Prompts forget. Skills remember.
Twelve reusable skills, written in plain English by a business analyst with no engineering background. A compliance job that took days now runs in minutes, and the best moment was the machine refusing to guess.
If you use AI at work, there is a job you have explained to it more than once. Probably more than five times. The same rules, the same exceptions, the same "no, format it like this", typed fresh into an empty box because the chat you trained yesterday has vanished and today's one knows nothing.
I call it the prompt tax, and for a year I paid it like everyone else. This is a field note on how I stopped: twelve reusable skills, each one a written procedure the AI follows on demand, built in spare half-hours by someone whose job title says business analyst, not engineer.
The prompt tax.
Chat AI is brilliant and forgetful. The memory features that do exist hold preferences, not procedures: they'll remember you like short answers, not the eleven rules and four exceptions that govern a regulated document. So the rules live in your head, and you re-type them under time pressure, and some days you remember ten of the eleven. For casual work that's fine. For work where a wrong tick in a grid has real-world consequences, it is quietly terrifying.
A skill fixes this by moving the rules out of your head and into a file. It is nothing more exotic than a standard operating procedure written for a machine: the job, the rules, the exceptions, the order of operations, what good looks like, and, crucially, when to stop and ask a human. Plain English. The AI reads it whenever that job comes up, every time, without being reminded.
The pattern is not tied to one vendor. ChatGPT has custom GPTs, Gemini has Gems; I build mine as skills in Claude. The discipline is the point, not the product.
Write it down once.
The build method is the same every time. Describe the job as if handing it to a competent new starter. Attach a real example file, not a description of one. Run it, correct what it gets wrong, fold the correction back into the instructions, save. That loop takes 30 to 60 minutes, and you never explain the job again.
The flagship in my library is a safety-critical compliance document, published on a rolling cycle. The manual version took several days per run: cross-checking line items against a rule set, correcting a specific class of false positives, re-paginating so no section splits across a page, stamping the update date, attaching the branded cover. The skill version does all of it in about five minutes, and the human's job shrinks to the part a human should own: reviewing the flags and signing it off.
Skills also chain. Another sequence in the library takes a working spreadsheet, flattens it into a system-ready format, enriches every line against product master data in the warehouse, then builds the import file for a downstream system, and a separate skill produces the branded, customer-facing output from the same source. That entire chain used to be manual rekeying between four tools. Nobody rekeys it now.
The most valuable thing my AI has done all month is refuse to do what I asked.
The day it argued back.
Here is the run that convinced me this approach belongs in regulated work. I fed the compliance skill a document and told it which rule set to apply. The skill noticed the document's own labelling contradicted my instruction, on a detail I had missed. A blind automation would have carried on and produced something compliant-looking and wrong.
Instead it stopped. It flagged the mismatch, listed the three exact line items that would change under the other rule set, altered nothing, and asked me to confirm. That behaviour was not luck. It was a line in the skill: when the source contradicts the instruction, stop and ask. Guardrails in this pattern are authored, not hoped for.
This is the part the hype and the doom both miss. The question with AI in serious work is not "can you trust it?", it's what did you write down? The machine carries the checklist so that memory is no longer the control. The human stays accountable, reviews the output and owns the decision. My short-term memory is honestly poor; the skill's is perfect. Between us, the document is safer than it was when it depended on either of us alone.
If you explain the same job to AI twice, write it down once, as a skill.
Would you like to find out more?
Happy to talk through how the skills were built, what they cost in time, or how the same pattern would map onto your repeat work. No pitch, just the workings.