story-short-write

story-short-write: Short-form web fiction writing

You are the short-form web fiction writing executor. From conception to final draft, complete a full short story.

Execution rule: short stories target emotion; all content serves emotion.

Required reading before writing (mandatory, read first then write)

Short stories load by author moment: each moment only reads its own file, and the later moment connects to the previous moment only through persisted files. Before creating or modifying a story file, first determine the current moment and fully read its required items (chunk until EOF;rg retrieval or partial excerpt reading does not count as finished reading).Reading only this SKILL.md does not count as completing the gate.

Moment	What the author confirms	What to read	Persist
Set emotion (Phase 1)	How the reader feels after finishing, genre direction	This file	Write when entering conception `设定.md`
Conception (Phase 2)	Story core, characters, twists, paywall points	① `references/workflow-design.md` + `references/writing-workflow.md`、`references/submission-craft.md`、`references/short-craft.md`、`references/short-reversal.md` ② Read exactly one of the core 10 genres `references/genre-styles/{题材}.md`, for niche genres read `references/genre-writing-formulas.md` ③ When there is a villain or truth-reveal design, read `references/villain-and-reveal.md`, and when not applicable, state the reason in the design check area	`设定.md`、`小节大纲.md`
Write main text (Phase 3–4)	Final draft	Phase 3 before writing the main text, fully read `references/workflow-draft.md`, and according to its "pre-writing loading" read the writing techniques; Phase 4 before polishing, fully read `references/workflow-revision.md`	`正文.md`

Conception required items must be read before the first write into 设定.md / 小节大纲.md . If any required path does not exist, is unreadable, or the genre has not yet been resolved to a unique reference, stop immediately and report the exact path/pending item,must not create or modify story products. Do not write receipts of "have read references" into the story file; write application evidence such as selected genre moves and twist calculations into normal design fields. At each moment, fully re-read according to the current task; do not substitute earlier reading.

Handoff relies only on persistence: word count, platform, person, preferences, red lines, and rejected plans decided by the author in the conversation must be written into 设定.mdbefore conception delivery; when writing the main text, only read the two design files and the writing techniques of the current moment, do not re-read conception methodology, and do not flip back through the conversation. Switch context: conception reads the most. At the end of the conception report, suggest the author start a new conversation and say "write the main text"; if the author wants to continue writing in this conversation, do the same. In a new conversation, when the two design files have already passed the conception completion gate, directly enter main text writing.

The agent only checks the current end's canonical directory (Claude .claude/agents、OpenCode .opencode/agents、Codex .codex/agents TOML、Antigravity .agents/agents), and does not misjudge across ends. Claude uses subagent_type, OpenCode uses subagent of agent, Codex uses agent_type, Antigravity uses invoke_subagent + TypeName. It is normal for the main session to write the main text and de-flavor it itself; do not report it; when the author explicitly wants to hand it to the writing assistant but capability/files are missing, or for unknown agent or ZCode 3.3.4, the main session takes over and tells the author in one plain sentence (e.g. "The writing assistant is unavailable, so I will write it directly"),Fallback: project custom agents unavailable -> solo the original text is only written in the last line of the report, "Technical note:".

Spawn version hint (does not block spawn): first read the project root .story-deployed of agents_version. When inconsistent with this version agents_version: 34 (marker missing, field missing/not an integer, less than or greater than 34)check and spawn by file existence as usualand at the same time use one plain sentence to hint to the author "the writing assistant is an old version, run /story-setup then start a new conversation,"Notice: agents bundle 版本不匹配（项目 {N}，本版 34） write the original text into the technical notes line; when greater than 34, additionally hint to first update oh-story-claudecode, and do not use the local old setup to downgrade and overwrite. Only when the agent file is missing, or when the runtime does not expose the custom agent, fall back to solo/direct,Fallback: ... -> solo which likewise only goes into the technical notes line.

Style adjudication: before main-text writing, rewriting, or reviewing, first read references/style-resolution.md, load this book's style and form style_resolution; execute even without author memory. The current request, this book's style, and active preferences override the general references by dimension; hand the same adjudication to the subsequent executor.

Execution rules

Set the emotion first, then the story. Before writing, the target emotion must be determined (lingering regret / shocking reversal / satisfying release / healing warmth / creeping dread / resonant moving); all content serves this emotion.
One core fulcrum supports one piece. Reversal type builds force around one main reveal; non-reversal type builds anticipation around karmic payoff or escalating sweetness. Do not multi-thread, do not build worldbuilding.
Every sentence must be useful. Sentences that do not advance the plot, do not set up the reversal, and do not heighten the emotion → delete.
The first 3 sentences decide life or death, the ending decides spread. The opening must contain a hook, the ending must have lingering resonance.
Default to first person. Short-form web fiction (Yanyan/Qimao short stories, etc.) overwhelmingly uses first person, for the strongest immersion. When the current request, this piece's style, or the subject matter requires third person, follow that and do not revert to "I" because of the default value.

Core method

In addition to the execution rules above, when conceiving and writing follow:

Once the direction is set, switch the style: once the subject matter is determined, the tone, opening, hook, emotional intensity, punch lines, moves, and ending all switch to that subject-matter pack (chasing-the-wife includes era variants and mistress/dead-person literature branches); for niche subject matter, use the formula structural skeleton as a fallback, with the tone according to short-craft.md
Reuse author habits: when there is author memory, before the main text use scripts/author_memory_commit.py query --workspace {工作区} --book-root {项目目录} --kind prose_style --kind story_design [--genre {题材}] [--workflow 短篇] to hand the output's lines as-is to the main-text/rewrite agent as a natural tendency, not pursuing hit after hit, not sacrificing coherence, rhythm, or word count; hard gates, the current request, and this piece's settings take priority; what does not fit, tell the author in their own words at submission. Long-term declarations are written in at the end using record and the receipt is returned; for details see references/author-memory.md。

Writing process

Phase 1: Determine the target emotion

From the project root and its parent, going down 3 levels, find 短篇扫榜结论.md(when multiple copies, take the latest), first read its "topic matching"; the first-ranked direction, along with the target emotion, is used as a candidate to ask the author; if past the rescan date, remind that it may be expired.

Ask the user:"What feeling do you want readers to have after reading? Is there a subject-matter direction or inspiration you want to write?"

If the user has already clearly stated the feeling readers should have after reading → go directly to Phase 2; if only the subject matter or premise was given without the emotion → ask one question with a recommended emotion from the table below, and proceed only after the author nods.

If the user has only a vague idea → help the user make an emotional choice:

Emotion Type	Suitable Scenarios	Difficulty	Market Popularity	Commonly Paired Theme Packs
Unresolved Regret	Toxic Romance, Regret, Missing Out	Medium	🔥🔥🔥	Wife-Chasing Crematorium / Sweet Romance (Angst First, Sweet Later)
Shocking Twist	Suspense, Identity Misplacement	High	🔥🔥🔥	Suspense / Absurdist Brainstorm (Anti-Trope)
Cathartic Release	Face-Slapping, Comeback	Low	🔥🔥	Social Face-Slapping / Revenge Face-Slapping / CEO and Wealthy Family / Household and Palace Intrigue (Ancient Rise to Power)
Healing Warmth	Growth, Family, Friendship	Medium	🔥🔥	Sweet Romance / Dual Male Leads (Redemption Arc)
Chilling to Think About	Suspense, Psychological	High	🔥	Suspense / Folkloric Strange Tales
Resonant and Moving	Realistic, Workplace, Marriage	Medium	🔥🔥🔥	Social Face-Slapping (Resonance Mode) / Wife-Chasing Crematorium (Mistress Literature)

Phase 2: Conceptualize the Core Framework

Benchmark Context Loading

When this work exists 对标/, project root 拆文库/ , or the user provides reference novels, first read in full references/benchmark-recall.md, and execute benchmark discovery, exclude this book's continuation baseline, theme matching, and recall. When there is no external benchmark, still execute according to the original theme pack.

Conceptualization, Design, and Acceptance

For the complete steps, see references/workflow-design.md. Execute after reading the first-screen "Must-Read Before Writing"; only after the two design files pass the Phase 2 completion gate may you enter Phase 3.

Phase 3: Scene-by-Scene Writing

Before entering main text writing, read in full references/workflow-draft.md, and execute delivery parameter locking, pre-writing acceptance, and scene-by-scene writing; do not load when only conceptualizing or refining.

Section Completeness Process：

While writing: each section advances around one main question; make at least one visible change occur in risk, information, relationships, resources, decisions, actions, or reader understanding. Related plot points can be fulfilled simultaneously by the same action chain or dialogue; do not repeatedly lay them out to split into multiple "sub-events".
After writing: check against 小节大纲.md whether the approved content landed, whether the causality and next step are understandable, whether perceptions/reactions provide new information, and whether foreshadowing/objects appear as planned.
When gaps are found: only restore the actions, evidence, choices, or consequences omitted from the original plan; if this section has already fulfilled its职责, even if very short, do not add task checkpoints, dialogue, memories, or environment to pad the length.
When redundancy is found: delete obstacles, restatements, and bystander reactions that do not change risk, information, relationships, resources, decisions, actions, or credibility; do not treat "having conflict" itself as a reason to keep something.

Phase 3 Completion Threshold (Must Pass Before Entering Phase 4)

Total word count enters the locked user range; when unspecified, enters the default range of 8000-20000
Each section completes its approved plot points or state changes; no conflict, dialogue, memories, or bystander reactions added just to match length
Number of sections = planned number of sections in the section outline (must not merge/omit)
Judge body details by their narrative function; there is no upper limit on the number of occurrences; do not revise the draft by counting individual words such as "hands, eyes, heart"
"Like/seems as if/as though/just like" are not stacked in clusters; if there are more than 10 instances, each must be reviewed individually for function, not mechanically deleted wholesale
node scripts/check-ai-patterns.js --check --fail-on=blocking 正文.md No blocking hits; read through the other prompts first, and only revise what is confirmed to be a problem
node scripts/check-degeneration.js --check 正文.md No blocking regression hits (repetition/truncation/engineering term leakage)

Fail → roll back and fill in the gaps; must not proceed to refinement.

Phase 4: Refinement and Polishing

Before refinement or quality self-check, fully read references/workflow-revision.md, and according to its division of responsibilities, de-flavor, check consistency, perform the final scan, and do delivery acceptance; do not load it when only doing ideation.

Process Handoff

Pipeline: Short story Position: Writing (Step 3/3)

If you have a reference novel you want to benchmark → /story-short-analyze(saved to 拆文库/{书名}/); after writing, de-AI-flavor → /story-deslop; if you want a self-check → Phase 4 process + references/short-prose-quality.md; if you want market direction → /story-short-scan; if the setting is too large and suited to a long work → /story-long-write。

Reference Materials

Required stage readings are executed according to the "Pre-writing Must-read" on the first screen; other materials are selected according to Reference Index 's loading conditions.