Why Stackkeel Exists

For thirty years, the winning move in software was less process. Small teams beat big ones. Lightweight beat heavyweight. Every methodology that added coordination steps eventually lost to one that removed them, because process overhead is human coordination cost, and human coordination cost dominates almost everything else.

AI coding agents quietly broke that rule.

The cost structure flipped

When an analyst pass, an architecture review, a written design, and an independent QA verdict each cost a meeting, you skipped most of them and shipped. When each one costs an API call, skipping them stops being thrift and starts being negligence. The practices we abandoned as too heavyweight, staged reviews, written handoffs, explicit verification, are suddenly the economical choice. The price chart changed. Most of our instincts have not caught up.

Stackkeel’s author spent three decades as a CTO on the “less process” side of that fight. Then an agent flagged a missing file during a routine bug fix, something entirely outside the framing of the request. The agent that flagged it wasn’t writing code. It was running the first phase of what became a six-phase pipeline.

That observation became a workflow, the workflow became conventions, and the conventions, refined across several real products, became this kit.

What that means in practice

Artifacts are the substrate. Work logs, decision records, and design docs are not overhead here. They are how stateless agents hand work to each other, and how a project remembers anything at all. Codified conventions are durable; uncodified ones are not, because the agent has no working memory.

Roles create pressure. The same model prompted as an adversarial analyst finds different problems than the same model prompted as an implementer. Stackkeel’s six-phase pipeline is role-shaped pressure applied on purpose: analyst, architect, tech lead, implementer, QA, and a final shipped-versus-intent verdict.

Enforcement beats intention. Conventions that live in a prompt evaporate. Conventions that live in git hooks, tripwire scripts, and CI hold no matter which assistant, or which human, is typing. That is why the kit’s gates are mechanical: commit grammar, work-log-before-code, secrets scanning, audit coverage, brand scope.

The bottleneck moved. With execution cheap, framing is what fails. A bad assumption in phase one cascades confidently through everything downstream. The pipeline exists to catch bad framing early, while it is still cheap to fix.

The thesis, executable

Everything above ships in the repository as working machinery rather than advice: the pipeline in AGENTS.md and the agent roster, the memory in docs/, the enforcement in scripts/ and CI, and the lifecycle that keeps projects and kit exchanging improvements. Stackkeel was built, reviewed, tested, and released by the same workflow it contains.