AI-Enabled Creative Operations

The problem it solves

Most teams adopt AI the same way: individuals start using it privately, everyone develops their own prompting habits, and output quality becomes a function of who happened to write the prompt. Six months later the team is faster and less consistent, which is the worst of both.

The failure is that the knowledge stayed in people's heads. A prompt that works lives in one person's chat history and dies there, and the same corrections get made every week by different people who don't know the others made them.

Automation on the front end, human gates on what matters

The design principle is a split, and getting it wrong in either direction is expensive.

The pattern in the right column: every item is a judgment call with consequences the model can't feel. Deciding whether a fix is good enough is the work. An accelerating pipeline can't make that call, and a team that hands it over stops having a point of view.

What a skill contains

A skill isn't a prompt. A prompt is an instruction; a skill is a small operating manual with the reference material attached and a defined way to improve.

• A trigger description. When this should be used, written broadly enough that it fires in the situations you'd want it to and not otherwise. Getting this wrong is the most common reason a good skill goes unused.
• A resolution step. Before producing anything, the skill states its assumptions (audience, funnel stage, boldness tier, brand, theme) so a person can correct a setting instead of rewriting the output. Half of all voice problems are a piece written at the wrong setting, not written badly.
• Two modes. Write mode produces. Check mode grades an existing draft, states what it inferred, scores against a rubric, and rewrites only the worst lines. Check mode is where the value compounds, because it works on other people's drafts.
• Reference files. The voice guidelines, the proof library with hedging rules, the placement caps, the inclusive language rules. Loaded conditionally so the skill reads what's relevant rather than everything.
• Hard gates. Rules that block output regardless of quality score. Legal exposure, accessibility requirements, expired statistics.
• A feedback loop. The part most teams skip, and the part that determines whether the skill is alive in a year.

The feedback loop is the whole thing

A skill improves by being corrected. When someone says "no, we don't say it that way," that correction is the most valuable thing they'll produce that week, and it dies in the chat unless it gets logged.

Worth logging: someone corrects the same thing twice in a session; someone overrides a setting and gives a reason; someone says "we don't say it that way" about something not in the skill; someone asks a question the skill can't answer.

Two rules about how changes happen:

• Never edit reference files from a live session. All changes go through a scheduled review. A skill anyone can silently change is a skill nobody can trust.
• Version it and keep a changelog. Including the known gaps. A skill that documents what it can't do yet is more useful than one that pretends to be complete.

Workflows worth building first

Content decay detection

An automated pass identifies pages losing position or going stale, drafts the update against the current messaging architecture, and routes it to an editor to polish and publish. The team stops choosing between net-new content and maintaining what already ranks, a choice most teams resolve by quietly abandoning maintenance.

The gate: a human decides what actually gets worked and what ships. The automation only ever produces a queue and a starting point.

Design to MVP stage

Projects arrive roughly 80% complete before anyone touches them, which changes what designers spend their day on. Instead of producing the fiftieth variant of the same layout, they're refining the things that need judgment. This is also the best defense against the commoditization of design work. A designer who owns the why behind a decision is doing something the pipeline can't.

Voice as an invokable standard

Once voice is calibrated, encode it. Every piece of copy across the team resolves against the same tier ladder and cap table, and check mode grades drafts before they reach review. Consistency stops depending on who's writing.

Structured content for answer engines

Search increasingly resolves to synthesized answers rather than a list of links, and the levers for being cited are different from technical SEO. What matters is expert citation, third-party proof, customer stories, and content structured so a model can extract a claim with its attribution intact. That's brand and content work, not a web team task.

What this doesn't do

Worth being direct about the limits, because the category is full of overclaiming.

• It doesn't produce a point of view. Convictions come from customers and from people who've been in the room. A skill can enforce a position; it can't have one.
• It doesn't fix a missing strategy. Faster production against no strategy produces more of the wrong thing, sooner. This makes an existing system faster; it doesn't substitute for one.
• It doesn't reduce headcount on its own. The realistic claim is that a lean team ships at a level that used to require a larger one, and the work left over is more interesting.
• It requires real investment. Expecting tech-company output while funding the lowest tier of tooling is a choice, and the results follow the choice.

How to know it's working

• Someone who didn't build the skill uses it without being told to.
• Review comments start citing the standard rather than personal preference.
• New hires produce on-voice work in week one instead of month three.
• The correction log has entries from people other than the skill's author.
• The team's floor rises. Ceiling still comes from people.