Content & Media Creation · dontbesilent2025/dbskill

dbs-learning

Name: dbs-learning
Author: dontbesilent2025/dbskill

Breaks down a topic into a continuous series of learning articles, adjusting the depth, angle, and pace of subsequent posts based on user feedback. Used when users request systematic study of a topic, want to continue to the next article, or need to adjust the course based on feedback.

L2 Local · Restricted workspaceV1 Commit-PinnedWrite FileDual-pass Audit

Author dontbesilent2025GitHub dontbesilent2025/dbskill ↗Stars 10121License CC BY-NC 4.0（仓库根 LICENSE：Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)；署名 + 禁止商业使用；Copyright (c) 2026 dontbesilent）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

A pure prompt-orchestrated 'adaptive course engine': all states persist in local fixed files (00-学习计划.md + two-digit numbered articles + assets/ + optional INDEX.md), allowing the model to resume each round by reading these files back.

skills/dbs-learning/SKILL.md

你维护的是一个自适应学习梯度。

Directory priority: user-specified directory > current project directory /学习课题/{topic_name}/ > fallback ~/Documents/dbskill-learning/{topic_name}/.

The key to the closed loop lies in the 'feedback extraction rules': 4 prompt questions are embedded at the end of each article, and when generating the next post, template lines must be filtered out so only user-added text is treated as feedback.

skills/dbs-learning/SKILL.md

只有用户在提示问题下面新增的文字，才算真实反馈。

Accompanied by a rule: 'If there is no content after filtering template lines, treat it as no feedback'—preventing the model from mistaking templates for user responses or hallucinating feedback.

Maps feedback signals to six adjustment strategies for difficulty/perspective (table-driven), achieving 'adaptive learning' rather than a rigid syllabus.

skills/dbs-learning/SKILL.md

| 没看懂、概念混乱、问题很多 | 降低抽象度，补例子，放慢节奏 |

The table also covers branches like 'Understood but found it boring → shift angle of approach' and 'Clearly mastered → increase concept density'.

Knowledge baseline confirmation is a dedicated phase (Phase 1.5) that only activates for new topics; it conducts scoped lightweight retrieval on the current project first, then asks the user to choose 'project-based / from scratch / specified materials'.

skills/dbs-learning/SKILL.md

本阶段只用于新课题。

Retrieval scope is strictly limited: prioritize reading AGENTS.md/CLAUDE.md/README.md/SOURCE_OF_TRUTH.md and directory indexes; 'do not traverse or read through the entire project'.

Evidence hierarchy is built into the workflow: formal content written/confirmed by the user counts as strong evidence, AI drafts count as weak evidence, and third-party materials do not count toward user knowledge baseline.

skills/dbs-learning/SKILL.md

AI 生成的草稿、分析报告和未确认内容：只能作为弱证据。

Includes two granular inference rules: 'A concept appearing in a file only proves the user might have encountered it', and understanding can only be inferred if 'invoked repeatedly across multiple independent scenarios'.

Retrieval incorporates strict sensitive exclusion and project-rules-first precedence: skips .git/.env/.private/keys/credentials/PII/financial records/client confidential files by default, giving priority to prohibited reading scopes defined in project rules.

skills/dbs-learning/SKILL.md

默认跳过 `.git/`、`.env`、`.private/`、密钥、密码、凭证、身份信息、财务原件、客户私密资料和其他明确标记为敏感的内容。

Also specifies that sensitive content 'even if relevant to the topic, can only be read after explicit user authorization'.

Writing style is governed by a 'prohibited sentence patterns list', explicitly banning 6 categories including 'not... but...' and their variants, while enforcing Chinese full-width punctuation and spaces between Chinese and English text.

skills/dbs-learning/SKILL.md

默认禁止使用以下句式及其近似变体：

Alternatives are provided: stating conclusions directly/causal clauses/conditional clauses/action clauses/concrete examples; exceptions apply only when imitating original text upon request, quoting original text, or analyzing sentence structures.

File naming and sequence progression follow strict mechanical rules: learning plan fixed at 00-学习计划.md, articles use two-digit numbers, next article = current max sequence + 1, skipping numbers and Chinese filenames are prohibited.

skills/dbs-learning/SKILL.md

下一篇序号 = 当前课题目录中最大文章序号 + 1

Article body follows a fixed structure: 'Problem to solve in this piece / Body / Summary / Next preview / Learning feedback', with the feedback section serving as input for the next round.

2Core Capabilities

01Topic initialization: create topic directory, 00-学习计划.md, 01.md, assets/, and index records

02Knowledge baseline confirmation (project retrieval + 3-choice prompt), written into the learning plan once confirmed and inherited without re-prompting each time

03Progress continuation: read learning plan → locate highest numbered article → extract actual feedback → summarize current comprehension state in 3-5 points

04Adaptive adjustment: alter abstraction level, angle, and information density of the next article based on feedback signals

05Article generation: fixed 4-part structure + end-of-article learning feedback section (including 4 prompt questions and input slots)

06Plan and index maintenance: update current progress/current topic/prior feedback summary/next article direction/update timestamp, and sync the INDEX.md table

07Safe fallback disk writes: avoids polluting generic/system directories by redirecting writes to ~/Documents/dbskill-learning/ and notifying the user explicitly

08Style discipline: bans patterns like 'not... but...', enforces Chinese typesetting rules, and never outputs body text in English

4Recommended Governance & Evidence

Suggested Controls for Safe Operation

For Agent Runtime

Restrict working directory

For Administrators

Retain operation logs

Evidence & Triggers (File Paths & Excerpts)

本地工作目录: File I/O and artifact output: {topic_directory}/00-学习计划.md, {seq}.md (starting from 01.md, two-digit sequence), assets/

Audit Notes: Evaluation target is this skill's own directory: contains no scripts, no network endpoints, and no credential reading. However, SKILL.md explicitly requires the agent to create and continuously write local files (topic directory, 00-学习计划.md, serialized articles, assets/, INDEX.md), classifying it as 'local read/write only, no egress'. Privacy note: it needs to read project files to determine the knowledge baseline—while it excludes .git/.env/.private/keys/credentials/PII/finances/client confidential data by default and defers to project rules, the actual read scope ultimately depends on host permissions and model compliance.

5Second-Pass Verification

[ok] 外部资源/工具：是否存在未声明的依赖 — 目录内仅 SKILL.md + agents/openai.yaml；无 URL、无包、无 CLI。external_deps 空数组成立。
[ok] 安全结论：敏感文件排除是否为可执行规则 — 排除项以编号 4 的形式写进「任何项目检索都必须遵循以下范围」，并给出 5（敏感内容需明确指定）、6（项目规则优先）两条兜底；验收用例 4 要求「不在知识基础摘要中暴露敏感文件名、内容或路径」。规则可判定、有对应场景，非泛泛声明。
[ok] 实现原理是否成立：反馈驱动是否真能改变后续深度 — Phase 3 给出六行映射表（反馈信号→处理方式），Phase 2 要求先读上一篇文章末尾反馈并概括为 3-5 条理解状态，Phase 5 要求把「用户上一轮反馈摘要」写回学习计划——形成读反馈→判断→写文→回填计划的闭环。用例 2/3 给出了两种典型反馈下的期望输出，可验证。
[ok] 功能声明 vs 实际能力：「系统学习一个主题」是否被夸大 — skill 不承诺教学效果、不做外部资料检索（不抓网页、不引第三方教材），只做「连续文章编排 + 反馈调整」；这与 description 一致，无越界承诺。
[unlocatable] 证据定位：仓库级元数据（stars/last_push）与 installs — stars=10121、last_push=2026-09-07 取自站点池 data/pool.json（GitHub API 采集），pin 源码内无法逐字定位；installs 未采集记 null。license 可在仓库根 LICENSE 逐字命中。

6Verdict & Scenarios

Turns 'course serialization' into a state machine: plan file + numbered articles + feedback section, resumable across sessions while inheriting confirmed knowledge baselines without repetitive questions

Explicit anti-hallucination feedback extraction (recognizes only newly added text below prompt questions), grounding 'adaptivity' in reality rather than model speculation

Prudent use of local materials for baseline confirmation: categorizes evidence strength, differentiates 'encountered' from 'mastered', excludes sensitive files by default, and prioritizes project rules

Verifiable writing style: prohibited patterns list + alternatives + three exceptions, avoiding synthetic 'AI-sounding' instructional prose

Testing-oriented design: built-in 6 acceptance test cases, turning critical branches like new topics, sensitive materials, and feedback types into verifiable scenarios

Best for: Suitable for users who want to systematically develop a topic into a serialized learning resource: self-learning a subject, breaking down methodologies into courses, or creating training series for teams—especially when needing difficulty adjustments based on understanding. Works best paired with dbs-knowledge to index topic directories into knowledge base navigation.

Not for: Not suitable for ghostwriting one-off standalone content (the skill explicitly states 'responsible for curriculum sequences, not ghostwriting single articles'); not suitable for business diagnosis or operational execution (delegate to corresponding dbs-* skills); not suitable for research learning requiring external web retrieval or citations (does not scrape or cite external sources, only orchestrates sequences); not suitable when creating new files in the working directory is disallowed.

Install Agent-Ready

① Verified Mirror 2026-09-15

Option A · Direct DownloadDownload dbs-learning.tar.gz

sha256: 1f686116522fc75b…

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.