One Sentence Worth Thousands of Lines of Prompts

I recently read an AI workflow write-up from a developer in China. The author joined his company recently, doing Agent development while maintaining backend services. Because he delivered fast and well, he became the primary owner of several projects, and his boss scheduled a one-on-one specifically to ask about his methods — which led to this write-up. It has five parts and is information-dense. I treated it as a specimen: verifying every tool it mentions, then examining what it gets right.

Fact-Checking First

Every skill and tool named in the article really exists, and the descriptions are accurate. Pinning down the ecosystem coordinates: the think skill he uses for plan alignment and the check skill for code review come from tw93's Waza repository, 7,300 stars; grill for requirement clarification, implement for implementation, and handoff for cross-session handoffs come from Matt Pocock's skills collection, 295,000 stars; ponytail, which cleans up AI over-engineering, is an independent project with 166,000 stars, created only this past June; and ego lite, the open-source browser for end-to-end testing, has 17,600 stars. One author's perspective, but everything cited is battle-tested in the ecosystem.

The context checks out too: when OpenAI released GPT-6 Astra, they did publish an official migration guide. Its core message in one line: shift from *teaching the model how to work* to *telling the model where the boundaries are and what completion looks like*, and clean outdated constraints out of AGENTS.md and skills. The model vendor itself is urging developers to subtract.

The Rise and Retirement of Heavy Process

The article's thesis is the inflation and retirement of constraints.

A year ago, Jesse Vincent's Superpowers pushed skill engineering into the spotlight: a complete development methodology built around Spec and TDD, heavy on process and constraints, designed to stabilize the models of that era. The repo now has 301,000 stars and is exactly one year old. It worked — for the models of that time.

The problem is the models caught up. The article puts it bluntly: even when the context is sufficient and the model could act directly, workflows like Superpowers make it walk through every prescribed step. Tokens burn fast; output doesn't grow. After the GPT-5.6 generation shipped, once-useful constraints started becoming noise.

The most interesting evidence comes from skill authors themselves. tw93 wrote: "A skill's value lies in what it lets the model do, not what it saves the model from doing. Install less, hold loosely, leave judgment to the model." The person selling skills is telling you to install fewer skills. This is the same message as OpenAI's official guide: the day a new model ships is the day to clean up your prompt debt.

Cognitive Shortcuts

The most information-rich section of the article is his set of short prompts: first principles, adversarial review, ablation experiments, Occam's razor, high cohesion / low coupling. Each is one sentence — not an incantation, but a precise direction of attack. Have the model stop marching down the preset plan of "add a Redis cache" and re-analyze from first principles why the API is slow; don't let it argue for the plan — prioritize counterexamples that could overturn the core assumption; three optimizations shipped together — design an ablation experiment to see where the gains actually come from.

This observation deserves elaboration. These methodology names work in one sentence because they are highly standardized in human knowledge: decades of papers, textbooks, and code-review culture have polished them into common nouns. The model saw massive amounts of related material in training; the name itself is an index key — invoke "Occam's razor" and the whole apparatus of judgments about deletion, assumption counts, and complexity costs activates. Hardcoding a workflow of "step one: analyze; step two: check assumptions" discards what the model already knows and re-teaches it with your own few hundred lines.

So what should constraints say? The article's operational answer: specify thinking directions with nouns; frame task details with boundaries and acceptance criteria. Neither absent nor overreaching. This aligns with a June study: removing three components from an evolving agent *improved* its performance — averaged instructions blur directional signal in text space. Human workflow and automated evolution converged on the same rule.

The Unspoken Half

One observation deserves to be pulled out on its own: the bottleneck has moved from writing code to reviewing it. Agents can run several tasks in parallel, but a human's speed at understanding requirements, judging designs, and confirming delivery doesn't multiply along. Writing more just means a bigger backlog of code awaiting review.

His response is to layer the judgment: problems that type checks, tests, and business assertions can catch — let the agent close the loop itself; human attention is spent on only three things — whether the requirement is understood correctly, whether the key business logic holds, and which risks remain unverified. Test environments, read-only databases, and log traces are all exposed as agent-callable tools so it can find its own failure evidence before handing anything over.

One step deeper is the closing insight: workflows themselves have a half-life. Some steps exist to patch a particular model generation's weaknesses; when the model changes, those steps' value needs re-auditing. The author's method is periodic subtraction — after Astra shipped, he deleted a batch of overly strict constraints from his AGENTS.md. The environment moves; a constraint's shelf life is indexed to the model's version number.

One year, 301,000 stars, and the line "install less, hold loosely." Both facts being true at once is what makes this specimen worth recording.