Blog

Project DNA: Agents Can't Write Against Genre

NovelKit Project DNA is the contract that locks genre, tone, and style — agents on Hermes cannot write against the DNA.

NovelKit Project DNA (PROJECT_DNA.md) is the first file in every novel workspace. It's not a marketing form — it's the contract injection for the entire AI novel writing process: the Lãng Khách orchestrator, the genre squad, and the novelkit_* tool registry all read the DNA before plan_next.

PROJECT_DNA structure

The NovelKit template covers: story info, target reader, creative premise, worldview (locations, power system, cultivation clock), main characters, arc map, style taboos, and a worldbuilding guide pointer. Every section has a verification checkbox before the pipeline runs.

The DNA references genre-specific consistency rules: Xianxia, Romance, Meta Genre/System, Urban, Sci-fi... in the skills/novelkit-canon/ skill bundle.

DNA and agents on Hermes

The context-engine plugin injects the DNA + canon files into the retrieval budget. The memory plugin isolates per novel — Title A's DNA never leaks into Title B. Prose Writer resolves Author Style from the DNA; it falls back to Nhĩ Căn if none is specified. Quality Auditor checks output against the DNA and the database.

Creating a Project DNA on Studio

Sign in at studio.novelkit.cc → Quick Setup or Full Setup → fill out the DNA → run bootstrap. Guide: Writing guide. Related: Story bible.

Can the DNA be edited mid-project?
Yes, but it has to go through a controlled sync; doctor checks coherence after any major change.

Read more: The 9-step pipeline · What is NovelKit · The writing system

DNA injection in the Hermes pipeline

Every task runner receives a creative_input_bundle with a pointer to PROJECT_DNA and the genre consistency rules. The context-engine plugin injects a DNA summary into the retrieval budget — agents always see the genre lock before generating. Changing the DNA mid-project means changing the contract; it has to go through an impact review against existing canon.

Author Style / Đại Thần in the DNA

The DNA locks the style_model (GHTK, MHTK, NC...). STYLE_ENGINE uses a terminology database and dialogue generator matched to the profile. Prose can't blend Đại Thần A's and B's voice in the same title unless the DNA allows a rule-based hybrid.

Key DNA fields worth noting

  • Genre + style_model
  • Target reader + tone
  • Creative premise / hook
  • Worldbuilding guide path
  • Cultivation speed lock (xianxia)
  • System type (system genre)

Template: skills/novelkit-canon/templates/PROJECT_DNA_TEMPLATE.md in the technical repo.

NovelKit Project DNA: a creative contract locked before the pipeline runs

NovelKit Project DNA (PROJECT_DNA.md) is the single source creative contract: genre, style_model (GHTK, MHTK, NC...), target reader, premise, worldbuilding guide path, cultivation speed lock, system type. Every task runner receives a creative_input_bundle with a DNA pointer — the context-engine injects a summary into the retrieval budget before the agent generates. Changing the DNA mid-project means changing the contract; it requires an impact review against existing canon.

Lãng Khách reads the DNA at bootstrap — dispatching the right genre squad: xianxia, romance, isekai, system, urban, sci-fi. A vague DNA ("xianxia + romance + system" with no locked ratio) dilutes the voice and triggers a DNA Depth Audit SOFT-FAIL from Quality Auditor.

DNA injection in the Hermes pipeline

Bootstrap: novelkit_dna plus the character/world tools run in parallel. Outline stage: Plot Weaver reads the DNA + master outline constraints. Write: Prose Writer + the STYLE_ENGINE terminology database follow the style_model. Review: Quality Auditor checks 5 DNA Depth fields — MC Archetype, Hook Strategy, style_model, worldbuilding_guide, world frame; 2 weak fields trigger SOFT_FAIL_STYLE.

DNA fields Publishers should review during a pilot

  • Genre + service line lock
  • style_model and the Author Style file pointer
  • Creative premise / hook one-liner
  • Worldbuilding guide path in the database
  • Cultivation speed / system type if applicable
  • Target chapter count and arc structure hint

DNA and canon authority

Authority order when there's a conflict: (1) PROJECT_DNA + database canon, (2) genre consistency rules, (3) retrieved memory/index, (4) model suggestion. The DNA sits in layer A canon — RAG can never override it. Technical template: skills/novelkit-canon/templates/PROJECT_DNA_TEMPLATE.md.

6 service lines and the DNA lock

NovelKit serves 6 service lines: xianxia, romance, isekai, system, urban, sci-fi — each with its own genre squad and consistency rules. The DNA must lock genre + style_model (GHTK, MHTK, NC...) + the blend ratio if mixing genres — a vague DNA ("xianxia + romance + system" with no lock) dilutes the voice and triggers a DNA Depth Audit SOFT-FAIL from Quality Auditor.

DNA injection in the creative_input_bundle

Every task runner receives a creative_input_bundle with a DNA pointer — the context-engine injects a summary into the retrieval budget before generation. Bootstrap: novelkit_dna plus the character/world tools run in parallel. Write: Prose Writer + the STYLE_ENGINE terminology database follow the style_model. Review: Quality Auditor checks 5 DNA Depth fields — MC Archetype, Hook Strategy, style_model, worldbuilding_guide, world frame; 2 weak fields trigger SOFT_FAIL_STYLE.

GenreRequired DNA fieldGenre squad file
Xianxiacultivation speed lockXianxia_consistency_rules
Romancelead pairing's Want/Need/Lieromance depth contract
IsekaiMC knowledge boundaryMeta Genre Pitfall Watchlist
Systemsystem type + UI rulesSYSTEM_CHEAT guard

Changing the DNA mid-project means changing the contract; it requires an impact review against existing canon, a doctor run, and amending the outline if needed. Publishers piloting should review the DNA before bootstrap — fixing the DNA after 20 chapters costs far more rework than fixing it before the pipeline starts.

Cluster: story bible, canon file-first, pipeline, writing guide, Studio.

FAQ: Project DNA in a B2B pilot

Can the DNA change after 20 chapters?
Yes, but it requires a canon impact review, amending the outline, and running doctor — it costs rework.
Can multiple genres be blended?
The ratio must be locked in the DNA — otherwise the Auditor triggers a DNA Depth SOFT-FAIL.
Does the DNA override RAG?
Yes — it's layer A canon; a retrieved index never beats PROJECT_DNA (principle P5).

Studio bootstrap reads the DNA before every task — the context-engine injects a summary into the retrieval budget. Quality Auditor's DNA Depth Audit checks 5 fields; 2 weak fields trigger a SOFT_FAIL_STYLE before scaling to a batch.

enrich_dna and bootstrap planning

The novelkit_dna tool can enrich a DNA from a deconstructed reference — but it still outputs a PROJECT_DNA.md file, never chat-only. Bootstrap planning docs (outline pointers) spawn after the DNA is verified. Every step has an artifact path.

Story bible, Studio setup.

DNA versioning and mid-season changes

Changing the tone or power system mid-arc requires a documented DNA amendment plus a doctor pass. Agents read the new DNA starting from the next task — it doesn't retroactively rewrite old chapters automatically. The Publisher approves the amendment before the next sync batch.

Author Style / Đại Thần

The DNA can point to a style model (a voice reference, for example). Prose Writer resolves the profile; it falls back to Nhĩ Căn if none is specified. A style coherence cron runs every 10 chapters to catch drift from the DNA lock.

Commonly left-blank DNA fields

Target reader, taboo list, cultivation speed, HE/BE contract (romance), system type (system genre). A blank field means the agent defaults to something generic → review SOFT-FAIL. Fully filling out the DNA is the highest-ROI investment you can make.

Story bible, Writing guide.