feat(linkedin): add persona-reviewer agent (2 modes) + fixture (S5)
This commit is contained in:
parent
be03d44c52
commit
1faffac303
3 changed files with 392 additions and 0 deletions
|
|
@ -0,0 +1,106 @@
|
|||
# Persona-Reviewer Fasit Fixture
|
||||
|
||||
One reader persona, one sample draft, the six judging axes, and the two review
|
||||
modes — used to sanity-check the `persona-reviewer` agent. This is a *fasit*,
|
||||
not a test harness. The structural lint lives in
|
||||
`agents/__tests__/persona-reviewer-fixture.test.mjs`. Whether the agent's live
|
||||
flags actually match the directions below is `[GATE]`/`[OPERATØR]` — it is not
|
||||
self-certified here.
|
||||
|
||||
**The jury judges; the editor writes.** Every expected output in this fixture is
|
||||
**direction, not rewritten copy**. A correct agent run hands back flags and a
|
||||
verdict — never edited text.
|
||||
|
||||
---
|
||||
|
||||
## Persona under review (primær)
|
||||
|
||||
Drawn from `config/personas.template.md` — the primær Linjeleder. Field keys are
|
||||
lowercase to match the library contract.
|
||||
|
||||
- **rolle** — Mellomleder med fag- og personalansvar i offentlig virksomhet;
|
||||
skal beslutte om og hvordan AI tas i bruk i egen enhet, uten dyp teknisk
|
||||
bakgrunn.
|
||||
- **avkobler** — Teknisk dypdykk uten «hva betyr dette for meg og mine»;
|
||||
frykt-retorikk; abstrakt policy; språk som forutsetter at hen kan koden.
|
||||
- **overbeviser** — Konkrete eksempler fra arbeidshverdagen, et klart ansvars- og
|
||||
dømmekraftsbilde, og en leder-takeaway hen kan handle på allerede i morgen.
|
||||
- **ekspertise** — Lav-til-middels teknisk; høy på ledelse og forvaltning.
|
||||
Trenger oversettelse, ikke nedlatenhet.
|
||||
- **sjargong** — Lav toleranse for teknisk sjargong; setter pris på presise,
|
||||
hverdagsnære formuleringer.
|
||||
|
||||
> Primær trumfer: a primær NO is not accepted — the text is revised until this
|
||||
> reader reaches a clean JA. A sekundær NO from a role or expertise ceiling is a
|
||||
> SIGNAL the gate works, not a defect.
|
||||
|
||||
---
|
||||
|
||||
## Sample-tekst
|
||||
|
||||
The draft excerpt this persona reads, read-only. Deliberately mixed: a workable
|
||||
human angle ("you don't outsource judgment") buried under one wall of jargon —
|
||||
exactly the case where the jury should flag direction without touching the copy.
|
||||
|
||||
> Transformatorarkitekturen vår utnytter selvoppmerksomhets-mekanismer over en
|
||||
> 175-milliarders parametermodell for å maksimere inferens-gjennomstrømning på
|
||||
> tvers av distribuerte GPU-clustere. Men det egentlige poenget er enklere: en
|
||||
> språkmodell tar ikke beslutningen for deg. Den foreslår — du svarer for
|
||||
> resultatet. Dømmekraften kan ikke settes ut. Spørsmålet er ikke om verktøyet
|
||||
> er smart nok, men om du fortsatt eier valget når det teller.
|
||||
|
||||
---
|
||||
|
||||
## The six axes
|
||||
|
||||
The persona-reviewer judges the sample on exactly six axes and returns **at most
|
||||
five flags** as direction (the sixth that passes cleanest is simply not flagged):
|
||||
|
||||
1. **Krok** — does the hook hold for THIS reader in the first two lines? Here:
|
||||
IKKE — the opening jargon wall ("transformatorarkitektur… inferens-
|
||||
gjennomstrømning") hits `avkobler` head-on; the linjeleder stops reading
|
||||
before the real point.
|
||||
2. **Resonans** — does the central point land for this reader? DELVIS — the
|
||||
"judgment can't be outsourced" core is exactly their concern, but it arrives
|
||||
too late to land.
|
||||
3. **Tone** — right for this reader (no condescension, no fear-rhetoric)? LØST —
|
||||
tone is respectful and non-alarmist once past the opening.
|
||||
4. **Troverdighet** — does the reader believe it? DELVIS — the closing claim is
|
||||
credible but abstract; no lived example from a leader's workday to anchor it.
|
||||
5. **Leder-takeaway** — one concrete thing this reader can act on tomorrow?
|
||||
IKKE — there is an insight but no action the linjeleder can take in their
|
||||
own unit.
|
||||
6. **Lengde/driv** — does it keep moving or sag? DELVIS — the front half drags
|
||||
under the jargon; the back half drives well.
|
||||
|
||||
Each flag is **direction**, not a rewrite: "the hook hits `avkobler` — open on
|
||||
the decision the leader owns" is correct; supplying the new opening line is not.
|
||||
|
||||
---
|
||||
|
||||
## The two modes
|
||||
|
||||
Both modes run the same persona but differ in scope and output.
|
||||
|
||||
### Resonans-modus (before lock)
|
||||
|
||||
Runs at the newsletter pipeline's pre-lock resonance sweep. Judges all six axes
|
||||
and returns ≤5 flags as direction, each tracked **LØST / DELVIS / IKKE**. The
|
||||
primær must reach a clean JA before the draft is locked. For this sample, the
|
||||
primær verdict is **NEI** (Krok + Leder-takeaway both IKKE) → REWORK, with the
|
||||
two IKKE flags as the priority directions.
|
||||
|
||||
### Konverter-modus (after lock)
|
||||
|
||||
Runs at the post-lock conversion sweep. Judges the **hook only**, binary:
|
||||
«would YOU click?» **JA / NEI**. No axis scoring, no copy — just the click
|
||||
verdict and a single reason. For this sample's current hook the verdict is
|
||||
**NEI** — "I'd scroll past; the first line is machinery, not me."
|
||||
|
||||
---
|
||||
|
||||
## Convergence loop
|
||||
|
||||
Re-run per persona until the primær returns a clean JA. Each flag is re-judged
|
||||
LØST / DELVIS / IKKE against the editor's revision. The jury never writes the
|
||||
fix — it only re-judges whether the revision now lands.
|
||||
Loading…
Add table
Add a link
Reference in a new issue