Building Reusable AI Workflows, Not One-Off Prompts

A prompt gets you one good answer. A workflow gets you that same quality of judgment every time you need it, not just the one time you got lucky with your wording. That’s most of what “AI direction” means in my work, and it’s easiest to explain with two things I’ve actually built.

Two skills, one habit

The accessibility QA agent and the small-business website audit skill look like they solve different problems. One checks a site against WCAG. The other reviews a small-business site for clarity, conversion, and trust. But under the hood they run on the same three habits:

1. Check the real thing first. Each skill runs a script against the actual HTML before any judgment call gets made, so the starting point is evidence, not a guess.
2. Sort every finding, don’t just list it. Confirmed issue, warning that needs a second look, or something that still needs a human to check by hand. Nothing gets reported as fact unless it was actually verified.
3. Say what wasn’t tested, not only what was. Neither skill claims full compliance just because an automated pass came back clean. That limit is built into the skill itself, not tacked on afterward as a disclaimer.

That third one is the habit I care about most, because it’s the one an AI system won’t enforce on its own. Left alone, a model will happily tell you a site “passes accessibility” because a script found zero errors. The skill exists specifically to stop that sentence from getting written.

Why this counts as a communications skill too

Every one of those three habits is really an editorial habit. Cite your evidence. Keep confirmed fact separate from your best guess. Say where your limits are instead of letting silence imply certainty. That’s the same discipline I was trained on for copywriting, now applied to how an AI agent gets instructed and double-checked.

It’s also why I build skills instead of asking for one-off outputs. A one-off prompt only carries the judgment I put into that single request. A skill carries the judgment forward every time it runs, whether it’s auditing a nonprofit’s donation page or drafting a coaching client’s messaging around health claims that could create real legal exposure if they’re overstated. That second case wasn’t hypothetical. I used this same evidence-first, say-what-you-can-actually-say approach a few weeks after building the QA agent, and the stakes were a lot higher than an accessibility report.

So that’s the actual offer here. Not “I used AI to write this,” but “here’s a process you can hand the same job to twice and trust it both times.”