From a Word file to a cited critique, one step at a time
There is no black box here. Three processes run in order — import into the special
format, knowledge-graph construction, then graph parsing and critique. The step names
below are read directly from the pipelines the workers execute, so this page cannot
drift from what actually happens to your manuscript.
.stybld is the native
AIStoryBuilders format. AIStoryBuildersPro adds the graph and report in a
.stybldpro project.
Inspection is free and deterministic. Nothing is charged and no AI runs until you approve the maximum cost.
How a manuscript becomes a structured project. Upload: Word, PDF, text, or Markdown Then Inspect: Deterministic scan. No AI, no charge. Then Approve: Maximum credit cost shown up front Then Convert: Twelve steps create indexed paragraphs Then Package: .stybld story plus .stybldpro project
Process 2
A typed knowledge graph is built from the structured story
The graph is not a word cloud. Its nodes are the concrete things in your story and its
edges are the relationships your prose asserts. Every node and edge carries evidence:
the chapter and paragraph where the claim was made.
Nodes are the concrete things in your story; edges are the relationships your prose
asserts. The amber nodes are the evidence anchors that make every critique citable.
A typed story knowledge graph. A Chapter node has beats through HAS_BEAT. A Beat node is
ordered against other beats through PRECEDES and is mentioned in Paragraph nodes through
MENTIONED_IN. A Character node, whose aliases have been merged into one identity, appears in
the chapter through APPEARS_IN, relates to other characters through RELATES_TO, is placed at a
Location through LOCATED_AT, and carries checkable facts through HAS_ATTRIBUTE. A Location
node, also alias-merged, is mentioned in Paragraph nodes. Every node and edge terminates in a
paragraph, which is why every finding can cite its source.
Alias merging
One character, however many names your prose uses for them.
Evidence anchoring
Every fact points back at a paragraph, so every critique is citable.
Determinism
The same manuscript produces the same graph shape, so re-runs are comparable.
Process 3
The graph is parsed, and the critique is written on top of it
Deterministic queries run first. The model is only asked questions the graph has already framed.
How the graph becomes a cited critique. Query: Deterministic graph questions run first Then Check: Continuity analyzer compares facts Then Read: Model reads batches with graph context Then Score: Versioned rubric with rationales Then Anchor: Margin notes at exact character offsets
How the critique is assembled
Deterministic graph queries run first and produce findings that need no AI at all.
The continuity analyzer compares facts across chapters and reports contradictions with paragraph anchors.
The manuscript is batched and sent to the model together with the graph context, so answers are grounded in structure rather than vibes.
Every finding is scored and normalised against the versioned rubric.
Findings become margin notes with exact character offsets, plus the editorial letter and its Fix First list.
All of it is written to report.json inside your project package.
What you end up reading
Report areas
Report area
What it contains
Editorial letter
A named editor's letter, opening with a Fix First list
Story map
Cast, locations, timeline tracks, node map, and analysis cards
Story markup
Your manuscript with margin notes anchored to exact character spans
Rubric
Versioned scores with a written rationale for every item
Log
Every step, token count, and decision — full transparency
