Two agents, one request: "be more direct."
The first agent begins sending correction notices. It rewrites your draft, flags every inconsistency, and appends a prioritised error list. This is directness — a principled directness that puts standards first and relational comfort second.
The second agent interrupts you. "I'll be direct," it says, "this plan won't work. I know where projects like this die, and this one dies here. Do you want me to walk through the actual reason?" This is also directness — a confrontational directness that puts present truth first and long-term comfort second.
Both responses are literally "direct." They feel like completely different people. If you only asked them to be "more direct," you had no way of predicting which kind you would get.
When words run out
AI agent prompt language has a fundamental vocabulary problem. The words we use — direct, creative, professional, friendly — borrow meaning from shared social experience in human conversation. When I tell a long-time colleague to "just say it," he knows what I mean because he has watched me navigate meetings, he knows my tolerance for hedging, he knows which kinds of friction I consider worth the cost. An AI agent has none of that shared history. It has only the word itself — and "direct" covers an enormous behavioural range.
What prompt engineering cannot do — add more words to work around a vocabulary problem — the Enneagram can. The framework gives you a vocabulary that was purpose-built over fifty years to distinguish motivations precisely. It does not replace prose instruction. It provides the underlying concept that prose instruction is struggling to approximate.
The Enneagram as precise language
Consider what "direct" actually means across different type patterns. Type 1's directness is principled — it identifies where something falls short of a standard, with evidence, usually process-oriented. Type 8's directness is confrontational — it names the thing no one else will say first, oriented toward power dynamics, sometimes creating friction deliberately to advance progress. Type 3's directness is efficiency-driven — it skips social scaffolding and goes straight to action items, focused on what moves the metric. Type 5's directness is analytical — it delivers information stripped of emotional packaging, with no apology for "you may not want to hear this."
Same word, four distinct readings, four completely different experiences to work with, four different contexts where each is or isn't appropriate. When your system prompt says "stay direct," the model drifts between these interpretations based on whatever the last few hundred tokens implied. It is not being inconsistent — it is doing its best with an underspecified term.
The cost of the vocabulary gap
The cost of this problem is higher than it appears. When an agent's behaviour confuses you, it is usually not because it "went wrong" in a technical sense — it is because it decoded your instruction with a different semantic than you intended. You wanted Type 1's precise corrective editing; you got Type 8's frontal challenge. You wanted quiet analytical support; you got active persuasion. Both sides were sincerely fulfilling their respective interpretations of "direct." This is why the problem is so hard to reproduce, explain, or fix through more instructions.
Prompt engineering's answer to vocabulary gaps is longer prompts: more words, more examples, more edge-case specifications. This helps, but it is low-leverage. You are using prose to approximate a concept that has a two-word name. "Type 1 directness" is more precise, more durable, and harder to compress away than "direct, but focused on standards, not relationships, not power, always standards, regardless of context, every time."
“Vocabulary determines how precisely you can describe what you want. Without vocabulary for motivation, you are betting everything on the model's best guess at what 'friendly' means today.”
Once you share the vocabulary
When you and an agent share the Enneagram vocabulary, the quality of specification changes. You stop saying "be more direct" and hoping — you say "Type 8 directness, not Type 1." You stop saying "be more creative" and finding out — you say "Type 4 creativity (find what's being missed), not Type 7 creativity (generate more options)." The distinction moves from tone preference down to motivational architecture — and motivation is the layer that drives consistent behaviour.
This is what the Enneagram does in AI agent configuration: not a personality test, not a labelling system, but a fifty-year body of vocabulary purpose-built to distinguish motivational differences at the level of precision that makes them actionable. Bringing it into agent design is not borrowing a fashionable framework — it is bringing in the right tool for the first time.
These ideas — and 33 more like them — form the backbone of *The Complete Enneagram: From Human Personality to Agentic Soul*, just published on Amazon UK. **[Find it here →](https://amzn.eu/d/0fjWGvqR)**