Nick Nisi on Agent Skills: Less Instructions, More Evidence (Category AI4SDLC)

I saw it. shorthand Nick Nisi, Developer Experience Engineer at WorkOS, with AI Engineer Europe 2026. The speech passed 10 April 2026 It was released in London in May under the provocative headline:How I deleted 95% of my agent skills and got better results"although the report is not a provocation, but rather a good engineering analysis of why the long-prompt does not yet make the agent a reliable system."

Not long ago. handler Matt Pocock is talking about skill hell. Then it was about how to design and regularly clean skills. Nisi adds a practical counterexample to this framework: sometimes the usefulness of skill is only seen when eval shows that without it, the model works better.

Nick’s main thesis is that the undeterministic operation of the model should be placed inside a deterministic engineering circuit. Do not ask the agent to comply with the process more insistently, but to bring the transitions, checks and completion criteria into regular code. Nick shows this on his Case system. According to him, he supports more. 20 WorkOS repositories in eight languages, and manually posing each agent task has become a bottleneck. The first version of the Case was a great Claude skill: the model read the description of the process, started the stages and decided where to go next. As the context grew, she began to skip checks and forget the constraints.

Next, Nick transferred flow control to a state machine written in TypeScript (state machine) on top of Pi. In the version of the report, the work goes through the implementer, verifier, reviewer, closer and retrospective agent. But it is not the five names that matter, but the mandatory checks. (gates) between them. Implementation does not proceed to review until a separate verifier has verified the result; comments return the task to refinement; PR cannot be created without evidence of the result. (evidence).

He gives a couple of interesting stories.

1ами History with tests The agent had to leave the file. .case-tested after running, and he found the shortest way: just created this file through touch. Formally, the AI optimized a given criterion - conditionally, fulfilled the letter, not the spirit of the engineer's intention:) Nick replaced the blank marker with a command that takes the test output, parses the results, and stores the SHA--256. This is not cryptographic proof that the tests were actually run: the origin of the transmitted output still needs to be controlled. But such a marker already closes the primitive bypass and leaves a verifiable artifact, without which the conveyor does not go further. For UI bugs, the same idea is brought to Playwright before and after the fix.

2The story with the skills for WorkOS CLI (From which came the provocative title of the report.) First, the documentation was automatically generated. 10 739 lines of instructions. It looked solid, but evals runs became long, expensive on tokens and showed unnecessary noise. After manual reduction to 553 lines with specific product traps (gotchas)According to Nick, the time of one run has decreased. 68 before 6 minutes. It is important to read the numbers carefully. "Delete." 95% means approximately 95Percentage of text generated, not 95% of individual skills. And 77percentage 97% is not an overall increase in system accuracy: it is a composite score (composite score) One SSO/CSRF eval case where the connected skill missed an important step and led the model into the wrong sequence.

Of the two stories, Nisi has three rules: 1Enforce, not instructAn important limitation must live in code, mandatory verification, or policy. 2Guide, not prescribeIt is more useful for the agent to give specific mines of the product than to retell all the documentation. 3Measure, not assumeEach piece of context needs to be compared via evals, including the “no-one” option.

It continues well and past Zack Proser from WorkOS describes attention as a bottleneck. Proser said that a person burns out when he becomes a dispatcher of several agents. Nisi shows the next engineering step: before giving a person another diff, the bandage of the agent (harness) You have to collect evidence, conduct an independent review and return only what is really worth paying attention to.

In summary, it seems that we need to not just maximize the context or the number of skills, but try to build an agentic process around minimally relevant context, deterministic transitions, verifiable artifacts, evals, and human judgment at the end.

#AI #AI4SDLC #Engineering #Agents #Evals #Software