Column: Translation Loss and the YAML-First Principle
Why cleaner language reduces effectiveness in generative AI workflows
The prompt is not the character. The prompt is a projection of the character.
Observation
A recurring pattern in production generative AI workflows:
When a prompt is “cleaned up” — made more natural, more concise, or better English — the generation result often becomes less consistent, not more.
This is not a coincidence.
It is a structural property of how generative systems interpret input.
The Translation Loss Problem
When a structured character definition is converted into natural English prose, several things tend to happen:
Useful redundancy gets removed Repetition that appears unnecessary to a human editor is often functionally load-bearing in generation. The model uses repeated signals to weight reconstruction priorities. Removing them reduces constraint strength.
Prohibitions get weakened Explicit constraints (“never show X”, “viewer’s left arm must not cross viewer’s right”) tend to get softened or dropped during natural language editing. A prohibition that survives in YAML often disappears in English paraphrase.
Priority relationships become blurred YAML structure preserves ordering and hierarchy. Flat English prose collapses these relationships. The model cannot reconstruct which attributes are core and which are variable from prose alone.
Structural separation gets merged Ideas that are functionally separate — identity constraints vs. variable elements vs. prohibitions — tend to merge in natural language. Once merged, they are difficult to recover.
Cleaner wording is not always better wording. In generative AI workflows, naturalness and effectiveness are not the same thing.
The YAML-First Principle
The YAML-First Principle addresses this by treating the structured definition as the canonical form, and natural language as a downstream projection.
YAML (SSOT)
└── Structured definition
└── Preserved hierarchy
└── Explicit prohibitions
└── Useful redundancy intact
↓
English prompt (projection)
└── Delivery format only
└── Not the source of truth
Key rule: When generation quality degrades, return to the YAML source — not to the English prompt.
The English prompt is a lossy representation of the YAML. The YAML is the character.
Useful Redundancy
Redundancy in a character definition is often functional, not wasteful.
Examples of useful redundancy:
- Stating a constraint in both positive and negative form
- Repeating a priority attribute in multiple sections
- Explicitly listing what must NOT change alongside what must be preserved
The test for useful redundancy is not “does this say the same thing twice?”
The test is: “if I remove this, does generation drift?”
If drift increases after removal, the redundancy was load-bearing.
Do not remove redundancy because it looks clean. Remove redundancy only after testing that removal does not increase drift.
Human Idea Preservation
The YAML-First Principle is also a human idea preservation mechanism.
When a human author creates a character definition, the original structure encodes:
- priority decisions
- constraint logic
- variation boundaries
- prohibitions and their reasoning
Natural language editing — even well-intentioned editing — tends to erode this encoded structure over time.
YAML as SSOT preserves the author’s original intent against the drift introduced by repeated translation and editing cycles.
The goal is not to produce the best English prompt. The goal is to preserve the character definition long enough for the character to survive changing conditions.
Relation to CIP
This principle operates at the input layer of the CIP governance loop.
| CIP Layer | YAML-First Role |
|---|---|
| Minimal Prompt Principle | YAML → English projection preserves minimalism |
| Anchor Mechanism | YAML SSOT functions as a textual anchor parallel to the image anchor |
| Identity Gates | Structured definition makes gate criteria explicit and auditable |
| Hard Abort | Return to YAML source on quality degradation |
The YAML definition and the anchor image together constitute the full identity specification:
Identity Specification
└── Anchor Image (visual SSOT)
└── YAML Definition (structural SSOT)
Neither alone is sufficient. Together they define the character completely.
Practical Rules
Rule 1: YAML is the source of truth English prompts are output format. Never edit the YAML to match the English.
Rule 2: Preserve prohibitions explicitly Prohibitions that are not explicitly stated tend to disappear during translation. State them in YAML. State them again in English. Do not assume the model will infer them.
Rule 3: Do not remove useful redundancy Before removing any repeated element, test whether removal increases drift. If it does, the redundancy is load-bearing.
Rule 4: When quality degrades, return to YAML Do not attempt to fix a degraded English prompt by editing the English. Return to the YAML source and regenerate the projection.
Rule 5: Cleaner is not always better Natural language optimization is not generation optimization. These are different objectives.
Status
This column documents an operational pattern observed across production generative AI workflows.
Systematic validation has not been conducted. These observations are recorded for further investigation.
First documented: March 2026
Related: Minimal Prompt Principle — Glossary — White Paper