AI Guides › Skills & Agents
Building A Skill Around A Checklist You Already Trust On Paper
By Nigel Guy · 2 min read
A checklist that's worked for years on paper feels like the easy case for
turning into a skill — the thinking's already done, you're just automating
the execution. That's usually wrong. A paper checklist works partly because
of judgement calls a person makes silently at each step, and those calls are
exactly what goes missing the moment you translate "always worked" into
"runs unattended."
The rule: a checklist that works reliably on paper is a different claim
from a skill that runs it reliably on its own — the gap between them is
almost always the judgement calls nobody wrote down because they felt
obvious.
The mechanism
- Run the paper checklist a few more times, watching yourself do it.
Not to check whether it works — you already know that — but to notice
every small decision you make that isn't written on the checklist
itself: skipping a step because a condition clearly doesn't apply,
reading a form differently depending on context, catching an
inconsistency by eye.
- List every judgement call you find, however minor it feels. "I
glance at the total and know if it's obviously wrong" is a judgement
call, and it's one a skill can't do the way you do it unless you decide
what to do instead.
- For each one, choose: turn it into an explicit rule, or mark it as
still needing a human. Not every judgement call converts cleanly into
a rule a skill can follow — some genuinely need a person to look at the
result before it goes anywhere.
- Build in a check for missing or ambiguous inputs. A person doing the
checklist notices when something's missing and pauses. A skill needs
that pause built in explicitly, or it will proceed on incomplete
information and produce a confident, wrong result instead of a stalled,
honest one.
- Test the resulting skill against the same cases that proved the paper
checklist worked. If those cases were good enough to trust on paper,
they're the right first test for the automated version — don't invent
new test cases before you've confirmed the old ones still pass.
What to skip
Skip assuming a checklist that's "always worked" will automate cleanly on
the first pass — the track record is evidence the checklist is sound, not
evidence the automation will be. And skip trying to convert every judgement
call into a rule; some genuinely don't reduce to one, and forcing it
produces a rule that looks tidy and quietly gets things wrong.
Guardrails
- The judgement calls you can't convert into rules are the actual
guardrails section of the resulting skill — write down explicitly which
steps still need a human, rather than letting the skill silently guess
at them.
- A checklist that's trusted because "it's always worked" was trusted based
on a person running it, with a person's ability to notice something odd.
That trust doesn't transfer automatically to the unattended version —
it has to be re-earned through the testing above.
- Revisit this build the same way you'd revisit any skill: if the
underlying process the checklist describes changes, the judgement calls
you mapped may no longer be the right ones.
All 751 AI guides · JulieMango plans from £17/mo