The Fastest Way to Build a Skill Is to Stop Describing It
While we were finishing module 5 of the Freedom Ag training, a new item appeared in the Claude Skills library: Skill Creator. The description was short. You record your screen doing something you already do regularly, hand Claude the video, and it builds a skill based on what it watched.
We had just spent several weeks doing that same job by hand.
What We Built by Hand
The five skills we shipped for Freedom Ag were designed around a specific constraint: the people using them are ag sales reps, not developers. Each skill needed an interview gate before any output, guardrails against fabricating prices or product rates, and structured outputs matched to how the work actually gets used.
Writing those skills by hand meant articulating things that are easy to do and hard to explain. What does a good SPIN selling question look like for a grower conversation specifically? When should the influence audit call something manufactured versus genuinely true? The prompts in the finished skills are precise about these distinctions, which is why they work.
But precision requires articulation. You have to be able to say what the rule is before you can write it down. That's the step that's hard.
What Skill Creator Bypasses
Recording your screen bypasses the articulation step entirely.
There are workflows most people do well but cannot describe cleanly. The sequence of actions that goes into processing a set of price sheets. The way you triage an inbox before you respond to anything. The order in which you pull data from one system before updating another. These are process knowledge that lives in the hands, not in words.
Asked to describe it, you'd give a rough outline. Asked to record it, you'd just do it. The recording captures the actual sequence. Claude observes the video and writes the skill from what it saw rather than from what you said you do.
The two inputs are different. What people say they do and what they actually do are not always the same thing.
When to Use Which
The hand-crafted approach is still right when the skill needs strong guardrails. Our ag skills have explicit rules about never filling a gap with an invented rate, always flagging placeholders, and asking before stating any number not in the source material. Those rules require deliberate writing. A skill built from a screen recording does not automatically include guardrails you haven't built in.
The screen recording approach is right when you're trying to capture a workflow you haven't managed to document. If you can do it faster than you can explain it, record it first. You can edit the skill Claude produces to add guardrails afterward. Starting from a video of real work is almost certainly better than starting from a blank document.
We're planning to test this with a few of the more process-heavy workflows we build for clients. The screen-recording input should be particularly useful for things like portal navigation, multi-step data entry, and anything where the step sequence is what matters and the logic is already correct.
The articulation bottleneck slows down skill building more than anything else. Removing it is not a small thing.