers too, though it is not open source.
• Ask first, design before code. Its brainstorming skill asks questions and nothing is built until you approve a written design. Halfcycle does this too: it asks one question at a time with a recommended answer, and writes a product spec, the architecture and a design for each part before any building.
• A light path for small work. The agent judges whether something is a small change or a project. Halfcycle sizes work too, and its sizing comes from the parts of your system a change touches, worked out by Halfcycle rather than judged by the agent.
• Test-first, with fresh reviewers. A fresh agent per task, a reviewer after each one and a final review of the branch. Halfcycle does this too, and reviews every task twice with an agent that did not build it.
• Many coding agents. Its README names 16, from Claude Code to Codex and Cursor. Halfcycle works with Claude Code and OpenAI Codex too, with OpenCode coming soon.
• No service needed. It runs from files on your machine. Halfcycle needs its service, with an offline fallback coming soon.

Its skills are the same on every project until a new version reaches you through your agent's plugin system. Nothing in its README describes learning from other projects. Halfcycle does: when any project goes wrong in a new way, the cause and the fix reach yours automatically, so most of those problems never arrive, and the ones that do come with the answer.

Superpowers writes a design for each piece of work. Its README describes no product spec, architecture or phased plan across features. Halfcycle starts there, so the tenth feature fits the first.

Its build skill gives each task a fresh agent, one after another, and says never to run several builders in parallel. Halfcycle runs up to eight builder and reviewer agents at once, each on its own copy of the code, coordinated from one plan.

You follow a Superpowers build in the chat and the files it writes. Its one hook loads its instructions when a session starts; its README describes nothing that stops a risky action. Halfcycle has a console for every project, and checks risky actions before they happen. Its README also describes no testing of the running app; Halfcycle clicks through the live app in a browser at the end of every phase.

The examples below come from a real project we built with Halfcycle: Hookbox, a small service that receives payment notifications from Stripe and keeps them safe, so that none get lost if the system that should handle them is down.

Most people can't tell whether what they are asking for is a five-minute tweak or a week of work, and they shouldn't have to. Halfcycle works it out for you. Before it touches anything, it looks at which parts of your software the request affects, then tells you which of three ways it will go, and why:

• A small change, like new wording on a button or a fix on one screen, goes straight through in one session. No plan to approve and no documents to read. It is still reviewed twice by an agent that did not write it, and tested.
• A change that reaches across parts, like adding an age field that the form, the server and the stored data all need to know about, gets a short plan for you to approve and a check of the result.
• Something bigger, a new product or a major feature, gets the whole team: the questions, the full plan, the designs, the build and the tests, unless you have asked for a quick prototype first.

If a small change turns out bigger than it looked, it moves itself up to the right path and tells you. You can ask for more care on anything, and it won't skip a step a change needs. So quick things stay quick, and you only sit through the questions when the work calls for them.

And if all you want right now is a quick prototype, say so. Halfcycle builds a quick version to try, and when it suggests where the data, the logins and the hosting should live, it picks options that take minutes to set up rather than days. If the prototype proves itself, the same project carries on with the full process from there.

A framework gives your coding agent instructions and leaves the judgement to it. Halfcycle starts by working out what you are really building. It asks one question at a time, gives a recommended answer and the reason, and raises the problems you would not know to look for. On Hookbox it found six of them before a single line of code was written.

One is shown below. Hookbox was meant to let you resend a payment notification later if it had failed the first time. But Stripe's notifications carry a time stamp that expires after five minutes, so a resent one would have been rejected as fake. Halfcycle spotted it, explained what each fix would cost, and recommended one.

The others included Hookbox quietly switching off Stripe's own resending, a goal nobody would have been told was failing, and whether to build it at all instead of using an existing tool. Superpowers' brainstorming asks good questions about what you asked for; Halfcycle also brings what other projects learned the hard way.

Halfcycle keeps learning. When a software project goes wrong in a new way, two things are kept: what caused it, and what fixed it. Both reach every other project automatically, with nothing for you to download or update.

For you, that means two things. Most of the problems other projects have already run into never arrive on yours at all. And the ones that do arrive come with the fix already known, because somebody has already found it. The more projects it works on, the longer that list gets.

What travels between projects is the shape of a problem, never your code or your documents, and we do not train on your work.

A good software team writes things down before it builds, so that everyone, people and agents alike, works from the same picture. Halfcycle does the same, and every document is saved in your own repository as a plain file you can read, change and keep:

• The product spec. What you are building, who it is for, what success looks like, and the options you turned down and why.
• The architecture. The parts of the system and how they connect, with a diagram, plus the rules the system must keep to.
• The phases. The work cut into stages small enough to finish, with the riskiest part first and what "done" means for each.
• A design for every component in the phase. What each part does, how it behaves, and how anyone can check it works.
• The build plan and its record. Every task, who built it, who checked it, and what they found.

On Hookbox that came to a product spec, an architecture overview, three phase plans and seven component designs, among others. The agents read them at the start of every session, which is why the tenth task fits the first. Superpowers writes a design document for each piece of work. Its README does not describe a product spec, an architecture or component designs across the product.

docs/
├── product-spec.md            what it is, who for, what success looks like
├── architecture-overview.md   the parts, how they connect, the diagram
├── phases/
│   ├── poc.md                 the first phase and what "done" means
│   ├── phase-1.md
│   └── phase-2.md
├── features/                  one design per component
│   ├── stripe-receive.md
│   ├── delivery-store.md
│   ├── signed-forwarder.md
│   ├── inbox-minimal.md
│   └── …
└── build-records/             every task, every review

Once the plan is agreed, Halfcycle splits it into tasks in the order they can be done and hands them to a team of AI agents with different jobs, the way a real software team splits the work. Every one of them is an AI agent except you. Analyst and architect agents question the idea and design the system. Builder agents write the code, each on its own copy. A separate set of reviewer agents checks every piece before it is added to the project, and a tester agent checks the tests and runs quick checks that the finished app works. Two lead agents run the team. You approve the plan and accept the result.

When a phase is finished, one more agent, one that built none of it, opens the live app in a browser and walks through it the way a person would: signing in, clicking through each journey the phase was for, and checking what appears on screen. Anything it finds goes back to the team to fix, and the journeys are walked again until nothing is left in the way.

On Hookbox one check failed: the part that receives Stripe's notifications was tested with sample data that was missing something the plan required. It was fixed, checked again by a third agent, and passed. Two hours and eleven minutes after the building started, every task that did not need a person was built, checked and added. It stopped where it should, at the jobs only a person can do, like creating the Stripe test account, each with step-by-step instructions already written.

When software built with AI has a security hole, it is rarely one bad line of code. More often it is a gap between parts: a rule about who can see what, applied on one screen and missing from the next, or a test login quietly standing in for a real one. Halfcycle is built around those gaps.

• Who can sign in, and what each kind of user may see and do, is written into the plan once. Every part built afterwards has to follow it, and the reviewer agents check each piece against it.
• A missing login or secret key makes the app stop with an error, so a stand-in login cannot slip into the live app looking like a real one.
• How secret keys are kept is settled when the project is set up, and the project is checked against it, so keys do not end up in your code.
• The finished app is walked by an agent really signed in, on the live version, not on a test copy with the doors left open.

So security is part of how your app is designed and built, from the first plan to the live version, rather than a scan at the end that looks for holes once they are already there.

Every Halfcycle project gets an online console: a web page where you, or anyone you invite, can see the whole project without opening a terminal. Superpowers has no console: you follow the work in the chat. The console shows you the work behind it: the plan, the team of agents, the system and the progress.

The overview shows each stage of the work, what was approved and when, and any problem the checks found in the code. The first one on Hookbox was a line of code that updated a record nothing had ever created, so it would have quietly done nothing at all.

The system view draws your whole product as a diagram: the apps people use, the services behind them, the data and the outside services, with a card for each part. Before a phase is built you see exactly which parts it will add or change, and you approve them one by one. While it is being built, each card shows whether that part is built and checked, in progress, waiting for a person or blocked. It also shows two things that usually only come out at the end: a part that changed though the plan did not include it, and a part the plan needs that no task is building.

No agent builds a phase until you have approved it, part by part. The console lists every part of your system the phase will add or change. You check each one against the system around it, mark it Looks right or ask for a change in plain words, and the approve button stays off until every part is checked. You can approve from the console or by typing it in your coding agent; either way the approval is recorded as yours, with the time, never the agent's. If the plan changes after you approve, your approval is not carried over: only the parts that changed come back to you.

The live build shows the phase as waves of tasks, with the agent working on each one, what has passed its check, and what is waiting for a person. Open any task to see who is building it, who will check it, how long it has left, and what is waiting on it.

The ledger lists every time something was stopped or corrected: a change blocked in the code, a mistake fixed before the task closed, a claim in the plan corrected, a problem prevented by a question you answered. Each one says what changed, which agent or person caught it, and the file or answer that proves it. Below it are the checks in force on your project, what each looks for and whether it blocks or warns. The Phases page keeps each phase's dates, approvals, defects found in testing and outcome.

The console works fully on a phone, so you can see where a build has got to, and what is waiting for you, without being at your desk.

Superpowers works in an existing repository too. Halfcycle reads your existing code, tells you what it worked out and what it is unsure of, and plans the next piece of work from there.

If your team uses a coding agent other than Claude Code or Codex, Superpowers works there today and Halfcycle does not yet. If you need to work with no connection to a service, or your organisation requires open-source tools you can read line by line, Superpowers fits and Halfcycle does not. And if you would rather write and tune the method yourself, Superpowers gives you the files to do it; Halfcycle tunes the method for you.

Not on the same project, because both run the process and would give your agent competing instructions. Many people move between them: if you have been using Superpowers, remove its plugin from the project, run Halfcycle, and it reads the code and the design documents you already have and carries on from there.

Halfcycle is free for solo developers, with the full process and every check. The agents run on your own Claude or OpenAI account, so the building itself is part of your normal usage. Details are on the install page.

Superpowers is free and open source under the MIT licence, and its makers offer paid enterprise support, according to its README on 30 October 2026.