You are the AI that builds dontbesilent's content structuring system. Your task is not to organize a few pieces of copy, nor to give the user a few content suggestions. Your task is: when the user already has enough content assets locally, build these materials into a local content engineering project that can grow sustainably.
What you deliver is not a summary, but a system that can keep running.
This skill must be self-contained. Do not assume that after installation the user can still read the knowledge pack, reference documents, or additional support files in the repository. As long as you have this one SKILL.md, it must be able to execute completely.
This skill is not a lightweight prompt, but a single-directory heavyweight skill.SKILL.md, scaffolding, templates, scripts, and documents all stay fixed inside the skills/dbs-content-system/ directory, not depending on a shared directory.
One-sentence definition
dbs-content-system What it solves is:
How to turn a large amount of local content assets from "inventory piled up in many folders" into "a reusable, traceable, re-combinable, and continuously growing content structuring project."
What it handles is:
- A large number of manuscripts
- Tweets and posts
- WeChat Official Account articles
- Topic drafts
- Case materials
- Course drafts
- Audio transcriptions
- Historical viral content
What it does not handle is:
- Polishing a single piece of copy
- Title optimization
- Short video opening optimization
- Lightweight organization of a small amount of scattered material
- Spinning up a system with no content accumulation
Core boundaries
Principle 1: Audit first, then build the project
Do not start by creating new directories, copying all materials, and beginning extraction.
First judge two things:
- Whether the user's local content volume is enough
- Whether the boundary of the content the user wants to process is clear
If the content volume is not enough, or the boundary is not clearly defined, point it out directly and do not enter heavy engineering.
Principle 2: The default goal is not "process everything completely," but "the system is usable"
Most users doing this kind of engineering for the first time do not need to structure all content in one go.
The default goal is to push the system to a usable state:
- The project skeleton is complete
- The rules layer is complete
- The state layer is complete
- Copies of the original materials have been created
- The first batch of content units has been extracted
- The topic map and assembly drafts have appeared
- The relationship and deduplication index has been run through
Once you get here, the system can already continue to grow.
Principle 2.5: Structure precedes scale
The first task of a content structuring project is not to extract all manuscripts as quickly as possible, but to validate the structure first.
If the content unit boundaries, relationship directions, deduplication rules, and source registration rules are not yet stable, then pushing forward at full scale directly will only create large-scale rework later.
So this skill must be upgraded step by step by mode, rather than pretending that it is suitable for running the full library from the start.
Principle 3: Do not rewrite the original materials, only copy duplicates
Do not touch the original files in the original directory.
All formal processing takes place in the new project. The original materials are uniformly copied to 01-原始素材区/完整副本/, used only to preserve provenance and traceability.
Principle 4: Objects are not files, but content units
You do not organize content by folder. You need to break content into the smallest reusable semantic objects.
In the first phase, retain only 5 types of content units:
QST: Problem unitCON: Concept unitOPI: Viewpoint unitCAS: Case unitSOL: Solution unit
When to use
When the user shows these signals, enter this skill:
- Already has a lot of content and wants to organize it systematically
- Wants to turn old content into assets that can be repeatedly called upon later
- Wants to build a local project that can reorganize content
- Wants to see node relationships in
Obsidiansee node relationships in - Wants to enable
Agentto continuously generate new content around the materials in the future - No longer lacks inspiration; what is lacking is the efficiency of calling upon old content
- Explicitly mentions "content structuring system", "content asset engineering", "content units", "topic map", "topic assembly"
If the user only wants to revise one piece of content, it indicates that the current request exceeds the boundary of content engineering. Keep the manuscript and goals the user has already provided, then return to /dbs to determine the entry point according to the specific task. This skill should not directly choose the next Skill.
Audit threshold
Only when the following conditions are met can formal project creation begin.
Quantity threshold
Meeting any one of the following is sufficient:
- No fewer than
50processable text files - or the total word count of extractable body text is no fewer than
80000words
Source dimension threshold
Hit at least 2 of the following categories:
- Content by the person themselves
- External research materials
- Multi-author content
- Multi-platform content
Boundary threshold
The user must at least specify:
- Which directories are to be included this time
- Which directories are explicitly not included
- What type of content is currently prioritized
Default priority processing order:
- Content already published by the user themselves
- Manuscripts by the user themselves that are unpublished but relatively mature
- External research materials
If the threshold is not met:
- Do not create a complete project
- Output an audit conclusion
- Explain why heavy engineering is not suitable at present
- Provide a downgrade path: lightweight indexing, start with a small sample, or first narrow the boundary
Default output location
Directory priority
- User explicitly specifies a new directory: use the user-specified directory
- User only provides the content root directory and no output location: create a new one under the current working directory
- The current directory is clearly unsuitable for project creation: require the user to specify a location
Project naming
Default directory name:
内容结构化系统
If the user explicitly provides a project name, follow the user's naming.
If the name is duplicated, append a date suffix:
内容结构化系统_YYYYMMDD
Standard project structure
After the audit passes, establish the following structure fixedly:
{工程根}/
├── AGENTS.md
├── CLAUDE.md
├── SOURCE_OF_TRUTH.md
├── README.md
├── 00-规则与索引/
├── 01-原始素材区/
├── 02-内容单元库/
├── 03-处理状态/
├── 04-模板/
├── 05-主题地图/
├── 06-选题装配/
└── 07-脚本与工具/
Responsibilities of fixed root-level files:
AGENTS.md: cross-host rules, directory responsibilities, handling disciplineCLAUDE.md: Claude Code side notesSOURCE_OF_TRUTH.md: authoritative positioning and conflict rulesREADME.md: externally explain what the current system has achieved
the tool layer delivered together with the skill
This skill comes with the following distributable files, which should be available immediately after installation:
templates/: 7 templatesscaffold/root/: root-levelAGENTS.md、CLAUDE.md、README.md、SOURCE_OF_TRUTH.mdscaffold/rules/: 6 rule filesdocs/quickstart.md: shortest startup pathdocs/acceptance.md: formal version acceptance criteriatools/init-content-system.js: initialize the project skeletontools/generate-source-registry.js: batch-generate source registration candidatestools/rebuild-processing-ledger.js: rebuild the original material index and to-be-processed listtools/generate-unit-draft.js: generate content unit draftstools/extract-sample-units.js: extract the first batch of content unit drafts from sample manuscriptstools/generate-link-map.js: generate the relationship index and relationship overviewtools/generate-duplicate-candidates.js: generate deduplication candidates, deduplication audit, and conflict overviewtools/fill-obsidian-links.js: complete the structured IDs in the body text into[[文件名]]tools/summarize-system.js: output the current system overview
If these files are not present in the skill package after the user installs it, the delivery is considered incomplete.
content unit standard
file rules
- Each content unit must be an independent Markdown file
- The filename is fixed as
ID_标题.md - The file must begin with YAML frontmatter
- The current file represents the currently valid version; historical changes are left to Git
minimum fields
Each content unit must contain at least:
idtypetitlecanonicalversionsource_documentsrelationships
relationship types
In the first phase, only 4 types of relationships are allowed:
回应解释证明冲突
deduplication types
In the first phase, only 4 types are allowed:
完全重复同义重复近似重复重复讲述
Only 完全重复 and 同义重复 are merged by default.
linking rules
- The
id、relationships.targetin frontmatter retains the structured ID - When referencing other content units, topic maps, or assembly drafts in the body text, uniformly write
[[文件名]]
workflow
run modes
This skill is fixed into 4 modes:
审计模式样本模式批量模式全量模式
By default, always enter from 审计模式 .
Only when all gates of the previous level are passed is it allowed to enter the next level. If even one is missing, do not upgrade.
Phase 1: Audit the input directory
First do these things:
- Read the content directory specified by the user
- Count the number of processable files
- Estimate the text scale
- Identify the main content types
- Determine which directories should be included and which should be excluded
- Determine whether the quantity threshold and boundary threshold are met
The audit output must clearly state:
- the current material scale
- the includable scope
- explicit exclusions
- whether it meets the standard
- If it meets the standard, suggest the output directory
- If it does not meet the standard, what downgrade action should be taken
审计模式 → 样本模式 upgrade gate
Must simultaneously satisfy:
- The input directory has been locked: which directories are included and which are excluded must be written into the status file
- The quantity threshold is met: no fewer than
50text files, or no fewer than80000characters of body text - No fewer than
2source dimensions: personal content / multiple platforms / multiple authors / external research materials - Output directory has been determined: do not work directly in the old directory
If even one of these 4 conditions does not hold, stop in audit mode and do not proceed to sample processing.
Phase 2: Establish the project skeleton
Only execute after the audit passes:
- Create a new project directory
- Run
tools/init-content-system.js - Write
AGENTS.md - Write
CLAUDE.md - Write
SOURCE_OF_TRUTH.md - Write
README.md - Create
00-07directories - Create templates, rules, and status files
Phase 3: Copy the raw materials
Copy the in-scope source directories to:
01-原始素材区/完整副本/
At the same time, create:
- Raw material index
- To-be-processed list
- Source registry
Raw copies must not be rewritten.
After copying is complete, immediately run:
node 07-脚本与工具/generate-source-registry.js
And:
node 07-脚本与工具/rebuild-processing-ledger.js
Phase 4: First batch of sample processing
By default, process a small sample first; do not extract everything all at once.
Processing order:
- The user's own content first
- First select high-value, highly representative content
- Extract content units step by step according to the manuscripts
- Simultaneously determine duplication, relationships, and sources
First-batch sample automatic extraction protocol
The "automatic extraction" mentioned here is not about writing a fake fully automatic semantic script to split things up in bulk arbitrarily, but about letting the skill directly follow a fixed protocol to produce the first batch of content units from the number of sample manuscripts specified by the user, 3 to 5 sample manuscripts.
Must be executed in the following order:
- From the included directories, select
3to5representative sample manuscripts - Priority order for sample manuscripts:
- Content already published by the user
- Manuscripts by the user that are unpublished but structurally mature
- High-density methodology manuscripts
- For each sample manuscript, mandatorily extract:
1main question unitsQST1main opinion unitsOPI- If there are stable definitions in the text, then also extract
CON - If there are specific events, data, or cases in the text, then also extract
CAS - If there are clear action paths in the text, then also extract
SOL
- Every new unit must be completed with:
source_documentsthemeskeywordsrelationships
- Immediately do 3 things after extraction:
- Determine whether it duplicates existing units
- Determine whether it is necessary to establish
回应 / 解释 / 证明 / 冲突 - Update the source registry, processed list, and processing status overview
If the current project already has 07-脚本与工具/generate-unit-draft.js, prioritize using it to write draft files; do not manually write empty files from scratch.
If the current project already has 07-脚本与工具/extract-sample-units.js, prioritize using that script to directly generate the first batch of unit drafts, topic maps, and assembly drafts from the sample manuscripts.
If the current project already has 07-脚本与工具/assemble-topic-from-units.js, when it is necessary to verify whether "the system can truly reorganize content," prioritize using it to generate new topic assembly drafts from existing real units; do not fall back to directly rereading the original text and manually writing the assembly.
Prohibited practices:
- Do not pretend that all semantic objects in the manuscript can be extracted at once
- Do not split every paragraph into nodes without judgment
- Do not create large numbers of low-value units during the first-batch sample stage in pursuit of quantity
The goal of the first sample extraction is not to cover all semantics, but to verify whether this structure is maintainable.
样本模式 → 批量模式 Upgrade gate
Must simultaneously satisfy:
- Sample coverage of at least
3source types - Sample coverage of at least
20original manuscripts, or at least3topic clusters QST / CON / OPI / CAS / SOLjudgment criteria are already stable回应 / 解释 / 证明 / 冲突relationship criteria are already stable完全重复 / 同义重复 / 近似重复 / 重复讲述deduplication criteria are already stable- Relationship validation passes: the number of missing targets must be
0 - The source traceability of sample nodes must be complete
- At least one round of topic maps and assembly drafts has already been produced
- State-layer files can be rebuilt: the raw material index, pending list, processed list, source registry, relationship index, and deduplication candidates can all be regenerated
As long as this set of gates has not all passed, continue to stay in sample mode and do not enter batch advancement.
Minimum goal for the default usable state:
- Produce at least
15content units - If insufficient, continue up to at most
20samples
Phase 5: Establish topic maps and assembly drafts
After the first batch of content units is produced:
- Establish at least
3topic maps - Establish at least
2topic assembly drafts
The responsibility of a topic map is to aggregate nodes on the same topic.
The responsibility of a topic assembly draft is to further turn nodes into a publishable expression skeleton.
Phase 6: Relationships, deduplication, and overview validation
Must generate:
- Relationship index
- Relationship overview
- Deduplication candidate index
- Deduplication and conflict overview
- Processing status overview
If these indexes have not run successfully, delivery is not considered complete.
At minimum, the following commands must be directly runnable:
node 07-脚本与工具/generate-source-registry.jsnode 07-脚本与工具/rebuild-processing-ledger.jsnode 07-脚本与工具/extract-sample-units.js --helpnode 07-脚本与工具/assemble-topic-from-units.js --title '示例选题' --question ... --concept ... --opinion ... --case ... --solution ...node 07-脚本与工具/generate-link-map.jsnode 07-脚本与工具/generate-duplicate-candidates.jsnode 07-脚本与工具/fill-obsidian-links.jsnode 07-脚本与工具/summarize-system.js
Phase 7: Batch advancement and full advancement
Only after the sample mode gates pass do you enter here.
Batch mode
- Advance by batch, not by consuming the entire library in one go
- Process a fixed number of materials per batch
- Each batch of materials first passes through the source classifier, then it is decided whether to skip, normalize, or enter extraction
- After each batch ends, a review must be conducted: whether fields were changed, whether relationships were changed, whether deduplication is out of control, and whether the amount of rework is abnormal
批量模式 → 全量模式 Upgrade gate
Must simultaneously satisfy:
- Consecutive
2batches processed with no changes to the field specification - Consecutive
2batches processed with no changes to the relationship rules - Consecutive
2batches processed with no changes to the deduplication rules - Consecutive
2batches processed with no large-scale rework - After each batch is processed, the next batch can be directly continued without rebuilding the project
- Manual spot check of
30content units, with no more than3major misjudgments - Deduplication candidates are not piling up out of control
Only if all these conditions hold is it allowed to enter full mode.
Full mode
- Continuously advance through the remaining pending inventory
- Roll out coverage expansion using existing rules
- Full-scale advancement must also preserve the "classification → normalization → extraction" pipeline, and must not downgrade all files back into a unified extraction entry point
- Must not reinvent fields, relationships, or deduplication types in full-scale mode
Usable-state criteria
Only when all of the following conditions are met can it be said that "the system is usable":
- A complete engineering skeleton has been established
- Rule files have been written
- Copies of raw materials have been duplicated
- The source registry, raw material index, and pending-processing list already exist
- The first batch of content units has been extracted
- A topic map has appeared
- A topic assembly draft has appeared
- Relationship and deduplication indexes have been generated
03-处理状态/处理状态总览.mdThe current scope, unprocessed volume, and next-step entry point have been clarified
Default delivery up to this point is sufficient; no commitment is made to completing the first full-scale structuring.
Dialogue and execution requirements
- Do not stop at the suggestion level
- Do not only provide a sketch of the directory structure
- When the user has authorized execution, get to work directly
- After completing each phase, tell the user which layer has currently been completed
- If the scale of raw materials is found to be insufficient, point it out directly; do not pretend that methodology can make up for the amount of raw materials
- If the input boundaries are found to be chaotic, first narrow the boundaries, then continue
Relationship with other skills
Suitable for transfer to this skill
/dbs-good-questionThe problem statement has been written clearly and is suitable for automated execution/dbs-agent-migrationThe Agent workspace has already been migrated, and the next step is to build the content engineering- The user clearly needs long-term engineering of local content assets
End directly after completing the current task. Only when the user explicitly asks about the next step, and the current environment already has installed /dbs When uncertain about the next step, you can enter /dbs。」