LUCID
“The dominant constraint on real-world AI usefulness is not model capability. It is the loss of fidelity in the channel between the human and the model. Lucid closes that channel.”
Same prompt. Same model. With Lucid and without.
Real prompts from the eval set. The raw column shows what Sonnet 4.6 produces from the prompt as-typed. The Lucid column shows what the same model produces when Lucid shapes the input. Pick a domain.
Measured: 79.4% decisive win rate · 27 W · 7 L · 17 T · n=51 · position-debiased Opus 4.6 judge
Outputs above are hand-picked from /evals/results/skill-full.json for clarity · full eval JSON published with the repo
Listen, clarify, translate, validate.
Four steps between your request and the model. The protocol stays the same whether you install the skill or run the MCP server.
Listen
Extract true intent — including unspoken constraints — from the user's request. Audience, purpose, form, hidden requirements.
Clarify
Ask one or two pointed questions when an answer would change the output by 50%+. Never interrogate.
Translate
Convert intent into the prompt shape the model performs best on. Lead with what serves the form.
Validate
Grade output against intent and re-run on miss. Self-validation in the skill, separate validator pass in the MCP server.