datdang-dev/lewd-writer

Lewd Writer Engine

Overview

The Lewd Writer is the primary prose generation engine of LND Studio. Operated by the Suki agent, it takes a completed forensic-state.json (from Panel Forensic) and an optional knowledge_payload.md (from Transformation Engine) and produces a draft-prose.json containing fully realized R18 Vietnamese Light Novel prose.

The engine enforces a 10-step sequential pipeline with strict escalation loops, sensory density requirements, and format compliance checks. Output is validated against draft-prose.schema.json before handoff to the Quality Audit engine.

INPUTS & OUTPUTS CONTRACT

Inputs:

• forensic-state.json (required, object/file) — The output from the Panel Forensic stage containing visual details, layout, and character dialogue.
• knowledge_payload.md (optional, string/file) — Additional narrative transformation, beat sheets, and character trope cues.
• story_bible (optional, object) — Continuity book and character bios.
• user_fetish_profile (optional, object) — Kinks and core fetish metrics to prioritize in prose details.

Outputs:

• draft-prose.json (conforming to draft-prose.schema.json) — Contains:
  • metadata: manga_name, page_number, POV
  • prose: string containing fully realized Vietnamese Light Novel prose.
  • sensory_affect: audit score breakdown including olfactory, tactile, visual, dialogue_escalation, tsundere_breakdown, and stroking_metric.

On Activation

1. Load config from {{project_root}}/studio/config/config.yaml (resolve output_folder, communication_language)
2. Load forensic state from the provided forensic-state.json path
3. Load knowledge payload from knowledge_payload.md (if available)
4. Load story bible and continuity ledger (if available)
5. Verify output schema exists at {{project_root}}/studio/schemas/draft-prose.schema.json
6. Begin at steps/step-01-context-loading.md

Steps

| Step | File | Purpose | | :--- | :--- | :--- | | 1 | steps/step-01-context-loading.md | Load forensic report, knowledge payload, and story bible context | | 2 | steps/step-02-scene-planning.md | Plan scene structure, pacing, and escalation loops | | 3 | steps/step-03-environment-prose.md | Write environmental atmosphere and spatial setup | | 4 | steps/step-04-dialogue-driven-action.md | Generate dialogue-driven action sequences | | 5b | steps/step-05b-format-ensure.md | HARD GATE — enforce light-novel-prose template format compliance | | 5c | steps/step-05c-sensory-injection.md | Inject mandatory sensory markers (smell, touch, fluid, etc.) | | 6 | steps/step-06-aftermath-polish.md | Write post-climax aftermath and lingering effects | | 7 | steps/step-07-quality-check.md | Self-audit against quality gates before handoff | | 8 | steps/step-08-wiki-update.md | Update character wiki/bible with new state | | 9 | steps/step-09-final-report.md | Generate final prose report and draft-prose.json — WORKFLOW COMPLETE on output |

Note on 5b/5c naming: These are intentionally sub-steps of a single quality enforcement phase. Step 5b is a hard gate — prose CANNOT proceed to 5c if format violations remain.

Data Resources

| File | Purpose | | :--- | :--- | | data/gooner-manifesto.md | Core writing philosophy and style guide | | data/onomatopoeia-library.md | Vietnamese SFX/onomatopoeia reference |

Dependencies

• Agent: Suki (MOD — lewd-writer.agent.yaml)
• Input Schema: forensic-state.schema.json
• Output Schema: draft-prose.schema.json
• Modules: sfx-lookup, fetish-guidance, style-enforcer
• Upstream: Panel Forensic → Transformation Engine → Lewd Writer
• Downstream: Lewd Writer → Quality Audit

Quick Reference

| Intent | Trigger | Route | | :--- | :--- | :--- | | Full prose generation | /prose-adapter | Load steps/step-01-context-loading.md | | Resume from step | /prose-adapter --resume | Read frontmatter stepsCompleted and resume | | Modification pass | Agent receives rewrite instructions from audit | Re-enter at step 4 with fix notes |

---

🛑 HALT CONDITIONS (FAIL-EARLY GATES)

If any of the following conditions are met during execution, you MUST immediately halt processing and return a JSON error response instead of writing any output file:

1. MISSING_FORENSIC_DATA: If the input forensic-state.json is missing or cannot be resolved, halt with: {"status": "HALT", "error": "MISSING_FORENSIC_DATA", "reason": "Required forensic analysis file is missing or inaccessible."}
2. INVALID_FORENSIC_STRUCTURE: If the forensic-state.json is invalid, empty, or does not match its schema, halt with: {"status": "HALT", "error": "INVALID_FORENSIC_STRUCTURE", "reason": "Forensic data is malformed or invalid."}
3. MISSING_REQUIRED_SENSORY_TAGS: If the forensic data does not contain the mandatory sensory grounding elements (at least 2 sensory_tags per active panel), halt with: {"status": "HALT", "error": "MISSING_REQUIRED_SENSORY_TAGS", "reason": "Forensic panels lack the mandatory sensory grounding markers."}
4. UNSUPPORTED_TROPE_GENRE: If the dynamic context specifies a trope or genre beat sheet that cannot be located or is incompatible with the character profiles, halt with: {"status": "HALT", "error": "UNSUPPORTED_TROPE_GENRE", "reason": "Specified narrative genre trope or beat sheet is unsupported or missing."}