The Translation Was Really Bad
If your AI output is inconsistent, the reflex is to blame the model. Most of the time it is not the model. It is the missing contract — the glossary — between your product and the language you are generating.
I gave the agent a translation job without a glossary. The output was technically Persian and emotionally wrong. Verb conjugations mismatched their subjects. Terms were invented instead of chosen. Sentences closed in ways Persian sentences cannot close. When I read it back, I asked the obvious question: do we need to create a full dictionary for you to translate properly?
Without a dictionary, every translation pass is a first draft — including the one the model already forgot it wrote.
The Dictionary Is the Product Decision
Terms like role, skill, workflow, and agent had to be locked before a single sentence was translated.
The glossary is not a style guide. It is a contract between the product and the language. Agent is ایجنت. Skill is مهارت. Role is نقش. Workflow is ورکفلو. Session is جلسه کاری, never سشن. Once those are locked, the sentences can be rebuilt honestly.
This is the meta-lesson for anyone running a business with AI: consistent output does not come from a better model. It comes from a better contract. The dictionary is that contract.
A glossary is a decision log. Every entry records not just the right word, but the wrong words you must never use.
Translate the Meaning, Not the Grammar
English copy is context-free. Persian needs the context to be meaningful.
The skill codified what my best human translation instincts already knew. English removes context; Persian demands it. A bare "I spent ten years building systems" is meaningless in Persian — systems for whom? The Persian translation has to make the referent explicit even when English leaves it implied.
That is not a license to invent. It is a license to complete. Every added phrase must be supported by the source, and the result must stand alone — a Persian reader should never need the English to understand it.
Recover the meaning What is the source really saying? What context does Persian need that English omits?
Define the message Write the meaning plainly before translating anything.
Render it in Persian Now translate the message, not the grammar. Verbs before nouns, outcome before architecture.
If a Persian reader has to reverse-translate your sentence to understand it, rewrite it.
Plural Verbs, Concretely
The mechanical rules that catch the most embarrassing errors.
Persian subject-verb agreement was the biggest recurring bug. ایجنتها چی فراهم میکنه is wrong — plural subjects take plural verbs. The skill now mandates it: skills provide, agents execute, evals run. Every plural subject gets a plural verb, even when the English treats a product name as singular.
Good translation sounds like a smart Iranian founder speaking to another intelligent person. Anything else gets rejected.
The quality test is brutal on purpose. Does it sound like something a founder would actually say? Would it make sense without the English? Is every word doing work? If any answer is no, the translation is rewritten.
Build the Dictionary First
Before you generate a single word with AI.
Lock the nouns. Every product object — role, skill, workflow, memory, agent — gets exactly one term and a list of rejected alternatives.
Write the grammar rules. Plural agreement, verb-first construction, context completion. Codify the mechanical rules so they are enforced, not hoped for.
Test against reality. Read every output as a native user. If it sounds generated, it is rejected. The dictionary makes that rejection reproducible.
AI fails without a glossary because generation is a decision, not a rendering. The dictionary is the decision.
The dictionary fixed what the model could not.
Locked terms, explicit rejects, and a quality test that rejects translated-sounding Persian. This is how the whole site speaks one language.
Read about the RTL design system