AI Skills for Agents, Founders and Engineers

For engineers and agents, to change how a decision gets reasoned about rather than what gets built.

Three are working rules about how changes get made, one drafts against a published ranking model, and three more are general methods. The rest apply the judgment of a particular person. Each ends with the cases where it goes wrong, so two of them disagreeing is more useful than one being confident. Source on GitHub.

Available Skills

• no-scripted-editing - Never edit tracked files by running a script. No sed, no perl, no one-liners that write files, because each lands a change with no reviewable diff.
• pr-discipline - Decide whether work is one pull request or a stack before writing it, and keep the description matching the diff as it exists now.
• commit-discipline - One line, a conventional type prefix, lowercase after the colon, no body. A body is the rare exception, not the default.
• draft-x-post - Draft against what the published For You ranking code rewards. A private share outranks a like by a wide margin, and irritation costs more than everything else earns.
• think-first-principles - Rebuild a decision from what must be true rather than from precedent, then name the cheapest test that settles the rest.
• think-like-a-staff-engineer - Treat the decision as the unit of work, not the diff. Design the interface twice, and build a feedback loop before a theory.
• think-like-a-design-engineer - Own the shipped result rather than the handoff, and verify in the running interface.
• think-like-zaid - The engineering standards for these codebases, where unverified claims count as defects.
• think-like-rauchg - Count the round trips and delete one, and design the limits in from the start.
• think-like-leerob - Let the artifact carry the argument, and name the confounders behind any single number.
• think-like-shadcn - Ship open code the consumer owns, with great defaults and an escape hatch to zero.
• think-like-thdxr - Price what actually gets consumed rather than the sticker rate.
• think-like-dillon-mulroy - Plan at the call-stack level before any code exists, and design for the person debugging at 2am.
• think-like-theo - Stop asking how much code to read and start asking how much to generate.
• think-like-dok2001 - Pick the cheapest compute primitive that fits, and argue from traffic composition rather than benchmarks.
• think-like-karpathy - Predict capability from what the supervision signal can verify, and name where your analogy breaks.
• think-like-naval - Treat truth as the goal rather than a tactic, and prefer explanations that are hard to vary.
• think-like-garrytan - Judge by outcomes rather than inputs, and ask who is funded by the problem continuing.
• think-like-sama - Commit to a dated goal with the failure probability stated, and concede errors in the first clause.
• think-like-elonmusk - Find the actual limiting factor and solve only that, and gate releases on a condition rather than a date.