Loop & Gate

loop & gate · build kit

The Build Kit.

A judgment layer for building software with an agentic loop. Anyone can install the tools that write code. Almost nobody knows where to intervene. This kit is the map of those points, and how to work each one even when its expertise isn't yours yet.

Install it

what it is

the judgment layer, not the loop

An agentic build loop can run brainstorm → plan → build → test → review → ship on its own, and it should. But it makes wrong assumptions and runs with them, it overbuilds, it calls a one-off demo “tested,” and it can't tell a real problem from a loud one. So the loop stops at a fixed set of points and waits for a human. Those points are the gates. In Loop & Gate, this is the build half: an idea to shipped software.

Work them well and a mediocre tool stack ships good software. Work them badly and the best tools on the market ship you a fast, tested, well-reviewed mistake.

It is not the build tooling. It sits on top of a loop it doesn't ship — two free public plugins do that — and it is not an autopilot: every gate stops for your decision. The gates are the asset. The tools are a disposable appendix, and the kit's own reference files label the tool table as a dated snapshot for exactly that reason.

the eleven gates

∞ first, every time

Each gate names its lens — Business for product judgment, Engineering for developer judgment, or Both — so you know when you're on home ground and when to slow down and lean on the structure. Gate ∞ is the one you never hand to the agent, because it is the decision about how much to trust the agent.

the gates, in the order they run
GateThe questionLens
∞ How much process does this change deserve? Right-size the process to the stakes — how deep, and how wide. A typo fix doesn't get a spec and a security pass; a billing change gets all of it. Both, above the others
0 Should this exist? Does it move the product toward the outcome you care about, and is it worth your scarcest resource — your attention this month? “Would be nice” is not a yes. Business
1 Who is it for, and what's a win? A specific person and an observable success signal, not a segment and “customers will love it.” Business
2 Which direction? The one that fits where the product is going, not just the one that solves today's ticket. Business
3 Right plan, right scope? Approve, adjust or send back: does the scope match Gate 1, and is the approach sound enough to build on? Both
4 Is the architecture sane? Trading present convenience against future cost, before it becomes load-bearing. Engineering
5 Off the rails? Is the loop converging or spinning, and does a choice it made on the fly need your input? Engineering
6 Does the test prove it? An assertion that is still true next month, or just the agent's transcript of a run? Engineering
7 Is the cost right? Match the spend to the stage; cheap work goes to cheap models, and infra is sized for the load you have. Engineering
8 Risk acceptable to ship? Passing gates is necessary, not sufficient. Ship, hold, or ship narrowly. Both
9 Ship now, to whom? When it goes live and who gets it first — timing, communication, staged exposure. Business

Each gate in the reference also carries what good judgment looks like, how to work it when its lens isn't yours, and the failure modes of rubber-stamping it. That reference ships inside the plugin as reference/gates.md.

install

three steps, one paid plan

You need a Mac or Windows machine and a paid Claude plan — Pro or Max. The free plan does not include Claude Code. If you have never installed any of this, the Foundation's getting-started page walks the same ground more slowly, and the kit's own GETTING-STARTED.md assumes no coding background.

Add the kit

In the Code chat box, on the desktop app, the web app or the CLI:

/plugin marketplace add slogsdon/loop-and-gate-build-kit

Then click Install on the menu that appears. The loop-and-gate skill is now available.

Add the build pipeline

The gates are a layer on top of a build loop, so the loop has to be there too. Same place, one line at a time — the first pair is the core loop, the second fills the gaps it leaves:

/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace
/plugin marketplace add addyosmani/agent-skills
/plugin install agent-skills@addy-agent-skills

Both are public and free. Bringing your own loop instead is a supported answer, not a workaround: the Foundation's add-kits skill offers each kit's pipeline as a declinable choice, and this kit's snapshot notes which gates have no tool at all.

Run a change through it

Open the folder you want to build in, then:

/loop-and-gate

Tell it what you want to build, in plain words. It sizes the change first, then walks you through only the gates that change earns.

On pi, the install is one command and the skill is a /skill: command:

pi install git:github.com/slogsdon/loop-and-gate-build-kit

Or git clone the repo and use the skill directly; scripts/setup.sh is an optional convenience on that path.

a worked example

one repro, two sizes

A customer reports the lead export is missing a field. You run /loop-and-gate:

Gate ∞ first. What breaks if this is wrong, and is it reversible? An export change with no money and no auth in it is reversible, so it earns a few gates rather than all ten.

Gate 1. Who asked, why it matters, and what “fixed” means to them. Gate 3. Approve the plan — and make the agent list what it is not doing, which is how the missing backfill for existing rows surfaces before it ships. Gate 6. “If this breaks next month, what turns red?” No answer means it's a demo, not a test. Gate 9. Ship to the customer who asked first, tell them, then widen.

A one-line copy fix, run through the same skill, earns two gates and ships in minutes. That difference is Gate ∞ doing its job.

status and limits

as stated by the source

The asset — the gates reference and the operating skill — the business-context template, and the pipeline snapshot with its compose-your-own guide are all in the repo. The source's own status note lists two things still ahead: glue to read the Foundation's voice and taste profiles, and a final package pass. This page does not claim those are done.

The kit doesn't decide for you. It surfaces the call, applies the heuristics, puts the agent to work making the decision defensible, then stops. Use it to work each gate and you level up on the lens you're missing. Use it to skip the gate and you just fail faster, with nicer tooling.

PRs that remove a gate or auto-approve one will be declined: the gates are the asset.

where to go

after the install