The Dark Arts of Skill Engineering — Paul Bakaus, Renaissance Geek
Source
youtube.com
Author
AI Engineer
Date
Context
An agent skillA reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.Full definition → can include executable checks as well as written instructions. The publisher's description of this workshop presents Paul Bakaus's approach to making design guidance part of the software that runs an agent. One example is a agent hookA script the harness runs automatically at a fixed point in an agent's loop — before a tool call, after an edit, on session start.Full definition → that blocks a write when a check fails.
That changes when a rule takes effect: the workflow can stop an action at the point of execution, instead of relying entirely on the model to remember prose instructions. The description frames hooks as part of a broader principle that a gate which can be skipped will be skipped. It presents a design approach, without supplying measured evidence that the techniques improve design quality.
Key takeaways · AI-distilled
Paul Bakaus argues that banning an overused AI-design pattern doesn't produce originality — it just relocates the model to the next-nearest cluster in latent space, since 'slop is a moving target' driven by the model's statistical median.
One technique: run two sub-agents blind to each other's work, one as design director and one as a deterministic linter, because a single agent grading its own output tends to just say it did well.
Another technique forces divergence by having the model name its top three font choices, then discarding all three to avoid the 'safe prediction' each one represents.
Bakaus recommends emitting instructions on stdout rather than burying them in prose, since models follow structured outputForcing a model's response to match a schema, so downstream code can parse it instead of guessing at prose.Full definition → more reliably — though this comes at the cost of losing prompt cachingReusing the model's processed form of a repeated prompt prefix so subsequent calls skip re-reading it, cutting cost and latency substantially.Full definition →.
The underlying discipline across all nine techniques: enforce with hooks that block a bad write outright rather than complaining after the fact, because 'if a gate can be skipped, it will be.'
Terms in this piece · Glossary
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
agent hook — A script the harness runs automatically at a fixed point in an agent's loop — before a tool call, after an edit, on session start.
structured output — Forcing a model's response to match a schema, so downstream code can parse it instead of guessing at prose.
prompt caching — Reusing the model's processed form of a repeated prompt prefix so subsequent calls skip re-reading it, cutting cost and latency substantially.
videoReal AI Agent Attacks and What Would Have Stopped Them — Andrey Rybka, Bloomberg
videoInside Arca's Long-Horizon Agents That Run the Business - Aman Singh, Arca
videoWhy Compute Still Trades in Group Chats — Affaan Mustafa & Alejandro Soumah Fernandez, Jump Trading
videoHow OpenAI Built ChatGPT for Bankers and Analysts — Joseph Kim, OpenAI
