SemText

Reference · NPL

<npl-ref>

NPL element

PromptLingo reference ↗

You write the assessment, the improvements, the validation, and any eval notes. The element lays them out as a review card after the work.

Looks back at a finished turn: what went well, what to improve, and how it was checked. You write the assessment, the improvements, the validation, and any eval notes. The element lays them out as a review card after the work. one of `full`, `compact`, optional — `full` renders each section as a labelled block; `compact` renders one emoji line per section. Any other value resolves to `full`, and the resolved variant mirrors to `data-variant` (dual-form contract; the mirror is deferred past parse via afterParse, like sem-note, so a sem-source snapshot never sees runtime writes). The variant is read at upgrade; it is display-only and not switched after wiring. per the global attribute catalog (conventions.md §2) per the global attribute catalog (conventions.md §2) per the global attribute catalog (conventions.md §2)

Children

  • <assessment>what held and what did not across the finished turn.
  • <improvements>concrete changes the next run should carry.
  • <validation>how the result was checked.
  • <eval>the measured outcome (scores, counts, diffs).
  • <fine-tune>training or retrieval follow-ups.

Example

npl-ref.html
source
1 <npl-ref variant="full">
2 <assessment>the approach held; one gap in the eval loop</assessment>
3 <improvements>add a regression harness before the next run</improvements>
4 <validation>checked against three traces</validation>
5 <eval>quality 0.82 vs 0.71 baseline</eval>
6 <fine-tune>retrain on the two rejected traces</fine-tune>
7 </npl-ref>
render

Live render boots with JavaScript — the authored source is alongside.

variant: compact

One emoji line per section; the authored sections are hidden from the render (kept in the source). The glanceable form for a finished turn.

variant: compact
npl-ref-compact.html
source
1 <npl-ref variant="compact">
2 <assessment>the approach held; one gap in the eval loop</assessment>
3 <improvements>add a regression harness before the next run</improvements>
4 <validation>checked against three traces</validation>
5 <eval>quality 0.82 vs 0.71 baseline</eval>
6 <fine-tune>retrain on the two rejected traces</fine-tune>
7 </npl-ref>
render

Live render boots with JavaScript — the authored source is alongside.

Source: spec/schema/npl/npl-ref.md