菜单
Writes reference cards for projects, concepts, or people in the wiki. Projects get 800-1500 words with architecture, benchmarks, limitations, and alternatives. Concepts get 1500-2500 words. People get 150-250 words. Use when writing wiki reference cards for specific entities.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Show pending changes and LLM-powered impact analysis without compiling. Detects new/modified/deleted sources via content hashing, maps them to dirty buckets, reads changed sources to assess impact and detect contradictions with existing claims. Use to preview what incremental compilation would do before running it.
Incrementally recompiles changed parts of the wiki after new sources are ingested. The skill handles change detection, LLM-powered impact analysis, and user confirmation. All compilation is delegated to the script pipeline (bun run compile --incremental), which handles entity extraction, resolution with anchoring, synthesis with evidence registry, reference cards, claims, self-eval, and state saving.
Extracts atomic verifiable claims from synthesis articles and runs self-evaluation by verifying a sample against raw sources. Outputs build/claims.json and build/eval-report.json. Use when extracting claims, running self-eval, or verifying wiki accuracy.
Writes field-map.md, the flagship overview article that connects all taxonomy areas into one system. A 3000-5000 word systems map showing integration points, paradigm routing logic, and a concrete practitioner flow. Use when writing the field map or systems overview.
Compiles the wiki from raw markdown sources. Read `config/domain.ts` for the domain topic and taxonomy. Orchestrates a multi-phase pipeline using subagents: read sources, write synthesis articles, write reference cards, generate field map and indexes, extract claims, and run self-eval. Use when asked to compile, build, or generate the wiki from raw sources.
Deep-researches a project or paper for the knowledge base. Clones repos to read source code, fetches full paper text, and produces structured analysis files in raw/deep/. Use when asked to deep-research, deep-dive, analyze source code, or enrich a source with implementation details.
| name | compile-cards |
| description | Writes reference cards for projects, concepts, or people in the wiki. Projects get 800-1500 words with architecture, benchmarks, limitations, and alternatives. Concepts get 1500-2500 words. People get 150-250 words. Use when writing wiki reference cards for specific entities. |
Read config/domain.ts for domain context and taxonomy bucket definitions.
Write reference cards for projects, concepts, or people.
| Type | Words | Output path |
|---|---|---|
| Project | 800-1500 | wiki/projects/<slug>.md |
| Concept | 1500-2500 | wiki/concepts/<slug>.md |
| Person | 150-250 | wiki/people/<slug>.md |
raw/deep/repos/ and raw/deep/papers/Every card starts with <abstract></abstract> tags before any headings. One sentence, under 200 chars. State what it does plus the key differentiator. Not a generic description.
What It Does -- Concrete description. Not marketing copy. What problem it solves and the architectural approach that makes it different.
Architecture -- How it works. Name specific files, modules, algorithms from deep sources. Enough detail that a reader understands HOW, not just WHAT.
Core Mechanism -- The key algorithm or pattern with implementation detail from deep sources.
Key Numbers -- Benchmarks with credibility assessment: self-reported, peer-reviewed, or verified in code. If README claims cannot be verified, say so.
Strengths -- Each backed by evidence from a deep source.
Critical Limitations -- At minimum:
When NOT to Use It -- Operational conditions where this is the wrong choice.
Unresolved Questions -- 3-5 specific unknowns: governance, cost at scale, conflict resolution, architectural growth limits. What the documentation does not explain.
Alternatives -- Link to other project cards with guidance on when to pick each.
Definition -- What it is. Precise, no jargon layering.
Why It Matters -- The specific problem it solves for practitioners.
How It Works -- Implementation details from deep sources. Name projects that implement it and how their approaches differ.
Who Implements It -- Link to project cards with each project's specific approach noted.
Failure Modes -- When does this concept break down? Concrete scenarios.
Open Questions -- What remains unresolved.
Brief profile: who they are, key contributions to the field, notable work. Link to their projects and publications in the wiki.
[Source](../raw/type/file.md) or [Source](../raw/deep/type/file.md)[Mem0](../projects/mem0.md), [context graphs](../concepts/context-graphs.md)---
entity_id: <slug>
type: project | concept | person
bucket: <bucket-id>
abstract: "<1 sentence, under 200 chars>"
sources:
- raw/type/file.md
- raw/deep/type/file.md
related:
- projects/slug.md
- concepts/slug.md
last_compiled: "YYYY-MM-DD"
---
Read and follow all rules in stop-slop/SKILL.md. No filler, no passive voice, no em dashes, no AI tells.
See references/template.md for the full structural templates for each card type.