Skip to content

PromptEmbeddedStyleConstraints

Dennis Lee edited this page Aug 17, 2026 · 1 revision

title: Prompt-Embedded Style Constraints (ASD-STE100) radar_quadrant: Techniques radar_ring: Assess radar_position: outer

Prompt-Embedded Style Constraints (ASD-STE100)

LLM-generated writing tends toward a recognizable house style: hedging modals like "should" or "might," vague filler like "seamlessly synchronize," passive constructions, and long, sprawling sentences. In documentation and technical contexts, that style is a liability, not just an aesthetic complaint, because it obscures the exact condition or action a reader needs to act on.

The SimpleEnglish project demonstrates a technique for eliminating this pattern by embedding a fixed rule set directly into the model's instructions rather than relying on post-generation editing or a linter pass. It packages ASD-STE100 Simplified Technical English, a 1983 aerospace documentation standard, as 53 numbered rules across 9 categories: ban hedging modals, cap sentences at 20 words, require active voice, and put any risk or condition statement before the command it qualifies. The rules ship as a SKILL.md file for agent-skill-compatible harnesses, or as a plain system-prompt block for any chat interface. Compliance depends entirely on the model following the embedded directives; there is no automatic checking or rewriting step.

Benchmarks the project published across six Claude models measured a 72.9% reduction in STE rule violations per 100 words, with output token counts holding steady or declining rather than growing, suggesting the constraint tightens prose without adding verbosity.

Radar Assessment

Placed in Techniques / Assess / outer.

The technique of embedding a fixed style rule set directly in agent or system-prompt instructions, rather than filtering output after the fact, is a distinct and reusable pattern, and this project's own benchmark numbers give it more grounding than an unverified claim. It has not been applied in the user's own workflow, so it stays at Assess. Outer position reflects that this is a single documented instance of the pattern, not one already proven to generalize beyond aerospace-style technical writing.

Trial gate: apply an ASD-STE100 or similar prompt-embedded rule set to the user's own agent or writing workflow and confirm it holds up in practice.

References

Clone this wiki locally