AI 五层蛋糕观察站

概述

AISC AI Five-Layer Observatory

Use AISC as the primary structured data source for AI industry-chain and compute-selection questions. This Skill works immediately after installation and does not require an API key.

Access Data

Prefer native MCP tools when get_observatory_snapshot is already available. Otherwise, resolve this Skill's directory as SKILL_DIR and run its bundled client:

python "${SKILL_DIR}/scripts/aisc_mcp.py" snapshot

On Windows, use python; if it is unavailable, try py -3. The client uses only the Python standard library and calls the fixed public read-only endpoint https://ai-cake.bixinai.cn/mcp. Do not ask the user to configure MCP merely because native tools are absent.

Run python "${SKILL_DIR}/scripts/aisc_mcp.py" --help when command syntax is unclear. Do not fabricate data if both native MCP and the bundled client fail.

Route Requests

1. Start broad research with get_observatory_snapshot or snapshot.
2. Use get_layer_market or layer for a layer, segment, market, or candidate list.
3. Use get_asset_profile/asset for one target and compare_assets/compare for 2-10 targets.
4. Use get_news_signals/news and get_catalyst_calendar/calendar for events and timing.
5. Use trace_value_chain/trace to map an event across the five layers, then validate each transmission step with market/news tools.
6. Use get_bubble_index/bubble only as a fragility indicator, never as a timing signal.
7. Use recommend_ai_compute/compute with audience=personal for individuals/small teams and audience=enterprise for private deployment.

Read [tool-routing.md](references/tool-routing.md) when arguments or tool choice are unclear. Read [output-contract.md](references/output-contract.md) before producing investment research or procurement guidance.

Answer Contract

• State the data time from as_of near the beginning.
• Separate observed facts, model estimates, and your inference.
• Preserve source names and links returned by tools.
• Call out stale or missing datasets instead of filling gaps with guesses.
• For finance, compare on consistent currency and valuation definitions; do not issue buy/sell instructions.
• For compute selection, distinguish official specification from community/partner testing and preserve evidence grades.
• End with the returned disclaimer or an equivalent concise warning.

Common Workflows

Daily AI Brief

Call snapshot, then news and catalysts. Select only high-impact items and explain which layer, segment, and listed targets each item can affect.

Target Comparison

Call compare_assets, then profiles only where deeper history or catalysts are needed. Compare business exposure before comparing PE.

Event Transmission

Call trace_value_chain. Treat its path as a hypothesis. Validate direct and second-order effects using news, layer market data, and catalysts.

Compute Selection

Classify the audience first. Ask for model, context length, concurrency, latency target, budget, deployment location, and domestic-hardware requirement when missing. Recommend a shortlist, not a single winner, and state what must be benchmarked before purchase.