Why this exists

Most teams try to improve AI output by improving prompts. That works for a while, then stops. The reason it stops is that prompting is the last mile — the failures that persist after good prompting are almost always failures of the information environment the agent is working in: undefined terms, no clear source of truth, no record of where facts came from, and no way to look things up in the way the question actually requires.

This SOP addresses those. It is ordered by leverage. Step 1 will fix more output quality than steps 4–6 combined, and it costs the least.

A note on scope: this is written for teams where humans are still in the loop and one person (or one agent session) works on a thing at a time. That describes almost everybody right now. Step 6 is about what changes when that stops being true, and can be skipped until it applies to you.

Step 1 — Define your vocabulary before you write another prompt

Why

An agent will faithfully use whatever term you give it and quietly treat two words for the same thing as two different things. This is the single most common cause of output that looks complete and is wrong.

Examples of the failure:

• One person's brief says "revenue," another's says "net sales," a third says "shipped revenue." The agent summarising all three produces a number that reconciles to nothing.
• A retail team calls the same item a SKU, the marketplace team calls it an ASIN, and paid social calls it a catalog item. An agent asked "which products underperformed" answers for whichever slice it happened to see.
• "Conversion rate" means sessions→orders on a DTC site, glance views→purchases on a marketplace, and click→purchase in an ad platform. Same phrase, three denominators, no warning.

The problem isn't that the agent lacks intelligence. It's that nothing in its inputs tells it these are the same concept, or that these are different concepts wearing the same name.

How

Create one plain-text file. Call it whatever you like — vocabulary, entity glossary, definitions. Load it into every project, skill, or agent context where analysis or reporting happens.

It needs three sections.

Entities — the things that exist in your business, with one canonical name each and their aliases:

Product      — canonical: internal SKU             
aliases: ASIN, GTIN, variant ID, catalog item ID, MPN
Campaign     — canonical: platform campaign ID             
aliases: campaign name, initiative, "the launch"
Channel      — allowed values: DTC web, marketplace, retail, wholesale,paid social, paid search, email, affiliate, retail media
Customer     — canonical: email hash             
aliases: account ID, subscriber, lead
Creative     — canonical: asset ID             
aliases: ad, image, video, listing image, hero shot

Metrics — every metric with its formula, and explicitly, its denominator:

Conversion rate (DTC)         orders ÷ sessionsConversion rate 
(marketplace) purchases ÷ glance views
Conversion rate (paid)        purchases ÷ link clicks                             
 NOTE: these three are NOT comparable.                             
Never place in the same column.
Revenue (net)                 gross sales − returns − discounts
Revenue (gross)               gross sales, pre-return                              
DEFAULT: when unqualified, "revenue" = net
ACOS                          ad spend ÷ ad-attributed sales
TACOS                         ad spend ÷ total sales (all channels for that product)
ROAS                          revenue ÷ spend, platform-attributed                              NOTE: attribution window must be stated