Imported from tingaicompass/AI-Compass (SKILL.md). Install upstream with npx skills add tingaicompass/AI-Compass. Copyright stays with the author.

AI-Compass Local Knowledge Base

Treat the directory containing this SKILL.md as the repository root. Use AI-Compass as a curated Chinese knowledge base for AI learning, model and tool discovery, technical architecture, engineering examples, and weekly industry developments.

Knowledge Model

Use the repository as three connected layers instead of one undifferentiated document collection:

Never treat a weekly item as a durable consensus without checking its matching topic page. Conversely, do not present a long-term topic page as evidence of the latest news without reading the weekly index and issue.

Decide the Entry Point

Classify the request before opening files:

If a request combines time-sensitive and foundational questions, answer in two labeled parts: 本周增量 and 长期沉淀.

Topic Routing

Search the most likely topic home before using a repository-wide search. These are the preferred entry files:

Use rg -n to locate an unfamiliar term after choosing the most likely subtree. Do not load all topic files or every weekly issue just because the repository is local.

Weekly Item Routes

When a question targets a specific weekly item, check weeklyHighlights/topicRoutes/<issue>.json after identifying the issue. Schema version 1 maps each weekly section_id and title to:

• the item's anchor in the weekly issue;
• one primary topic path and its stable body anchor;
• zero to two related topic paths.

Use the primary topic as the durable full-text entry. Related topic paths are navigation only: follow their directory link to the primary entry instead of expecting or synthesizing a duplicate body. Treat the route file as a retrieval index, not as evidence for the underlying claim; cite the weekly issue or primary topic page that contains the text.

Route files are optional so older issues and older repository versions remain usable. If the file is absent, malformed, uses an unsupported schema version, or has no matching item, fall back to latest.md or INDEX.md, open the relevant issue, and use the preferred topic table plus a scoped rg -n search.

Retrieval Workflow

1. Identify whether the request needs current information, historical comparison, durable explanation, implementation detail, or a learning route.
2. For a specific weekly item, use its schema-v1 route entry when available; otherwise open the entry file selected above and use its structure, links, and headings to narrow the scope.
3. Read the minimum primary source passages needed to support the answer. Prefer an official link, repository link, or paper already recorded in AI-Compass when the user asks for further reading.
4. Cross-check claims that mix freshness and theory: the weekly issue establishes when it appeared; the topic page establishes how it fits the longer technical landscape.
5. State uncertainty when the repository has no entry, a weekly claim is unverified by longer-term material, or a version/date may have changed since the stored issue.

Task Playbooks

Latest AI developments

Read weeklyHighlights/latest.md, then the linked issue. Return the issue date and number, rank only the developments relevant to the user's request, and identify the affected audience or use case. Do not summarize all entries unless the user explicitly asks for a full digest.

Historical comparison

Use weeklyHighlights/INDEX.md to identify the relevant issues first. Build a short timeline from the selected issue files, separate repository facts from your interpretation, and explain whether the observed change is a model release, product iteration, benchmark result, or ecosystem shift.

Technical explanation or recommendation

Start from the routed long-term topic page. Explain the concept, its role in a system, tradeoffs, and suitable scenarios. When recommending a tool or model, give the selection criteria and distinguish documented capabilities from your inference; do not imply endorsement merely because a project is listed.

Learning route

Use README.md to identify the user's current level and goal. Propose a short ordered route through existing modules, state what each step teaches, and avoid presenting the full repository tree as a required curriculum.

Project or codebase analysis

Read the relevant module overview before source files. Describe purpose, main components, data/control flow, configuration and external dependencies, then point to the smallest runnable or explanatory file set. Never invent setup instructions, APIs, or test results absent from the repository.

Answer Contract

• Answer in the user's language; preserve technical names and file paths exactly.
• Label time-sensitive findings as 周度增量 and stable explanations as 长期沉淀 whenever both are used.
• Cite repository-relative paths next to factual claims, for example weeklyHighlights/49.md or 3.2 Agent/3.Agent.md.
• Prefer concise synthesis over link dumping. Include the fewest sources that make the answer auditable.
• Clearly distinguish direct repository facts, source-linked claims, and your own synthesis or recommendation.
• Never claim that a topic, model, or feature is current beyond the latest issue date recorded in weeklyHighlights/latest.md.

Failure Handling

• If a term is absent, say that it is not currently recorded in AI-Compass after searching the relevant subtree; do not fabricate an entry.
• If internal sources disagree, report the conflict with paths and dates rather than silently choosing one.
• If the user requests real-time external facts, explain that AI-Compass is a local snapshot, use the latest recorded issue as the repository baseline, and only browse externally when the task separately authorizes it.