SuperAPI AI Directory

crossframe-review

General & other

View repository

Updated 10-16 08:46

About this skill

CrossFrame Review

When to Use This Skill

  • Use only after explicit CrossFrame Review invocation or after crossframe-suite routes a CrossFrame output into the review gate.
  • Use to check reasoning fidelity, evidence boundaries, source anchors, concept drift, article-body collapse, and repair steps.
  • Do not use as a generic code review, prose review, or ordinary critique skill.

Packaged Source Note

This AAS-ready copy preserves the original CrossFrame skill body below. Chinese remains the canonical semantic layer; English metadata is only for discovery, installation, and repository review.

Limitations

  • The skill body is intentionally Chinese-canonical; English metadata is for discovery and does not replace the original Chinese terms.
  • Use only after explicit CrossFrame invocation or crossframe-suite routing; do not apply it as a generic default reasoning layer.
  • It structures analysis, drafting, and review, but does not replace source verification, domain expertise, or legal, medical, or financial judgment.

If the review object comes from a continuous workflow of multiple CrossFrame skills, first read ../crossframe-suite/SKILL.md to reconstruct the expected scheduling chain, then determine whether there are missed triggers, false triggers, or skipped quality gates.

This skill only performs review and repair suggestions; it does not replace crossframe generating diagnoses, nor does it replace crossframe-essay writing articles. Chinese is the authoritative semantics; English is only for skill ids, file names, and interface descriptions.

Lightweight entry rules

After each trigger, first read the canonical skills rather than copying their main text:

  1. Read ../crossframe/SKILL.md。
  2. Read ../crossframe/references/read-routing-map.md。
  3. Read ../crossframe/references/runtime-read-policy.md and ../crossframe/references/continuity-closure-map.md, determine which v5.0 continuous co-reading bundles should have been triggered; only when bundle descriptions, source anchors, or closure details are needed, then read ../crossframe/references/continuity-bundles.md or the specific bundle file in a targeted way.
  4. If the review object is a deep, audit, high-responsibility, public institution, intimate relationship, long-term evolution, or article-type output, read or check ../crossframe/templates/read-state-capsule.md as specified in v5-read-state-capsule whether it exists and is reused downstream.
  5. Read on demand ../crossframe/worksheets/source-continuity-check.md and ../crossframe/worksheets/source-anchor-integrity-check.md, check whether the closure is complete, and whether the central proposition and action boundaries can point back to the capsule source anchors.
  6. If the review object is an article, long-form piece, commentary, thought article, newspaper reply, or "modern editor comrade tone" output, read on demand ../crossframe-essay/SKILL.md。
  7. Read this directory's protocols/review-protocol.md and templates/review-report.md。
  8. If it involves article drafts, citations, retrieval materials, or voice, additionally read protocols/article-review-protocol.md。
  9. If it involves public facts, real institutions, platforms, policies, people, companies, latest facts, AI/process artifacts, critical articles, or source use, read ../crossframe/references/source-ledger-workflow.md, check whether the source ledger fields are complete.
  10. Read on demand this directory's references/ the scoring table, failure type table, and evidence boundary checklist.

Do not copy the CrossFrame main skill, article skill, eval, examples, or complete cases into this skill's output. During review, cite only necessary rule names, trigger points, and evidence locations. If v5-read-state-capsule already exists, downstream reuses the capsule by default, and the source index must not be repeatedly read in full blocks for the sake of review.

Review objectives

Determine whether an output truly completes CrossFrame's minimum reasoning chain:

  • Clearly diagnose or write the object.
  • Distinguish facts, explanations, evidence gaps, and judgment tiers.
  • Pass through the object gate, evidence gate, scale gate, responsibility gate, and observation gate.
  • Form at least two mechanism candidates, or explain why there can only be one.
  • Perform fidelity checks on high-risk concepts that carry a judgment role, rather than treating terminology as a conclusion.
  • Perform source structure continuity checks on high-risk concepts that belong to v5.0 continuous sections, rather than reading only a single concept card.
  • Perform source anchor integrity checks on the central proposition, mechanism candidates, high-risk concepts, action boundaries, article-type translation, and writing techniques; content that cannot point back to the capsule must not be written as the original meaning of CrossFrame v5.
  • Give retraction conditions, next-step observations, or low-condition action boundaries.
  • Article-type outputs must first have a structural insight draft, then write the main text.

Must-catch failures

Once the following problems appear, they must be clearly located in the review; when severe, directly judge as unqualified:

  • Concept stacking: only piling up words such as "undertaking, reflow, scale, anti-capture" without falling back to facts and behavior.
  • Pseudo-reasoning: give the conclusion first, then decorate it with terminology; no mechanism candidates, counter-conditions, or evidence gates.
  • Missing fact boundaries: treating rumors, AI reports, self-assessments, search summaries, or explanations as facts.
  • Skipping the draft: article-type outputs are written directly, with no structural insight draft or equivalent internal skeleton.
  • Character judgment: turning structural diagnosis into labels such as "this person is bad, lazy, incompetent, pathological."
  • Fabricated citations: inventing original text, page numbers, sources, or authors' views; writing uncertain original sentences as direct quotations.
  • Source-checking takes over the thesis: retrieved material determines the article's stance, and CrossFrame becomes merely a collage shell for materials.
  • Strong judgments overreach: judgments about discipline, reputation, rights, resources, or public memory lack proposition verification and appeal/counter-evidence entry points.
  • Scale whitewashing: using macro narratives to cancel low-scale suffering, chains of responsibility, or specific dereliction of duty.
  • AI compliance theater: treating AI-generated material, polished reports, or self-assessment texts as independent strong evidence.
  • Continuity fidelity failure: what should trigger continuity-bundles.md the co-reading package instead reads only a single concept card, a single protocol, or a single excerpt before making a judgment.
  • Missing capsule: tasks that should be generated by crossframe lack a capsule, causing essay/review to each reread the source index or invent routing. v5-read-state-capsule has no capsule for its task, causing essay/review to each re-read the source index or invent routing.
  • Source anchor failure: the central proposition, mechanism candidates, high-risk concepts, action boundaries, or article translation cannot point back to capsule source anchors, yet are written as CrossFrame v5 original meaning.
  • Downstream rereads the source as a whole block: when a capsule already exists, essay or review again reads the v5 source index, the complete co-reading package, or the material selection map as a whole block, causing source boundary drift.
  • Selector compression failure: the mode/role or article type selector does not fully render options, recommendations, and waiting for user reply.
  • Technique boundary-crossing failure: writing techniques add new facts, strong judgments, punchlines, or metaphorical proof, crossing the draft and capsule source boundaries.
  • Source-use boundary-crossing failure: writing popularity, institutional statements, PR copy, AI-generated material, self-assessment texts, or secondhand paraphrases as verified facts.
  • Missing source ledger: public, critical, article, or high-responsibility outputs involve real objects, but there is no source, time, source type, supported proposition, what it cannot prove, evidence tier, place of use, downgrade reason, and what still needs corroboration.
  • v5 reality protection failure: when involving AI procedural artifacts, weak signals, opacity, absence of institutional infrastructure, inability to exit, malicious compliance, metaphor drift, instrumentalization, or open assertion withdrawal, the corresponding v5 concept card and co-reading package are not read.

Output protocol

Use by default templates/review-report.md. The final review must include:

  • Review object
  • Fact boundaries
  • Trigger rules
  • Score/grade
  • Key issues
  • Evidence location
  • Repair suggestions
  • Whether qualified

If the user only wants a one-sentence conclusion, still retain the minimal structure of "whether qualified + main failure points + next-step repair."

Suite scheduling visibility

When this skill is invoked via crossframe-suite as the default quality gate, and the user has not explicitly requested "only a review/full review report/no article," the review does not take over the final output:

  • The review object is the already formed 结构洞察底稿 and 文章正文。
  • If it passes with A/B or minor revisions, feed the issue back upstream for correction; the final visible deliverable is still # 结构洞察底稿 + # 文章正文, with at most one added line of short quality-gate summary.
  • On C/D/F or hard failure, block publication and require going back to the corresponding upstream to supply the underlying draft, supply evidence boundaries, or rewrite the body; if the user did not ask for review-only, do not output only a review report in place of the repaired article.
  • templates/review-report.md Only when the user explicitly requests a full review report, reviews only existing output, or a hard failure requires explaining the blocking reason, should it serve as the main output.

Qualification determination

Scoring is only auxiliary; hard failure takes priority:

  • A: 90-100, qualified.
  • B: 75-89, conditionally qualified, usable after minor revisions.
  • C: 60-74, unqualified, requires major revision.
  • D: 40-59, unqualified, requires redoing the main reasoning.
  • F: 0-39 or triggering hard failure, high-risk failure.

When personality judgment, fabricated citations, skipping the article's underlying draft, overreaching strong judgments, complete absence of evidence boundaries, or continuity fidelity failure is triggered, it cannot be judged qualified even if the writing is fluent.

Repair principles

Review output should prioritize actionable repairs by default, not rewriting the full text. Unless the user requests "direct rewrite," only provide:

  • Factual boundaries that should be supplied.
  • Routing that should be read or declared.
  • Mechanism candidates that should be added.
  • Judgments that should be downgraded.
  • Personality judgments, fake citations, and jargon-stacking sentences that should be deleted or rewritten.
  • Structural insight underlying-draft items that the article should add.