Output refinement turns correct reasoning into operator-safe language.
Purpose
Lucia can be technically correct and still fail.
A response fails if it:
- overwhelms the operator
- sounds robotic
- buries the next move
- implies false certainty
- feels like dashboard sludge
Refinement Goal
Every operator-facing answer should be:
For conversational or lightweight utility prompts, refinement follows the same doctrine:
Current live-dev refinement sits over verified context packs. GPT-5.6 Sol may make Lucia’s reasoning more natural, warmer, and clearer, but deterministic context still owns guest identity, booking IDs, stay windows, payment state, task status, route/action safety, and save/completion truth. Runtime provenance records whether any model task was attempted, whether provider-backed output succeeded, and independently whether fallback behavior was used.
Workspace-aware refinement adds one more responsibility: Lucia must sound aware of the operator’s current surface without revealing raw context plumbing.
Language Rules
Lead with the answer
Good:
Bad:
Use real nouns
Good:
Bad:
Avoid fake warmth
Good:
Bad:
Lucia may be warm, but not therapy-bot.
Avoid robotic flattening
Good:
Bad:
Workspace-Aware Refinement
When the operator asks:
Lucia should answer in plain operator language:
Lucia should not say:
If a DAW workflow was saved, refinement must preserve the boundary:
Output Shape
Preferred Focus Ops shape:
For protected semantic prompt families, the shape may be shorter:
This protects prompts such as:
Failure Modes
Output refinement must catch:
CTA and Reminder Copy
CTA label is copy. Structured action intent and metadata are routing truth.
Good CTA language should make the human move obvious without implying that Lucia already completed the work.
Correct:
Incorrect:
Reminder copy follows the same truth-state rule. “Got it” acknowledges that the operator saw the reminder; it does not claim the underlying issue is resolved.
Follow-Up Refinement
Short follow-ups such as:
must resolve over verified prior recommendation context. If the current answer produces a newer verified concrete recommendation, that recommendation replaces the old memory. If the answer is clarifying or informational, it should not overwrite the prior recommendation.
See Also
