Every team that writes prompts eventually owns a folder of half-copies. One lives in a chat thread, one in a Notion page, one in a system prompt pasted into production by whoever deployed last. They drift apart within a month, and nobody can say which version produced the good output in the demo.
An AI skill file solves that by treating the prompt as a versioned artifact. MindFluence is one: a marketing skill that ships as a single SKILL.md, with the tables, examples and checks inlined instead of spread across eight include files that a model has to remember to open.
Three rules make that file survive a model swap, a context limit, and a reviewer who wants to know why a sentence exists.
Rule 1: the core has to work alone
A skill that depends on external reads breaks the moment the runtime does not implement them. Some clients load one file and stop. Some truncate the folder listing. Some will happily fetch a reference document and then forget it two turns later.
MindFluence inlines the parts that every task needs: a 38-row bias selection router, 14 pre-built bias combinations, 13 anti-patterns each with a detection rule, seven tones with their narrative minimums, and a post-generation checklist. A model can run a normal request with no file reads at all.
The reference folder still exists for depth, with a longer decision matrix, a cultural adaptation table, and 13 scenario playbooks for launches, pricing pages, cold outreach and crisis replies. The rule is that depth is optional and the core is not.
The trade-off is size. A self-contained file is long, and a long system prompt costs tokens on every call. That cost is predictable, which is worth more than a cheap prompt that silently degrades.
Rule 2: the checks have to be mechanical
“Make the copy more persuasive” is not something a reviewer can audit. “Does every authority claim carry a name and a year” is.
Each of the 13 anti-patterns in the skill carries a detection rule written to be run like a search, not like a judgement call. Vague social proof fails when “thousands” or “many” sits next to a customer claim. Fear without an exit fails when a threat and its remedy land in different paragraphs. Fake scarcity fails when a limit has no stated reason.
After the draft, the model runs seven checks and labels every bias it used. The labels are the point. They turn an argument about taste into an argument about whether the second paragraph is doing the job assigned to it.
Rule 3: provenance beats flourish
The most expensive failure in AI copy is an invented fact. A fabricated customer, a study that does not exist, a percentage with no source. It reads perfectly, it ships, and it costs the client credibility that takes a year to rebuild.
The skill treats provenance as a gate. Every number has to trace to the input, to a named source, or to an explicit placeholder. When no real case exists, the model writes a labelled [illustrative] example and says so, instead of inventing a person with a job title and a revenue figure.
That rule also fixes the structural weakness of data-only copy. A page made of aggregate statistics informs the analytical reader and leaves the rest cold. One named account, even a small one, carries further, and the skill requires the narrative to be either real or visibly marked.
What this looks like in practice
Four modes cover the work. Quick mode runs the full pipeline for a normal request. Deep mode asks five clarifying questions first, which is what a strategy brief deserves. Audit mode reads existing copy and names the biases and anti-patterns present. Optimize mode iterates on underperforming copy and produces variants for a test.
The file is plain Markdown and MIT licensed, so it drops into any client that accepts a system prompt. Nothing in it is vendor specific. That is the portability test: if the file only works inside one product’s dashboard, it is a config file, not a skill.
When the skill needs to know which language it is writing for, a detection call is cheaper than asking the user. The text language detection guide covers the thresholds, and the readability metrics guide shows how to check the result lands in the band you planned.
Background on system prompts: OpenAI prompt engineering guide and the Model Context Protocol specification for clients that load skills as tools.