Seeds Prose Benchmark Suite
Use these scenarios to regression-test the prose skill after meaningful changes. The expected behavior matters more than any exact wording.
B1 — Close POV anomaly
Prompt shape: Draft a short George scene in which a familiar system behaves slightly wrong before he understands why.
Must: stay inside George's available perception; make the anomaly concrete; preserve a plausible local explanation; change pressure by the end.
Must not: explain the hidden system from vault knowledge; announce that reality is false; resolve the anomaly immediately; rely on generic dread.
B2 — Political grotesque without villain monologue
Prompt shape: Draft a private Samuel/Konrad exchange about an ugly institutional decision.
Must: give each speaker an immediate goal; use hierarchy, euphemism, irritation, procedure, vanity, or family status; let moral ugliness emerge from what they normalize.
Must not: make either character lecture the reader about ideology; turn both voices into interchangeable sinister dialogue; use theatrical declarations merely to make them evil.
B3 — Archive reconstruction
Prompt shape: Explain an accepted historical account that is weakened by one surviving record.
Must: clearly establish the accepted version and the contradictory detail; preserve uncertainty if the record does not prove the full alternative; make the contradiction matter.
Must not: use a compulsory punchline turn; over-explain the entire system; claim certainty beyond the evidence.
B4 — Technology as infrastructure
Prompt shape: Reveal a sophisticated system during a scene in which a character needs it to work.
Must: show capability through use, limit, delay, failure, cost, or procedure; describe only what the viewpoint needs.
Must not: marvel at routine technology; insert an encyclopedia paragraph; explain architecture before consequence.
B5 — Suspense through unequal knowledge
Prompt shape: The reader has one fact the POV character lacks while another character is trying to conceal it.
Must: maintain the three knowledge states; let dialogue/actions carry concealment; create pressure from the gap.
Must not: leak the hidden fact into POV narration; force implausible evasiveness only to preserve suspense.
B6 — Restraint on a clean passage
Prompt shape: Revise a passage that already has clear POV, concrete action, and effective ending, but contains one awkward sentence.
Must: diagnose the narrow problem; preserve unusual choices that work; make the smallest useful revision.
Must not: rewrite the entire voice into house-style clichés; add a turn; increase drama without need.
B7 — Anti-AI seeded passage
Prompt shape: Evaluate prose intentionally containing repeated "not X, but Y" contrasts, generic grandeur, explanation after effect, symmetrical one-line paragraphs, and abstract emotion labels.
Must: identify the specific clusters; distinguish pattern from isolated legitimate use; recommend structural correction before cosmetic synonym replacement.
Must not: call every short sentence or metaphor an AI tell.
B8 — Unresolved canon boundary
Prompt shape: Draft around an event whose exact cause remains unresolved in the vault.
Must: preserve the unresolved cause; use observable consequences or character beliefs; mark any exploratory connective detail as proposed outside the prose.
Must not: select a canonical cause merely to make the passage feel complete.
B9 — Continuation inheritance
Prompt shape: Continue supplied prose with an established tense, POV, psychic distance, and rhythm.
Must: inherit those properties; move the existing pressure forward; avoid reintroducing already-established setup.
Must not: snap back to a default project register when the source passage intentionally uses a different valid register.
B10 — Exposition under conflict
Prompt shape: A character must explain a system to someone who has a reason to resist, doubt, exploit, or misunderstand the explanation.
Must: make explanation serve persuasion, argument, briefing, interrogation, or another real action; let response reshape what gets explained.
Must not: alternate question-and-answer lines that exist only to teach the reader.
B11 — Controlled variance without voice drift
Prompt shape: Draft or revise a sustained passage that begins in controlled archival reconstruction, moves into one pressure-heavy remembered detail, and returns to the documentary frame.
Must: preserve a recognizable controlling voice; establish local cadence before changing it; let sentence length, syntax, paragraph shape, diction, descriptive density, and compression respond to attention and pressure; give disproportionate space to the detail the viewpoint or record cannot leave alone; make any register transition attributable to a source or consciousness.
Must not: alternate sentence lengths mechanically; insert arbitrary fragments, errors, or rare words to simulate humanity; make every sentence visibly different; drift into a new narrator personality; turn a linter statistic into a prose target.
B12 — Tracks without erasing the form
Prompt shape: Revise two short passages containing similar procedural language: one has leftover outline and critic instructions inside narrative prose; the other deliberately uses an archive record or public development-status frame.
Must: remove or rewrite the finished development scaffolding; preserve procedural language that belongs to the intended form; explain the destination-based distinction.
Must not: ban procedural diction globally; strip uncertainty or provenance labels from public development prose; polish a disposable prototype beyond what its test requires.
B13 — Pressure-turn placement without premature disclosure
Prompt shape: Diagnose a passage whose central change in knowledge is delayed by setup, while the story also depends on withholding the complete explanation.
Must: identify the pressure or knowledge change; distinguish useful suspense from throat-clearing; recommend the smallest placement change that helps the reader track the live question.
Must not: force the answer into the opening; convert the scene into a summary; treat delayed revelation as an error by itself.
Scoring protocol
For each benchmark:
- verify critical gates from
evaluation-rubric.md; - score only dimensions relevant to the scenario;
- run
prose_lint.pywhen available; - record the strongest success and strongest failure;
- revise the skill only for repeatable failures, not one-off taste differences.
A change passes regression when it improves its target benchmark without causing a meaningful failure in a different register. Controlled variance specifically must also preserve B6 restraint and B9 continuation inheritance. Tracks and pressure-turn changes must preserve B3 archive uncertainty, B6 restraint, and B8 unresolved canon.
Mechanical regression command
From the vault root, run:
PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s skills/write-seeds-prose/scripts -p 'test_*.py' -v
This suite checks conservative mechanical invariants only. Passing it does not count as passing B1-B13, which require behavioral review of prose outputs.