Hawken Reads Your Prose. You Write the Book.
Hawken is the agent that reads what you wrote and tells you what's actually on the page. Scene scans, chapter scans, manuscript-wide reads. Style metrics, sentiment curves, voice drift, structural pacing. It does not generate prose. It does not draft scenes. It does not hand you something to keep. The book on the page stays yours.
A Reading Agent, Not a Writing One
Most "writing assistants" want to put words on the page for you. Hawken is the opposite. Its only job is to read what you've already written and report back. Style metrics, register classification, pacing curves, character screen time, voice drift, structural shape — measured against your prose, your Legendry, and your declared rules. You take the report and decide what to do with it.
- Reads Your Lore First
- Before Hawken weighs in on a scene, it pulls the relevant Legendry entries — characters, locations, factions, active rules. The findings are anchored to the world you actually built, not to whatever pattern a foundation model had on hand.
- Honors ProseGuard Rules
- Findings respect the style rules you've defined. A villain who speaks in sentence fragments stays speaking in fragments. An unreliable narrator stays unreliable. Hawken doesn't second-guess your declared voice.
- Never Touches Your Prose
- Hawken never edits the document. No background rewrites, no autocomplete, no sentence-completion fairy dust. You ask for a scan, you read the scan, and the writing on the page stays unchanged unless you change it yourself.
- Same Input, Same Result
- Every scan is reproducible. Run a scene scan today and again tomorrow on the same prose with the same rules, and the findings list is identical. There's no model temperature changing the answer between sessions.
14 Registers, Calibrated to Your Audience
Pick a register and an audience for the scene before you scan. Hawken calibrates the read against them — a Mythic Narrative scene gets evaluated as mythic register; a Conversational scene gets evaluated as conversational. Findings don't flatten what you wrote into one polished corporate voice, because that's useless feedback.
Six Sub-Reports Per Scene, In Parallel
A scene scan composes six sub-handlers in parallel and assembles a single report. ProseGuard rule violations, Worldspell voice and grammar checks, Insights structural reading, Craft Plot beat presence, scene-metadata contradictions, and the deterministic style metrics — all dispatched at once, all surfaced as one stream of findings.
ProseGuard Findings
Every active rule on the scene runs against the prose and surfaces violations with rule id, severity, matched span, and the project scope the rule was inherited from. The findings carry source-handler badges so you know which engine produced them.
Sentiment Curve
Sentence-level sentiment scoring across the scene with a visualized curve. Useful when the scene is supposed to escalate and the curve says it's actually flat.
Register Classification
Hawken names the dominant register from the 14 categories with a confidence percentage. When the passage you wrote doesn't match the register you intended, the gap shows up here.
Scene Metrics
Word count, sentence count, average length, Flesch-Kincaid grade, adverb density, filter-word density. Real numbers. Useful at first as a baseline; useful later when you want to see whether a revision actually tightened the scene or just moved the words around.
Chapter Scans and Whole-Manuscript Scans
Hawken's scene scan is the unit. Chapter scans fan that unit out across every scene in a chapter and compose chapter-level aggregates on top — pacing variance, tension beats, character screen-time imbalance, plot-point progression. Manuscript scans take it one level further: arc progression per character, beat coverage per plotline, voice drift across the whole book, and a six-section editorial report rendered to PDF.
- Pacing Profile
- Scene-length variance, sentence-length variance, the chapter's tension curve, and flatline-segment detection keyed to your strictness setting. The places where the rhythm goes flat surface as findings, not as a vibes-based "it felt slow" note.
- Character Screen Time
- Per-character mention counts, scene counts, and mention share. Flags when the protagonist gets crowded out by a side character, or when an antagonist disappears for stretches the plot needed them in.
- Theme & Motif Detection
- The manuscript scan ships a topic-modeling pipeline (UMAP + HDBSCAN + c-TF-IDF on bundled MPNet embeddings) that surfaces themes by chapter and recurring motifs with gap-run analysis. When a motif gets established in chapters 1-5 and then disappears for the middle third, the editorial report flags the drop.
- Editorial Report
- The manuscript scan composes a six-section report — Overall Assessment, Structure, Character, Pace and Tension, Theme and Voice, Top Priorities. Rendered to Markdown for in-app reading or compiled to a print-ready PDF for the editor you're sharing it with.
Strictness, Source Badges, and Authored Voice
Every scan accepts a strictness setting on the same four-step scale every Divinity Engine handler uses. Off, low, medium, high. The setting controls which severities surface, which thresholds the heuristics fire at, and how aggressively the report flags borderline cases. Per-panel overrides ride on top of a global default, so you can run Hawken at high while leaving Lorekeeper at medium.
- Source-Handler Badges
- Every finding carries a badge identifying which sub-handler produced it — ProseGuard, Worldspell, scene contradictions, Insights, Craft Plot. You always know which deterministic query is talking to you.
- Authored Library Voice
- Findings render through the bundled Authored Library — pre-written prose variants selected deterministically per finding. The tone is consistent across runs and editable through the Agent Personality tab if you want to write your own variants.
- Strictness Slider, Everywhere
- The same control appears on every analyze panel. Pick once and Hawken's scene/chapter/manuscript scans all run at that strictness. The localStorage override is per-panel, so each surface remembers your preference independently.
- Run It On Your Own Drafts
- Every scan is local. No prose is uploaded. No analysis fee. No quota. Run a manuscript scan ten times in one afternoon if you're chasing a specific finding down. The cost is the price you paid for Ishvana, full stop.