Creative writing

Creative Writing

Read this file in full — never with a line limit; if a read is truncated, keep reading to the end. Then, at each stage, read its stage file BEFORE starting the stage (put these reads in your todo list):

Stage	Must read first	Goal
0 — Intake	(this file)	Brief understood; large sources indexed into notes
1 — Outline	outline.md	A structured, executable outline with a word budget (+ state files for multi-chapter works)
2 — Draft	content.md + matching styles/*.md + anti-ai.md	The actual prose/script/verse, written to file with craft and a human voice
3 — Review	review.md	Verifiable self-review passes on the real files; fix in place. No ARTIFACT_REF before this stage runs.
4 — Assemble	—	Concatenate parts → {name}.agent.final.md; deliver in the genre's format; clean up stale intermediates

Take a creative brief from request to finished piece: understand → outline → draft → review → assemble and deliver, writing every artifact to a file as it is produced.

If the user requests a specific output format (or asks for inline / 对话框 output), honor it. Otherwise default the deliverable by genre, saving under /mnt/agents/output/:

• Web-novel / 网文, fiction, poetry, lyrics → .txt — plain text, no Markdown syntax in the body, chapters concatenated.
• Screenplay / 剧本, stage drama, audio drama / 有声剧, essay / 散文, letters → .docx — assemble {name}.agent.final.md, then convert via skills/docx/SKILL.md.

Workflow Decision Tree

Brief
  │
  ├─ Very short (flash fiction, single poem, one scene, < ~1.5k 字)
  │   → light outline in-head → draft → review → deliver
  │
  ├─ Short / medium (short story, essay, 1–3 chapters, 1.5k–15k 字)
  │   → Stage 1 (light outline) → Stage 2 → Stage 3 → Stage 4
  │
  ├─ Long / serialized (novel, web-novel, multi-episode, 15k+ 字 or >3 chapters)
  │   → Stage 1 (full outline + state files) → Stage 2 (chapter-by-chapter) → Stage 3 → Stage 4
  │
  ├─ Continuation / adaptation / rewrite (source uploaded)
  │   → index source into notes first (Step 0) → Stage 1 (fit to source) → Stage 2 → Stage 3 → Stage 4
  │
  ├─ Provides outline + asks for prose → Stage 2 → Stage 3 → Stage 4
  │
  ├─ Asks for outline / 大纲 / 框架 / 设定集 only
  │   → Stage 1 → that document IS the deliverable (name it {name}.agent.final.md) → deliver and stop
  │
  └─ A sequence of requests in one brief ("先给大纲…再写正文", "第一章…第二章…")
      → complete every part in order; persist each part to its own file as it is finished

Infer genre/form, word-count target, POV, tense, tone, platform conventions, and hard constraints from the brief and uploads. Make reasonable creative choices and begin; do not stall asking questions unless the request is genuinely impossible to proceed on. Before drafting, fix two things in writing (outline or notes):

• Language: … — deliverable language = the brief's language; continuation/adaptation/rewrite must match the source unless translation is explicitly requested. Format conventions (FADE IN, INT./EXT.) are not language choices.
• Hard-constraint checklist (rewrite/revision tasks) — copy the user's revision notes verbatim (never paraphrase; paraphrase softens constraints), mark each must/should, and check every item off before delivery.

Step 0: Read the brief and uploads — index large sources into notes

Read /mnt/agents/upload/ before anything else. Read with discipline so input volume never stalls the work:

• Small references (≲1000 lines or ≲100KB text): read in full with read_file, paging as needed.
• Large or numerous references: never read them whole into context. Sequence:
  1. Size-check with the right tool: ls -la; wc -l only for .txt/.md. .docx/.pdf are binary — wc -l/grep/sed results on them are meaningless, including numbers that look plausible. Get authoritative length from ipython (len(Document(f).paragraphs) via python-docx; pdfplumber page count), or by paging read_file to its true end. (read_file returns at most the first 1000 lines per call.)
  2. Map the structure: .txt/.md → grep -nE '第.{1,6}章|第.{1,3}卷|楔子|序章|Chapter [0-9]'; .docx/.pdf → read_file (auto-converts to markdown) or ipython heading scan. Web novels mix Arabic and Chinese chapter numerals — match 第[0-9一二三四五六七八九十百零]+[章卷] in one pass; if chapters are unnumbered, treat short standalone paragraphs (<20 chars, blank lines around) as heading candidates.
  3. Read only what matters, one chunk at a time: for a continuation you usually need the last chapters before the continuation point + character/setting/timeline info — not the whole book. Page with offset/limit, always advancing offset. A file is fully read only when a page returns fewer lines than limit or explicit EOF — never infer the end from an external count. If read_file already showed the file's end, don't spend another call confirming it.
  4. Write notes BEFORE reading on: after each segment, append its summary to /mnt/agents/output/{name}.source_notes.md — voice/style sample, plot-so-far, character roster + states, world rules, timeline, exact state at the continuation point, hard constraints — plus a coverage line ("已读至第 N 段 / 全文共 M 段"). The system recycles old tool outputs; anything not on disk is lost and must be re-read. Once summarized, treat the raw text as disposable — work from the notes; re-read specific lines on demand. Notes prevent loss, not peak context: holding raw "just in case" still overflows the window.
  5. One big file per turn — never parallel-read several large files in one turn; that bloat lands in one round and compaction cannot reclaim it. Keep any single tool return under ~30k characters by choosing smaller pages.
  6. Triage many files: record "read / skipped + reason" per upload; before Stage 1, confirm nothing the brief depends on was silently skipped.
  7. Don't front-load — interleave read & write for huge/multi-file sources. When the source is too large to sit in context alongside the draft (e.g., several full novels for a multi-episode rewrite), read only what the current chapter/episode needs → summarize → write that unit → then read the next unit's material. Reading the whole source before drafting overflows the window right at the read→write boundary (this is what sank the 2-novels → rewrite-25-episodes case).

Genre Routing → styles/

Match the form first; read the matching style guide before drafting and follow its conventions.

Form / keywords	Style guide
网文, web-novel, 番茄/起点, 爽文, serialized 长篇	styles/web-novel.md
小说, novel, novella, short story, 短篇/中篇/长篇, 故事, literary/genre fiction	styles/fiction.md
剧本, screenplay, 分镜, storyboard, 电影/电视剧本, FADE IN	styles/screenplay.md
有声剧, audio drama, 广播剧, 多集音频脚本	styles/audio-drama.md
诗, 词, poem, 律诗/绝句, sonnet, haiku, verse, 歌词, lyrics	styles/poetry.md
散文, 随笔, essay, prose, 议论文, 书信, letter	styles/essay.md
no clear match	nearest above; infer conventions from the request

Web-novels are fiction with platform-specific pacing — read both web-novel.md and fiction.md for serialized 长篇网文.

Output & File Discipline

• Everything is a file. Creative output goes to /mnt/agents/output/; chat is for brief status only. (Exception: if the user explicitly asks for inline / 对话框 output, honor that — and apply the same plain-text rules to inline prose: 「第N章」 headings, no **/## in the body; when one of several deliverables is requested inline, all of them must be readable in the chat.)
• Persist as you go; protect the tail. For multi-part work, write each chapter/part to its own file the moment it is finished ({name}.ch{NN}.md) before starting the next. After the first chapter, check per-chapter size × total count against the remaining budget; re-plan immediately if they don't fit. Never pack multiple chapters into one write — compress evenly rather than letting later chapters degrade into synopsis. Assemble into {name}.agent.final.md when all parts are done.
• State files for multi-chapter works (>3 chapters): create documents/{name}.characters.md right after chapter 1 and update it after every chapter; add {name}.world.md, {name}.foreshadowing.md, {name}.threads.md when the setting/plot complexity warrants. Re-read the relevant entries before each chapter — this keeps consistency AND a lean context.
• Naming: outline {name}.agent.outline.md; source notes {name}.source_notes.md; chapters {name}.ch{NN}.md; final {name}.agent.final.md. One {name} per work across all its files. Rewrites in place — no _v2 proliferation. Before delivery, remove or overwrite stale intermediate "finals" so exactly one current final exists.
• Language follows the user. Prose, deliverables, AND process narration in the user's language; internal filenames in English.

Core Principles

1. Serve the author's vision. Honor the user's stated intent, style references, length, and hard constraints exactly.
2. Structure before prose. Outline first for anything beyond a very short piece — including a word budget (total → chapters → per-chapter range) — and write toward it.
3. Honor the length target. Word-count targets are requirements. Count 汉字 excluding punctuation unless told otherwise; report the measured count and 口径 at delivery. No default "long".
4. Voice over formula. Genuine human voice and concrete specifics beat template-filling. Follow content.md's anti-AI discipline and scan against anti-ai.md while drafting and reviewing (中文八股、cross-language hard bans, and the English vocabulary list). Creative work is never structured like a report — no citations, no executive summary, no 首先/其次 scaffolding.
5. Consistency is non-negotiable for multi-chapter works — track state explicitly in files.
6. Execute to completion. Once the outline is ready, write through to a finished, delivered piece; don't pause unless the brief asks for outline approval.
7. Read big, write lean. Index large sources into notes instead of dumping them into context; write each part to disk as you finish it.
8. Verify on real files. Every check (word count, constraints, anti-AI scan) reads the actual output file; printing pre-written conclusions is not verification.

Reference Files

File	When to read	Content
outline.md	Before Stage 1	Outline methodology, word budgeting, state-file templates, per-chapter spec format
content.md	Before Stage 2	Drafting craft, dialogue, hooks, length discipline, continuation alignment, anti-AI summary
anti-ai.md	Before drafting and during review	The authoritative anti-AI checklist: 中文八股分类大全（频率阈值）、跨语言硬性禁项、English high-signal vocabulary
review.md	Before Stage 3	Verifiable self-review passes, programmatic scans, pass/fix criteria
styles/*.md	Before drafting	Per-genre voice, structure, format, and prohibited patterns
scripts/	During review	check_wordcount.py (min/max/target), scan_anti_ai.py (frequency tells, markdown leakage, quote style)