ai-content-kb

Imported from pandatalk8/ai-content-kb (AGENTS.md). Install upstream with npx skills add pandatalk8/ai-content-kb. Copyright stays with the author.

Agent Operating Rules

This repository is an AI-assisted personal content knowledge system. Follow CLAUDE.md and KNOWLEDGE_BASE_GUIDE.md.

If this is a new session, read START_HERE.md first.

Natural-Language Workflows

Treat the following phrases as repository workflow intents. They are natural-language shortcuts, not application slash commands. Accept close Chinese or English variants.

加入知识库 / add to knowledge base

When the user provides or references material:

1. Inspect the complete material and determine provenance.
2. Put owner-authored input under raw/; put external material under sources/.
3. Preserve the original meaning and record a source URL or origin note when available.
4. Do not copy non-redistributable full text into a public repository; store permitted metadata and an original summary instead.
5. Check existing aliases and related pages.
6. Create candidate wiki pages under .kb/staging/wiki/ and candidate sidecars under .kb/staging/links/ when useful.
7. Report created files, provenance decisions, uncertainties, and review items.

If no material, path, attachment, or URL is available, ask the user what should be added.

增加 Wiki 索引 / add wiki index

For the specified material or topic:

1. Read the relevant originals, not only existing wiki summaries.
2. Reuse canonical concepts and aliases where possible.
3. Create or update candidate concept, entity, map, or source-note pages under .kb/staging/wiki/.
4. Include repository-relative citations to raw/, sources/, or products/.
5. Create matching candidate graph sidecars under .kb/staging/links/.
6. Do not overwrite reviewed wiki pages or promote candidates automatically.

审核并发布索引 / review and publish index

Review the requested staged wiki pages and sidecars. Validate citations, aliases, paths, edge semantics, and source hashes. Promote accepted files to wiki/ and .kb/links/, then update wiki/changelog.md. Report anything left in staging.

查询知识库 / query knowledge base

Follow the Query workflow below. Answer with repository-relative evidence paths and distinguish owner judgment, external claims, published expression, and inference. Do not modify files unless asked.

在知识库中进行 AI 内容创作 / compose from knowledge base

For an article, course, newsletter, social post, or video script:

1. Establish the topic, audience, format, objective, tone, and important constraints from the request.
2. Search reviewed wiki pages and graph sidecars, then read the relevant originals under raw/, sources/, and products/.
3. Separate owner judgments, external claims, reusable published expression, and new inference.
4. Produce a source plan or outline before long-form drafting when the task is substantial.
5. Write unreviewed prose to .kb/staging/drafts/; put course outlines or control files in .kb/staging/course-drafts/. Include a Sources section with repository-relative paths.
6. Mark unsupported claims, evidence gaps, and questions for human review. Do not invent citations.
7. Do not publish to products/ automatically. After the owner accepts the draft, move it to raw/drafts/ for further development or to products/ only when publication-ready.
8. After publication, offer to backfill reusable concepts and relationships into staging.

回填知识库 / backfill knowledge base

Inspect the requested existing content using the Backfill workflow. Generate candidates and reports in .kb/staging/ or .kb/reports/; do not rewrite original or product bodies.

检查知识库 / lint knowledge base

Run deterministic checks described in Lint. Write a report under .kb/reports/ when the result is substantial. Do not apply semantic corrections automatically.

迁移旧知识库 / migrate legacy knowledge base

Follow docs/MIGRATION_GUIDE.md and its safety model:

1. Begin with a read-only inventory. Do not move, copy, delete, rename, or rewrite legacy files during inventory.
2. Generate an inventory report under .kb/reports/ and a proposed mapping under .kb/staging/migration/.
3. Classify files by provenance, review state, and intended use before topic. Mark ambiguity as needs_review.
4. Wait for mapping approval before copying content.
5. Pilot with 20–50 representative files. Copy rather than move, and never use destructive synchronization.
6. Keep generated wiki pages and graph data in staging.
7. Validate counts, hashes or bodies, frontmatter, links, attachments, aliases, privacy, and copyright risk.
8. Expand only in reviewed batches with one coherent Git commit per accepted batch.
9. Keep the old vault read-only until cutover and recovery checks pass.

Directory Semantics

• raw/: original personal input from the repository owner.
• sources/: external, citable source material.
• products/: reviewed, published, or delivered personal output.
• wiki/: human-readable knowledge index and synthesis layer.
• .kb/links/: reviewed, machine-readable relationship graph.
• .kb/staging/: AI-generated output awaiting human review.

Source Of Truth

• raw/, sources/, and products/ are authoritative in different roles.
• wiki/ is an index and synthesis layer, not independent proof.
• Non-trivial claims in wiki/ must cite raw/, sources/, or products/.
• Unreviewed AI output is never a source of truth.

Write Boundaries

• Preserve original files under raw/ and sources/ unless explicitly asked to edit them.
• Do not modify mounted or externally managed product files unless explicitly requested.
• Do not batch-insert Obsidian wikilinks or vault-specific metadata into product files.
• Put AI-generated bulk changes in .kb/staging/ first.
• Preserve human-reviewed wiki content.
• Do not store binary audio, video, archives, credentials, or secrets in the vault.

AI-Generated Content

• Put unreviewed article drafts in .kb/staging/drafts/.
• Put unreviewed course outlines or COURSE.md drafts in .kb/staging/course-drafts/.
• Put unreviewed wiki page candidates in .kb/staging/wiki/.
• Put draft relationship sidecars in .kb/staging/links/.
• Move a draft to raw/drafts/ only after the owner chooses to develop it.
• Move content into products/ only after human review and acceptance.

Graph Model

Represent important relationships with sidecar YAML under .kb/links/. Use the schema in docs/GRAPH_SCHEMA.md.

Preferred edge types:

• mentions
• explains
• uses
• derived_from
• part_of
• related_to
• alias_of

Do not use natural-language summaries as the only relationship record.

Ingest

1. Read the complete source file.
2. Extract its title, summary, concepts, entities, claims, and examples.
3. Check existing concept aliases before creating pages.
4. Create draft pages and sidecars under .kb/staging/.
5. Promote reviewed output to wiki/ or .kb/links/.
6. Record reviewed changes in wiki/changelog.md.

Backfill

• Sample structure first: title, frontmatter, headings, opening, section leads, ending, links, and image alt text.
• Read the full file when needed to verify a relationship or claim.
• Do not modify original article or course bodies during backfill.
• Keep candidate relationships in staging until reviewed.

Query

1. Start with wiki/ for orientation.
2. Use .kb/links/ to locate connected raw, source, and product files.
3. Return to those original files for factual detail.
4. Cite repository-relative paths.

Lint

Check for:

• broken source and target paths;
• missing citations;
• duplicate aliases;
• deprecated pages still referenced;
• source hash changes that require re-ingest;
• missing course or media control files;
• secrets, personal absolute paths, and binary files.

Review Rule

The deciding factors are:

review status
intended use
human ownership
publication readiness