SuperAPI AI Directory

story-review

General & other

View repository

Updated 10-16 09:45

About this skill

story-review: Multi-perspective adversarial review

Spawn version prompt (does not block spawn): first read the project root .story-deployed of agents_version. When inconsistent with this version's agents_version: 35 (marker missing, field missing/non-integer, less than or greater than 35)check and spawn as usual based on file existence, but only check the canonical directory of the current runtime; at the same time, in 「这次怎么审的」, use one plain-language sentence to remind the author 「审稿助手是旧版,运行 /story-setup 后新开会话」,Notice: agents bundle 版本不匹配(项目 {N},本版 35) write the original text into the technical notes line; when greater than 35, additionally prompt to first update oh-story-claudecode, and do not use the local old-version setup to downgrade and overwrite. The version prompt itself does not trigger degradation; if the file is missing or abnormal, the tool is unavailable, there is a recursion limit, or the call fails, handle it according to Phase 0, and write the technical notes as Fallback: ... -> solo。

You are responsible for coordinating the review, identifying problems in the novel's structure, characters, prose, and setting, and giving revision suggestions grounded in the original text.

Iron rule of execution: review is about finding problems, not verifying correctness.

Style adjudication: first read references/style-resolution.md, load this book's style, and reuse the caller's style_resolution and author preferences; when invoked independently, adjudicate according to the current request and this book's style. Hand the same adjudication to subsequent executors.

Author preference boundaries

Only accept the author preferences already available to the caller; do not query or record author memory; when invoked independently, do not load memory files. Preferences can only help interpret intent and organize the report; they cannot reduce rubric severity, judge factual conflicts as no problem, or skip platform gates; the current request still takes priority.


Review Mode selection

  • /story-review or /story-review full → prefer to spawn all 4 Agents; if currently already inside a sub-agent, the core Agent is not deployed/abnormal, or spawning fails, automatically degrade to solo.
  • /story-review lean → prefer to spawn story-architect + consistency-checker; if currently already inside a sub-agent, any required Agent is not deployed/abnormal, or spawning fails, automatically degrade to solo.
  • /story-review solo → do not spawn Agents; perform the basic review in the current session.

Phase 0: Pre-check and degradation (must be executed first)

  1. Determine the request mode: parse the full、lean、soloin the user input; when unspecified, the target mode is full。
  2. Confirm whether spawning is allowed: if currently already executing inside a sub-agent/Agent, do not recursively spawn, and directly degrade to solo。
  3. Identify ZCode capability boundaries: if currently running on ZCode and the project uses .zcode/, ZCode 3.3.4 does not execute project/plugin custom agents; do not attempt to spawn agents with the same names just because agent files for other platforms exist on disk; directly degrade solo, and write the technical notes as Fallback: project custom agents unavailable -> solo。
  4. Check core Agent deployment status(only check the canonical directory of the current runtime; do not misjudge because files for other platforms exist):
    • Claude Code checks .claude/agents/, OpenCode checks .opencode/agents/, Codex checks .codex/agents/, Antigravity checks .agents/agents/
    • full required agents:story-architect、character-designer、narrative-writer、consistency-checker
    • lean required agents:story-architect、consistency-checker
    • For each required Agent file:
      • Claude Code agent(.claude/agents/): Read the frontmatter and confirm that name: is exactly consistent with subagent_type; if the frontmatter is missing, unparsable, or the name does not match, treat it as a malformed agent.
      • OpenCode agent(.opencode/agents/): The file name is the agent name (OpenCode does not require writing name:in the frontmatter); read the frontmatter and confirm that mode: subagent and permissions: rule lists exist and are parsable (2.x uses the plural permissions:; the old singular permission: is treated as pending redeployment); if the frontmatter is missing or unparsable, treat it as malformed.
      • Codex agent(.codex/agents/): The file name is {agent}.toml, the TOML must be parsable, and it must contain name、description、developer_instructions;name must be exactly consistent with the target agent.
      • Antigravity agent(.agents/agents/): The path is .agents/agents/agent-name/agent.md(agent-name is the target agent name), the frontmatter must be parsable, and name must be consistent with the target agent,mainAgent: false、subagent: true、tools must be non-empty; if missing or inconsistent, treat it as malformed.
    • If any file required by the target mode is missing or malformed,do not attempt to spawn the missing/abnormal Agent; automatically downgrade to solo, and at the beginning of the report tell the author in one sentence, "The review assistant is missing or corrupted, so this time I will review on my own; after running /story-setup you can review from multiple perspectives"; write the downgrade reason missing agents -> solo / malformed agents -> solo and the problem file into the Fallback and Files columns of the technical notes line.
  5. Confirm that the Agent tool is available: Claude/OpenCode/Codex require the current runtime's sub-Agent/Task invocation capability; Antigravity requires invoke_subagent; if unavailable, directly downgrade to solo, and write in the technical notes Fallback: agent tool unavailable -> solo。
  6. Runtime failure downgrade: If any Agent spawn returns a failure,subagent_type / agent / agent_type / TypeName is unavailable, frontmatter/TOML runtime parsing fails, or the sub-Agent cannot start, stop spawning further, switch to solo to re-review, and write in the technical notes Fallback: spawn failed -> solo and the name of the failed agent; do not treat partially successful Agent results as full/lean conclusions.
  7. Determine the actual mode: Write both the requested mode and the actual mode into the technical notes line at the end of the report.

Review baseline and reference material rules (must be followed)

Reference files supplement the review standards; when unreadable, use the built-in baseline package below.

The report is for the author (must be followed)

The report is written for the author: what was reviewed, what needs to be changed, why (explain in terms of reader experience and story consequences, with citations from the original text), what the author needs to decide, and next steps. Reviewer names, S1–S4, Gates, detector category names, script names, PASS/FAIL, and file field names do not enter the main text; for locations, write "Chapter N 'quote'" or "Chapter N, paragraph M". Convert priorities into plain language (section headings follow the template): S1, S2 → Must fix,S3 → Suggested fix,S4 → Optional. The execution path is written only on the last line of the report, in a fixed format:

技术备注:Mode {请求}→{实际} · Fallback {none | project custom agents unavailable -> solo | missing agents -> solo | malformed agents -> solo | agent tool unavailable -> solo | spawn failed -> solo | subagent recursion guard -> solo} · Rubric {fanqie | qidian | zhihu | generic} ({file | embedded})[ · Files {缺失或异常的 agent 文件}][ · Notice {版本不匹配原文}]

Reference material parsing order

When reference files can be read, try in the following order, and use the first match:

  1. {项目根}/.claude/skills/{规范路径}(installed within the Claude Code project)
  2. {项目根}/.opencode/skills/{规范路径}(installed within the OpenCode project)
  3. {项目根}/.codex/skills/{规范路径}(installed within the Codex project)
  4. {项目根}/.zcode/skills/{规范路径}(installed within the ZCode project)
  5. {项目根}/skills/{规范路径}(OpenClaw / Reasonix / generic deployment, also the development environment for this repository)
  6. {项目根}/.agents/skills/{规范路径}(the real skill root inside an Antigravity project; Codex / Reasonix may also scan this directory or its symlink)
  7. The directory where the current runtime loads this skill, or a directory of the same name in its accessible global skill search path {skill-name}/... directory

It is normal for it not to exist in the preceding layers; this is not deployment corruption./story-setup will actually copy the 13 skills to .agents/skills/for Antigravity, copy to .zcode/skills/for ZCode, and copy to skills/for OpenClaw / Reasonix / generic. Codex project deployment does not copy the skill body itself; this skill is loaded by Codex from the skill root, and references usually hit layer 6 or layer 7. Do not manually copy references/ into .codex/skills/—manual copies are not managed by story-setup and will silently become outdated after upgrades.

The canonical paths are as follows; writing only bare file names is prohibited, and misreading another skill's references across skills is prohibited:

Purpose Canonical path
General quality checklist story-review/references/review-quality.md
General content scoring rubric story-review/references/quality-rubric.md
Methods for removing AI flavor story-review/references/expression.md · story-review/references/narrative-focus.md · story-review/references/story-logic.md
Plot cycle/climax formula story-review/references/plot-core-methods.md
Character relationships/favorability story-review/references/character-relations.md
Dialogue quality story-review/references/dialogue-mastery.md
Review forbidden words story-review/references/banned-words.md
Platform rubric story-review/references/rubrics/{fanqie,qidian,zhihu}.md
Punctuation precheck script story-review/scripts/normalize-punctuation.js
AI sentence pattern precheck script story-review/scripts/check-ai-patterns.js

Built-in review baseline package (mandatory when paths are unreadable)

If the above reference files are unreadable in the current project,do not downgrade the review to having no rubric, and do not say in the report "unable to load the specific rubric" and then stop using the standard. You must use the built-in baseline package in this section, and write the Rubric source in the technical notes line as embedded。

General web fiction content rubric:

  • Core selling point: whether this chapter advances around a clear selling point; if the selling point cannot be seen, at least S2.
  • Conflict advancement: whether this chapter has obstacles, choices, costs, or relationship changes; if it only explains/chats/summarizes, at least S2.
  • Task blockage point: when a character is stuck doing something, whether the blockage produces changes in information, relationships, costs, choices, or foreshadowing; if the blockage contains only process details and deleting it does not affect the story, at least S3.
  • Emotional curve: whether there is setup, escalation, release, or reversal; if the emotion is flat or abrupt, at least S2/S3.
  • Hook and expectation: whether the beginning or ending creates follow-up questions; if there is no suspense or unfinished expectation, at least S2.
  • Opening freshness (only for the opening/first 3 chapters): does the opening have a specific character/situation entry point, or is it the default formula of the same genre (can it be swapped wholesale into any similar book)? "Has a hook/does not open with weather" does not exempt homogeneity; a formulaic opening is at least S3 even if it has a hook, and if it clashes with the same-genre template overall, S2.
  • Character motivation: whether behavior fits goals, personality, situation, and relationship pressure; distortion for the sake of plot is S1/S2.
  • Dialogue quality: whether there is subtext, information control, and character differentiation; manual-style dialogue is at least S2.
  • Setting consistency: does not violate established rules, timeline, or character attributes; clear factual conflicts are usually S1.
  • Prose naturalness: concrete, perceptible, action carries information; AI tone, clichés, and summary style are rated S2/S3 according to impact.
  • Sentence-length rhythm: Judge according to the book's style, semantic continuity, and scene function; mechanically choppy sentences, action lists, or bloated long sentences are rated S3/S2 based on impact, and are not reviewed against a uniform word-count or sentence-length template.
  • Punctuation rhythm: Whether punctuation serves the tone/character voice; turning everything into periods, randomly piling up question marks/exclamation marks, or residual ……/—— forced pauses are rated S3/S2 based on impact.
  • Verification of specific word-count expressions: When the main text uses specific word-count expressions such as "these five characters / just four short characters / three characters and it lands / eight characters slam down" to evaluate dialogue, inscriptions, letters, thoughts, or danmaku, the counting basis, machine verification result, and narrative necessity must be confirmable; when the word count cannot be ensured to be correct, handle it as a text naturalness issue and suggest changing it to non-specific-number expressions such as "the moment this sentence landed," "those few characters," or "as the words fell."
  • Format readability: Short paragraphs, independent dialogue, no extra blank lines; if formatting obstructs reading, rate S3; if severely chaotic, rate S2.
  • Plot loop: goal → obstacle → action → cost/feedback → new expectation; missing goal/obstacle/feedback is usually at least S2.
  • Climax construction: Check buildup, emotional落差, and payoff; strong reversals may use buildup → false victory → collapse → reversal/payoff, and missing foreshadowing or payoff is usually S2/S3.
  • Relationship progression: The scale of interaction must match the current relationship stage; boundary-crossing intimacy, sudden trust, and sudden hostility all require foreshadowing, otherwise rate S1/S2 based on impact.
  • Foreshadowing status: Foreshadowing status must be trackable; foreshadowing density serves only as a structural risk indicator, and unless it directly causes comprehension confusion, it is not escalated to S2+.

Three-layer de-flavoring fallback quick reference:

  • Expression: Are stock phrases, summary sentences, and overly neat sentence patterns mechanically repeated or repeating the meaning? The words themselves are only clues; direct emotion writing, error-correction contrasts, and narration allowed by the style may be retained.
  • Drafting: Are key confrontations or decisions summarized, and are secondary processes overly detailed? Both compression and expansion should preserve the character experience and necessary continuity.
  • Story: What is the reason for the action, where does the basis for resolution come from, and what changed before and after? Quiet scenes can carry relationship and cognitive changes, and conflict is not forced.
  • Handling principle: Only output findings when there is textual evidence; point out reading loss, minimal suggestions, and material limitations; report only, do not make changes, and do not default to splitting short, adding actions, or adding new plot.

Platform fallback summary:

  • Fanqie: Strong opening, strong conflict, high-frequency satisfying points/emotional feedback, low comprehension threshold.
  • Qidian: Self-consistent setting, upgrade path, long-term expectations, worldview carrying capacity.
  • Zhihu Yanyan: Short-story hooks, reversal density, emotional payoff, advancement through information gaps.

Phase 1: Collect content to be reviewed

  1. Determine review scope:
    • The user specified a chapter/file → review only the specified content.
    • The user did not specify → prioritize reviewing the most recently modified main-text files (git diff --name-only main text/setting/outline-related files), otherwise review the current chapter of the current book.
  2. Scope delivery strategy:
    • Prioritize passing file paths, chapter names, and line-number ranges to the reviewer; do not copy the entire book or a large number of chapters in full into every prompt.
    • A single file or short excerpt may include a key excerpt of 300-1200 words.
    • Reviews of multiple chapters/whole volumes/whole books must be batched: split by chapter or file group, output independent findings for each batch, then synthesize; before batching, first read references/batch-review.mdin full (cross-batch state, continuity, out-of-order reminders).
  3. Read relevant supporting materials: main text, related settings, character profiles, outline, tracking/context, foreshadowing files; when missing, mark insufficient evidence in the report.
  4. Identify the target platform and load the rubric:
    • Prefer the platform explicitly specified by the user.
    • Next, read the 目标平台 / 平台 field in the project documentation, for example 设定/题材定位.md、大纲/、拆文报告 , etc.
    • Do not treat .active-book as a platform source; it can only assist in locating the current book title directory.
    • Fanqie Novel → prefer reading story-review/references/rubrics/fanqie.md; when unreadable, use the built-in Fanqie fallback summary.
    • Qidian → prefer reading story-review/references/rubrics/qidian.md; when unreadable, use the built-in Qidian fallback summary.
    • Zhihu Yanyan → prefer reading story-review/references/rubrics/zhihu.md; when unreadable, use the built-in Zhihu fallback summary.
    • Unrecognized platform → prefer reading story-review/references/quality-rubric.md; when unreadable, use the built-in general web fiction content rubric; in the technical notes line write generic and file | embedded。
  5. Form a review baseline package summary: compress the loaded file contents or built-in fallback summary into 5-12 review criteria; both subsequent solo and sub-Agent work must use this summary. The summary must retain one sentence-length criterion: narration defaults to long comma sentences; choppy sentences and telegraphic style are handled at the same level as AI tone, and are not passed just because they are "short".
  6. Deterministic pre-check (report only, do not modify): when the review scope includes local main text file paths, run the script bundled with this skill:
    node scripts/normalize-punctuation.js --check <正文文件...>
    node scripts/check-ai-patterns.js --check --fail-on=blocking <正文文件...>
    node scripts/check-degeneration.js --check <正文文件...>
    
    • Merge the ellipsis、double-hyphen、markdown-divider results as format findings into the report.em-dash For dashes, only adopt the semantic rewriting suggestions from check-ai-patterns.js (see the next item);normalize-punctuation.js the same position reported as em-dash is deduplicated and discarded during merging, to avoid two conflicting findings at the same place, namely "mechanical replacement" and "rewrite by function". In addition, manually check whether punctuation rhythm is periodized throughout or randomly piled up; the script does not replace tone judgment.
    • check-ai-patterns.js Both blocking / advisory are first treated as prose candidates, and are reviewed item by item according to style_resolution, the original function, and the three-layer method; after confirming that specific reading loss is caused, assign severity according to the definitions below. Functional error correction, narration, and pauses are retained, and script levels are not directly converted into findings that must be modified. Detector fixes are for reference only and do not override style and factual boundaries. For the complete semantic review, see expression.md、narrative-focus.md、story-logic.md; report only issues and suggestions, do not make edits.
    • check-degeneration.js Report model degeneration (verbatim repetition/truncation/placeholders/engineering term leakage), each with severity: blocking|advisory: blocking (repetition/truncation/tier1 engineering terms) as S1/S2 prose findings, with the fix suggestion being "regenerate this passage, not rewrite it"; advisory (tier2 chapter/ambiguous words) as S3.
    • These three pre-check scripts are read-only;story-review do not modify the main text, settings, or outline files; when automatic main text repair is needed, it is recommended to switch to /story-deslop. In full / lean mode, only the "Tracking File Maintenance" below allows modification of 追踪/; all modes of batch review may write .story-review/state.mdaccording to batch-review.md, and solo does not write project content except for that status.
    • Default --quote-mode keep, do not treat the Zhihu Yanyan short story's 「」 Treat as an issue; only check the corresponding conversion suggestion when the project explicitly specifies a quotation style.

Unified Findings Schema (must be used by all modes)

All reviewers (including solo) first organize issues according to the unified structure, then convert them for the author according to the Phase 4 or solo template; internal fields are used for merging and sorting, and are not pasted verbatim into the author report.location Must use the tool to read the original file line numbers shown in the results; do not delete blank lines and renumber.

For consistency / factual / causal / rule_boundary -type findings, thefix field should only state the factual unification direction (for example, "unify as the old left-arm wound, and synchronize conflicting places in the main text/settings" or "a single source needs to be adjudicated in timelines A/B"), and should not include literary creation suggestions.

- severity: S1 | S2 | S3 | S4
  category: structure | character | prose | consistency | platform | factual | format | causal | rule_boundary
  location: 文件路径:行号 或 章节/段落描述
  evidence: "引用原文或具体证据"
  issue: "问题描述"
  fix: "可执行修改建议"

Severity definitions (same scale as long-form writing and detectors: S1/S2 = must fix, S3 = recommended to look at, S4 = informational only):

  • S1: will break the mainline, character motivation, world rules, or reader trust, and needs priority fixing.
  • S2: clearly affects chapter effect, retention, pacing, or character credibility, and should be fixed this round.
  • S3: local quality issue, such as wording, minor formatting, or local pacing, can be scheduled for fixing.
  • S4: suggestion or style fine-tuning, does not block release.

Phase 2: Parallel Spawn Agent (full/lean mode)

When the actual mode is still full/lean after executing Phase 0, read references/agent-prompts.md and spawn the four prompts in parallel according to the invocation rules therein; the comprehensive adjudication and report templates are also therein; do not spawn missing Agents. Each Agent does not inherit the parent conversation context, and the prompt is self-contained with paths, scope, and the unified Findings Schema.

Phase 3: Comprehensive Adjudication (full / lean mode)

After collecting reviewer results, merge, deduplicate, and present disagreements according to "Comprehensive Adjudication" in agent-prompts.md.

Phase 4: Output Report (full / lean mode)

When the actual mode is confirmed to be full/lean, use the report template in agent-prompts.md; when degraded to solo, use the solo.md template instead.


solo mode

When degraded or specified as solo, read references/solo.md, and execute according to the process and output format therein.

Tracking file maintenance

For long-form projects, when wrapping up a full/lean review, read references/review-tracking.md。

Process linkage

Pipeline: general Location: review (after writing)

Timing Jump to Command
Need to fix issues found in the review story-long-write / story-short-write Return to the corresponding writing skill to modify
Found AI flavor that needs cleanup story-deslop /story-deslop
Need to re-decompose the benchmark book story-long-analyze / story-short-analyze /story-long-analyze or /story-short-analyze