We Replaced Our SOP Docs with AI Skills (Starting with the One Nobody Wanted to Do)

This year the whole cycle, for every employee, took three days. Not because we lowered the standard, but because we stopped treating the evaluation as a document and started treating it as a skill. That shift, from SOPs that describe work to skills that do it, has quietly changed how Talas runs.

The problem with SOPs

Every company has them. The standard operating procedure, the process doc, the wiki page that explains exactly how a task is supposed to be done. And every company has the same problem with them: a document can only describe the work. It can't do any of it.

So the SOP sits there. Someone still has to read it, remember it, and grind through every manual step the same way every time. For something like a performance evaluation, that means the quality of the outcome depends entirely on how much energy whoever's stuck with it has left. The doc is a description of good work, not good work itself.

What a skill actually is

The way I think about a skill is like a tool in a toolbox. It sits there until a specific job calls for it, and then it does that one job well. The difference from an SOP is the verb. An SOP tells a person how to do a task. A skill lets Claude actually do the task with you.

But the part that makes me trust a skill isn't the automation. It's the barriers. When I build one, I set limits and I test the output, because I want to understand how the skill gets to its answer, not just that it produced one. I'd rather know where it's solid and where it's reaching than get a clean-looking result I can't verify. That one principle is what made the performance evaluation skill work, and it's worth seeing up close.

The one that started it: performance evaluation

The reason evals were the right place to start is that they're a logic problem wearing a human costume. Most of the month wasn't judgment, it was process: matching answers against criteria, consolidating reviews, keeping scoring consistent from one person to the next. That's exactly the kind of structured, repetitive reasoning a skill is good at.

So I built one. It helps craft the criteria, works through the interview and transcription, then evaluates each answer against the criteria and offers a judgment. The whole cycle, for every employee, now takes three days instead of a month.

But here's the part I'm proudest of, and the reason I actually trust the output. The skill doesn't just hand me a score. It hands me a score with a confidence level. It tells me, in plain percentages, where it's standing on solid ground and where it's making an assumption. A high-confidence read on whether someone met a clearly defined criterion is something I can lean on. A lower-confidence read is a flag that says "look closer here, I'm reaching."

That changes the relationship entirely. Claude isn't making the decision. It's showing me its work and pointing me to exactly where my own judgment is needed. And there's a lot it can't see: the relationship and communication skills I've watched up close, whether someone aligns with our culture, the historical performance and growth I've witnessed over years. The skill handles the structured evaluation so I can spend my energy on the human judgment only I have. Final say was always mine.

What's telling is where the time actually went afterward. With the consolidation and scoring handled, the only real bottleneck left was conducting the interviews themselves. The one genuinely human part of the whole process, sitting across from someone and having a real conversation, became the thing that took the most time. That's exactly as it should be. The skill didn't replace the judgment. It cleared away everything that was getting in the way of it.

That one started as a skill and grew into a full system through Claude Code. Which is usually how it goes.

It wasn't just one

Once the pattern clicked, it spread. A few of the skills running at Talas now:

• Lead-list refinement, which cleans and qualifies raw prospect lists against an ICP so our outreach starts from a sharp list instead of a messy one.

• UGC and influencer matching, which pairs the right creators to a brand instead of doing it by gut and spreadsheet.

• The business stress-test, a skill Marcelo built to pressure-test our own strategies and deliver harsh, unsentimental criticism, so we find the weaknesses and risks in a plan before the market does.

Different jobs, same shape every time: take the tedious, structured part of the work, hand it to a skill with clear barriers, and keep the human on the decisions that need a human.

The turn we didn't expect: building skills for clients

The interesting part is what happened next. Once we'd built enough of these for ourselves, we realized the skill itself was something we could build for clients, not just internally. That's now a Talas Labs offering.

A recent one: we helped a retail client build a skill that evaluates their inventory, flags styles that should go on discount, and handles uniform issuance decisions. Work that used to be manual judgment over endless spreadsheets, turned into a tool their team can run on demand. The skill encodes the logic once, applies it consistently, and frees the team for the calls that actually need a person.

That's the moment the whole thing stopped being an internal efficiency trick and became a service.

What this actually means

The lesson underneath all of it is simple. An SOP describes good work. A skill does it with you. The first is a document that depends on someone having the energy to follow it. The second is a tool that does the structured part reliably, shows you where it's confident and where it's guessing, and leaves you the decisions that need real judgment.

That's what AI-native operations look like in practice. Not replacing people, but building tools that take the tedious 80 percent so people can spend their attention on the 20 percent that actually moves things.

If you want a marketing partner that builds systems like this, for your operations as readily as for our own, that's what Talas Labs does →.