Home / All Skills / Content & Media Creation / dbs-ai-check
English (EN)简体中文 (ZH)日本語 (JA)
Content & Media Creation · dontbesilent2025/dbskill

dbs-ai-check

Scans copy for AI writing characteristics and outputs a detection report, diagnosing only by default without rewriting. Used when users request to check for AI tone, AI traces, or mechanized expressions.

L1 Basic · Source registration onlyV1 Commit-PinnedDual-pass Audit
Author dontbesilent2025GitHub dontbesilent2025/dbskill ↗Stars 10121License CC BY-NC 4.0（仓库根 LICENSE 首行「Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)」，条款为 Attribution + NonCommercial，Copyright (c) 2026 dontbesilent。**非商业授权**：个人/学习/研究可直接使用，商业用途需单独授权；GitHub API license.spdx_id 为 NOASSERTION（CC 系非 SPDX 标准标识））Commit 8b8e33f1ec
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Rejecting 'anti-AI de-flavoring' is the product stance: identifying AI traits aims to help authors see their own writing clearly rather than pretend to be human; thus, the default mode only diagnoses and does not rewrite.

skills/dbs-ai-check/SKILL.md
**默认只识别，不改。**
What it does: redefines 'AI flavor' as 'too perfect'—written too well, too smoothly, too evenly, without a single spot where even the author hasn't fully thought it through (Principle 1). The detection goal shifts from 'removing traits' to 'uncovering identifiable authorial presence.' It also clarifies that 'spending time de-AI-ing is worse than spending time figuring things out,' preventing it from devolving into paraphrase-based rewriting.

The detection mode is a point-by-point scan of '22 traits × sequential text order': the report does not group by trait category, but quotes passages sequentially along the original text while labeling trait numbers and severity in place.

skills/dbs-ai-check/SKILL.md
按文本顺序逐处指出问题，不按特征分类。
The report structure is hard-coded as '**Hit X AI fingerprints**' + '> Quoted original text' for each instance + a one-sentence problem explanation + backtick line `Trait #N Trait Name Severity`. Scanning sequentially rather than clustering by traits allows authors to revise progressively along their own text without jumping back and forth among 22 categories.

Three severity levels separate 'almost only AI does this' from 'humans do this too,' avoiding mechanical judgment.

skills/dbs-ai-check/SKILL.md
- 🔴 强信号 — 几乎只有 AI 会这样做
The three tiers are 🔴 Strong Signal / ⚠️ Medium Signal (frequent in AI, but humans may also do it) / 💡 Weak Signal (requires genre and context analysis). This is the key difference from standard 'AI detectors': it grades diagnostic strength first, leaving the decision to rewrite up to the human.

Guided rewriting mode (triggered after the user says 'I want to revise') never ghostwrites: it queries the intention behind each trait sequentially by severity, asking only one question at a time and providing 'revision directions' rather than rewritten text upon answer.

skills/dbs-ai-check/SKILL.md
3. 用户回答后，给出修改方向（不是修改结果）
A 4-step workflow is hard-coded in SKILL.md: query traits by severity (High → Low) → ask only one question per trait and await response → provide direction, not finished copy → offer specific edits only when the user explicitly requests 'help me rewrite,' which must incorporate the user's expressed preferences. The revised copy remains the author's own, aligning with 'revisions must be based on user preference.'

A 22-item follow-up mapping table is embedded directly in SKILL.md, detailing the 'underlying intent' and a sample prompt for each item, turning abstract traits into actionable questioning scripts.

skills/dbs-ai-check/SKILL.md
「1.7 秒」「2.3 秒」这种数字你量过吗？你当时真的感受到的是什么？用你嘴上会说的词来描述。
The table schema features three fields: 'Trait N — Trait Name / Underlying Intent / Follow-up Question' (e.g., Trait 13 'Every paragraph ends with a punchline' → Underlying intent 'Memorable takeaway in every section' → Follow-up 'What is the single most important sentence in your article?'). It also notes: 'Tailor wording based on specific copy and hits during actual questioning; do not recite templates mechanically.'

False-positive mitigation is built into the criteria: thresholds and exceptions are defined for 5 traits (e.g., 'not X but Y' counts only if appearing 3+ times within 800 words; concluding punchlines only apply to long-form WeChat articles).

skills/dbs-ai-check/SKILL.md
鲁迅、李敖、罗翔都用。低密度是正常修辞
The false-positive table consists of three columns: 'Trait / Trigger Threshold / Explanation', covering #8 Not X but Y, #7 Speaking foolishness on behalf of the reader, #13 Concluding punchline, #1 Preempting counterarguments, and #5 Naming ritual. It also includes an explicit warning: '⚠️ False-positive warning: consecutive punchlines per paragraph are a stylistic requirement for short-video scripts, not AI tone. Clarify the genre with the user before detection.'

Genre identification comes first: applicable rules are adjusted across four categories (short-video scripts / long-form WeChat articles / posts / formal academic writing); when unstated, the model's inferred genre must be noted in the report.

skills/dbs-ai-check/SKILL.md
检测前先判断文案体裁，不同体裁的判定标准不同
Adjustments are mapped to specific rule numbers (#13 is disabled for short videos, prioritizing #16/#17; #14 is excluded for posts; #1/#2 are skipped for academic writing). Explicitly modeling whether 'the same phrasing counts as AI tone across different genres' significantly reduces false positives.

Self-check rules for follow-up questions: questioning copy must never exhibit the targeted AI traits (e.g., no multiple-choice formats guessing answers for users, no 'not X but Y' phrasing).

skills/dbs-ai-check/SKILL.md
不用选择题结构（A 还是 B？）→ 这是在替用户想答案
Comprises 5 self-check rules targeting the model's own phrasing. This rare 'meta-discipline' ensures that the tool's interaction style conforms to the very standards it advocates, preventing the methodology from contradicting itself.

2Core Capabilities

0122-trait sequential point-by-point detection report (original quotes + inline trait ID and severity + text-order traversal)
02Guided rewriting: sequentially queries intent by severity, one question at a time, providing directions rather than finished copy
03False-positive mitigation: trigger thresholds and exception rules for 5 traits
04Genre adaptation: specific disabled/prioritized criteria tailored across four distinct genres
05Special guardrails: direct counter-responses to requests like 'help me de-AI this' or 'rewrite this to look human'
06Negative constraint: acknowledges natural writing when no AI tone is found, never fabricating report hits

4Recommended Governance & Evidence

Suggested Controls for Safe Operation
For Agent Runtime
Can be auto-loaded
For Administrators
Register source and version
Evidence & Triggers (File Paths & Excerpts)
  • SKILL.md: Pure prompt directives and rule text, without script execution or external impact
Audit Notes: Self-verification: The dbs-ai-check directory contains only SKILL.md and agents/openai.yaml, with no scripts, references, templates, or data files. It involves no script execution, network endpoints, credential access, or local file writes; all capabilities are executed via the prompt criteria matrix and output templates in SKILL.md. The sole input is the text pasted by the user in the current session. It ends with a single soft suggestion ('When unsure about next steps, type `/dbs`') without triggering any tools. Boundary check (under the 'whether endpoints constitute necessary external calls' standard): contains no URLs or backend names, and zero CDN, font, or API references. The remaining risks reside purely at the prompt layer and output quality (see risks).

5Second-Pass Verification

  • [ok] 「22 个特征」数量属实 — 对 SKILL.md 逐条计数，`**特征 N —` 形式的条目恰为 22 条（特征 1 堵住所有反驳 … 特征 22 对「深刻」的过拟合），与 description 一致，无夸大。
  • [ok] 无脚本/无网络/无凭证/无写入 — 目录仅 SKILL.md + agents/openai.yaml；token 扫描（http|curl|wget|API_KEY|token|secret|\.env|bash|exec|python|~/\.）零命中；SKILL.md 无任何工具调用或读写文件指令。
  • [ok] 严重度三档确实写死 — 「🔴 强信号 — 几乎只有 AI 会这样做 / ⚠️ 中信号 — AI 高频但人也可能做 / 💡 弱信号 — 需要结合体裁和上下文判断」三行逐字存在。
  • [ok] 「默认不改」是否被机制保证 — 开篇「**默认只识别，不改。**」，报告模板末尾写「以上是检测结果，不涉及修改」，改写必须由用户说「我想改」触发；三条互相印证，非空口宣称。
  • [ok] 误伤警告与体裁调整不是装饰 — 误伤表 5 行含具体阈值；体裁表 4 行含具体条目号禁用（#13/#16/#17/#14/#1/#2）；特征 13 行内另有一条独立 ⚠️ 提示。
  • [ok] 与 marketplace 条目描述一致 — marketplace.json 描述「扫描文案中的 AI 生成痕迹，输出检测报告。默认只诊断不改。」与 SKILL.md description 及正文一致。
  • [ok] 元数据（license/stars/last_push/installs/commit） — GitHub API：stars 10121、pushed_at 2026-09-07T03:48:59Z、license.spdx_id NOASSERTION（根 LICENSE 为 CC BY-NC 4.0，非商业）；本地 HEAD = 8b8e33f1ecaed8cee606fe950c4426b525ead314 = 任务 pin。
  • [discrepancy] meta.installs 值域校正（批4 质检 2026-09-15） — 原记录为 skills.sh 仓库级聚合徽标（579.6K（skills.sh 徽标 /b/dontbesilent2025/…），并非本 skill 的单体安装量；站内惯例只记可定位的单 skill 数值 → 校正为 null（其他 25 份 dbs 报告本就为 null，修此 8 份以齐平）

6Verdict & Scenarios

  • Clear positioning that avoids paraphrasing traps: clarifies that 'AI flavor = too perfect' and anchors users to substantive content via 'time spent de-AI-ing is better spent clarifying thoughts.'
  • High report usability: quotes original text sequentially with inline trait numbers and severity levels, allowing authors to edit continuously along their manuscript.
  • Controlled false positives backed by quantitative thresholds and genre exceptions rather than intuitive 'looks like AI' tagging.
  • Non-intrusive rewriting: questions intent and suggests directions without ghostwriting, ensuring the revised copy remains authentically the author's.
  • Built-in meta-discipline: ensures interaction prompts strictly adhere to the evaluated stylistic standards.
    • Best for: Ideal for self-media creators, long-form article writers, and short-video scriptwriters conducting a 'robotic tone health check' before publishing—especially when readers note it 'sounds like AI' without pinpointing where, as the report isolates specific sentences. Also suited for authors wanting to refine drafts without losing agency, using structured questions to keep creative control in their hands. For developers designing detection skills with false-positive controls and meta-disciplines, it serves as a fully structured reference implementation (comprising trait matrix, threshold rules, genre mappings, and inquiry scripts).
    Not for: Not suitable for users seeking one-click paraphrasing or automated rewriting (requires diagnosis first, followed by sequential questioning, modifying copy only upon explicit instruction via an intentionally paced workflow); not suitable for scenarios requiring objective, verifiable metrics (contains no statistical benchmarks or model-based detectors, relying purely on heuristic criteria); not suitable for compliance audits of academic papers or legal filings (rules #1 and #2 are deliberately disabled for academic writing, disclaiming formal argumentative liability); and not suitable for processing confidential drafts when intra-session quoting of source text is prohibited.
    Install Agent-Ready
    ① Verified Mirror 2026-09-15
    Option A · Direct DownloadDownload dbs-ai-check.tar.gz
    sha256: 2ebc4b2111d7ab06…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit 8b8e33f1ec.