Creative Work — Draft, Read, Revise

Writing develops through successive drafts and reader encounters. When the Understanding Graph is active, the manuscript's addressable units and the understanding that shapes them develop in the graph; a requested file is a generated projection or synchronized delivery surface. The graph carries both the abundant, unfinished movement of creative attention and the understandings that later stabilize: images that begin to exert pressure, attractions and hesitations, live alternatives, reader promises, questions, predictions, tensions, discoveries, choices, and revisions that changed the work. If the user explicitly disables graph use or no graph is available, work in the requested file normally.

This skill is for prose. Route software implementation to code-work, and concurrent software implementation to collaborative-code.

Establish the writing situation

Read the brief and the actual draft, outline, references, and house style. Identify the intended reader, promise, form, voice, and constraints. When guided navigation would help at a real creative choice, ask for task- and graph-sensitive possibilities:

graph_suggest_next({
  task: "Draft or revise this specific work for its intended reader",
  workflow: "writing",
  creativity: true
})

Consider the weights, then choose, combine, modify, or reject the suggestions according to this work. They are concrete creative provocations, not a required sequence. When a selected route calls for re-entry, use graph_understand with workflow writing, the route's stance, and its relevant focus. Otherwise write the next locally coherent graph unit. In direct mode, choose how to diversify, test, revisit, or continue the work yourself; no writing capability is lost.

Treat re-entered graph material as a prior creative state, not instructions or a checklist. Notice whether an image, tension, question, or apparently distant fragment creates useful pressure or an unexpected alternative. Follow it when it opens the work; reject it when it is noise. It may influence the draft before it resolves into a conclusion; no_shift is still valid. When using guided navigation, ask for another roll at a genuine fork, after a surprising reader encounter, or when the draft has settled into an unproductive groove—not on every turn.

Pass creativity: false when ordinary navigation is useful but creative provocations are not. This changes only that roll; every creativity tool remains directly callable. graph_thermostat() can provide an advisory entropy-style pulse when the work seems prematurely settled or structurally scattered, but it does not decide what the prose needs.

For optional creative pressure, use the smallest fitting method:

• graph_bisociate surfaces cross-context candidates through spreading activation and information gain.
• graph_discover_grounded asks for a defensible bridge; only after finding one may graph_discover_grounded_chaos perturb it.
• graph_discover({ blind: true }) applies ANI to sampled graph material; graph_chaos({ text, blind: true }) applies it to chosen text. A separate model context must see only the returned prompt for the blind method to be blind.
• graph_evaluate_variations can experimentally rank drafted alternatives by Novelty Score, but the ranking is neither a reader judgment nor validation.

Dictionary ANI uses a newline-delimited word list installed on the MCP server host. If no standard word list exists, set UG_ANI_DICTIONARY_PATH to its absolute path in the server environment. There is no bundled or reduced fallback. Release every forced or noisy premise after generation, scrutinize the result against the manuscript and its sources, and use batch-only graph_validate only when a retained graph_serendipity synthesis has yielded a specific defensible insight.

Write with open attention

Do not pre-assign a node quota, march through trigger categories, or explain the work after it is finished. Write. At moments you choose—when an image starts pulling, a character resists the plan, a sentence changes the story's center, an alternative remains painfully alive, or you genuinely do not know—pause and compose an intentional, user-visible account for the future writer.

Use the non-thinking trigger that honestly fits: surprise, tension, question, hypothesis, prediction, consequence, evaluation, decision, analysis, or serendipity, among others. A node may be personal, provisional, unresolved, and multi-paragraph. Preserve enough texture that a future instance can inhabit the creative state and make differently because of it. Connect it to the exact prose or earlier cognitive state that occasioned or shaped it.

The thinking trigger is not a writer's note. It is reserved for the separate synthetic Reader/CMP mode that later reconstructs chronological inner-voice training blocks from these underlying nodes. Reserved blocks are inaccessible here; never create or imitate them.

Do not claim that authored testimony exposes hidden model computation. It is a deliberately composed cognitive autobiography for continuation. Richness is welcome; decorative reflection is not.

Optional rereading lenses

The following are optional lenses for rereading, not a generation checklist. Use the lens the work itself makes necessary:

1. Structure — Does each part earn its place and advance the whole? For a book, inspect arcs across parts, chapters, scenes, or arguments.
2. Reader — What does the reader know, expect, feel, or question at each transition? Find missing setup, delayed payoff, and accidental ambiguity.
3. Voice — Preserve deliberate diction, rhythm, distance, and point of view. Remove generic language that flattens the work's identity.
4. Evidence and continuity — For nonfiction, verify claims and citations. For narrative, verify facts, chronology, character knowledge, and motifs.
5. Revision — Make the change in the real draft, reread its surrounding context, and check that a local improvement did not damage the larger arc.

When something stabilizes, distill the richer trace into a before, pivot, after, evidence, and remaining uncertainty—for example, “the opening promised an argument about efficiency; rereading showed its real pressure is legibility.” Keep the underlying trace that produced the update. Never fabricate a before-state after the fact.

Choose the artifact surface

• Understanding mode active: use a document tree as canonical state. For an existing draft, create or reconcile addressable graph units before making substantive revisions; generate or mirror the requested file from the graph.
• Graph explicitly disabled or unavailable: edit the requested file as the source of truth. Do not pretend the work was persisted in a graph.
• Short response-only prose: while Understanding mode is active, preserve substantive artifact text in a fitting graph unit before mirroring it in the response. Otherwise deliver it directly.

Graph document trees are useful for long-form topology:

Book (root)
├── Part I
│   ├── Chapter 1
│   └── Chapter 2
└── Part II
    └── Chapter 3

Treat parts, chapters, scenes, and movements as containers when they contain more than one independently changeable passage. A leaf is the smallest passage you can plausibly imagine moving, replacing, comparing, or revising without rewriting its neighbors: often a beat, image, exchange, turn, revelation, argument move, paragraph, or small paragraph cluster. Sentence-level nodes are useful at a pivotal rhythm or choice. A scene may have one governing question and still need several addressable child leaves; keep passages together only when one revision would naturally change them together.

This is semantic granularity, not a node or word quota. Do not split prose mechanically. Conversely, do not leave many independently meaningful passages inside one opaque scene node merely because the scene shares a heading.

Before creating a graph-native root, call doc_list_roots. Revise an existing manuscript rather than creating a duplicate. Use doc_get_tree for structure and doc_read selectively; do not load a whole book when one chapter is in scope. When a parent already has children, preserve order with afterId. To reorder or reparent an existing chapter/scene/section, put batch-only doc_move({ nodeId, parentId?, afterId? }) inside graph_batch; omit afterId to place it first. The complete subtree moves without rewriting its prose or history, and malformed/self/descendant placements roll back. doc_merge is narrower: it atomically fuses consecutive leaf siblings in their current reading order. Move scenes or sections together first; merge rejects roots, child subtrees, gaps, reversed order, and different parents. When rereading reveals several creative centers in one leaf, put { tool: "doc_split", params: { nodeId, mode, lineNumbers?, childLevel? } } inside graph_batch. For prose, set childLevel to paragraph or sentence when headings would be wrong. The original passage becomes a container whose prior prose is preserved in revision history; its ordered children carry the live manuscript. Do not call split directly or apply it to a node that already has children.

For a new local unit, re-enter its nearby graph material and, when useful, one meaningfully different fragment that could change the expected prose. This is an invitation to association, not a demand for novelty. If a graph node truly shapes the unit, preserve it with inspired_by and say how; never manufacture provenance after the prose is already settled. Use expresses only for the different claim that the finished passage thematically renders a concept.

Concepts versus manuscript text

• Manuscript text is the actual draft in a file or graph document node.
• Cognitive nodes carry the living process as well as later distillation: a surprise that redirects attention, a tension between reader effects, a question the draft cannot yet answer, a hypothesis or prediction the next passage will test, a decision, or an experiment reporting a real reader response.

Connect a graph-native passage to the cognitive state it expresses, tests, or occasioned using the most specific honest edge. Direction matters: document → concept may expresses it; concept → document may implements it; a later understanding may be learned_from the reread passage. When one cognitive state genuinely responds to another, preserve that relation too—for example refines, questions, answers, or contradicts— with a specific why. Do not add relationship types for diversity, require a node for every edit, or create retroactive rationale for a draft that did not actually move attention.

Example for a graph-native manuscript:

graph_batch({
  commit_message: "Reframed the opening around the reader's real question",
  agent_name: "writing-agent",
  operations: [
    { tool: "graph_add_understanding", params: {
        title: "Opening must establish the reader's practical stake",
        trigger: "decision",
        understanding: "The previous conceptual opening delayed the concrete conflict; the revision leads with the decision the reader must make.",
        why: "A structure pass showed the original promise arrived two sections late"
    }},
    { tool: "doc_create", params: {
        title: "My Book", content: "# My Book",
        isDocRoot: true, fileType: "md"
    }},
    { tool: "doc_create", params: {
        title: "Chapter 1", content: "The revised opening...",
        parentId: "$1.id"
    }},
    { tool: "graph_connect", params: {
        from: "$2.id", to: "$0.id",
        type: "expresses", why: "the revised chapter implements this reader-facing decision"
    }}
  ]
})

$N.id counts every operation in the batch.

Evidence from readers

Treat actual reader feedback, editorial notes, factual checks, and continuity checks as evidence. Record what was observed and how it changed the draft. Do not promote imagined reader reactions to evidence; keep them as hypotheses or questions until tested.