TrueAI Skill

Helps an agent (Claude / Cursor / Cline / ChatGPT / your own pipeline) recommend, look up and compare AI SaaS apps from the TrueAI 真选AI catalog without ever leaving the chat.

The catalog ships with rich structured fields documented in ai_tool_schema.json — name, slug, tagline, description, developer, form_factor, is_free, pricing_info, five sub-scores (overall / usability / effect / price / speed), official_url, categories, audiences, optional external_reviews, etc.

When to use this skill

Trigger it any time the user is doing one of the following:

1. Asking for a recommendation by need. _"我想做一份中文 PPT"_, _"need

a free AI to clean up my photos"_, _"我刚开始学编程，要免费的 AI 工具"_.

1. Asking about a specific app by name. _"Cursor 的定价是多少？"_,

_"通义千问支持 API 吗？"_, _"What is Genspark?"_.

1. Asking about an app by URL. The user pastes https://cursor.com

and wants to know what it is.

1. Browsing / comparing inside a category. _"Top 5 free AI coding

tools"_, _"列几个国内能直接用的图像生成工具"_.

Do not use this skill for non-AI products, general web search, news, or anything that doesn't map onto the TrueAI catalog.

Setup (one-time)

The skill talks to a running TrueAI backend. Pick a base URL:

# Public hosted endpoint (recommended — nothing to install):
export TRUEAI_API_BASE=https://www.shiflowai.cloud

# Local dev (if you're running the TrueAI repo yourself):
# export TRUEAI_API_BASE=http://localhost:8000

There are two equivalent ways to invoke the skill — pick whichever your agent supports best.

A. Native MCP (recommended if your agent speaks MCP)

The TrueAI backend exposes two MCP transports — pick whichever your client supports:

In Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json), Cursor (~/.cursor/mcp.json), Cline, Continue, Windsurf:

{
  "mcpServers": {
    "trueai": {
      "url": "https://www.shiflowai.cloud/mcp"
    }
  }
}

(For local dev replace the URL with http://localhost:8000/mcp.)

You'll then see four tools available to the agent:

B. Plain HTTP (works with any agent, no MCP needed)

The same data is also reachable via the REST API. Any agent that can run shell commands or do HTTP calls can use it directly.

All URLs below assume TRUEAI_API_BASE=https://www.shiflowai.cloud.

All responses are JSON. Schema: see docs/ai_tool_schema.json.

Recipes for the agent

Recipe 1 — User describes a need → recommend tools

curl -sG "$TRUEAI_API_BASE/api/search/chat" \
     --data-urlencode "q=我是新手想做小红书封面，要免费的 AI"

Response shape:

{
  "intents":    { "categories": ["image","design"], "audiences": ["beginner"], "free": true, "no_vpn": false },
  "candidates": [ /* Tool briefs, already ranked by overall_score */ ],
  "external":   [ /* fallback nav sites only when candidates is empty */ ]
}

Present the top 3 candidates with a one-sentence rationale per item, then link the official_url. If candidates is empty, surface the external nav sites as a fallback rather than apologizing.

Recipe 2 — User names a specific app

# By slug (preferred, exact)
curl -s "$TRUEAI_API_BASE/api/tools/cursor"

# By name — fall back to fuzzy search and take the top hit
curl -sG "$TRUEAI_API_BASE/api/tools" --data-urlencode "q=通义千问" \
  | jq '.items[0].slug' \
  | xargs -I{} curl -s "$TRUEAI_API_BASE/api/tools/{}"

Recipe 3 — User pastes a URL

# Extract the brand keyword from the host (cursor.com → cursor) and search
HOST=$(echo "https://cursor.com" | awk -F/ '{print $3}' | sed 's/^www\.//')
BRAND=$(echo "$HOST" | awk -F. '{print $(NF-1)}')
curl -sG "$TRUEAI_API_BASE/api/tools" --data-urlencode "q=$BRAND" \
  | jq -r '.items[] | select(.official_url and (.official_url | contains("'"$HOST"'"))) | .slug' \
  | head -1 \
  | xargs -I{} curl -s "$TRUEAI_API_BASE/api/tools/{}"

(Or just call get_ai_tool over MCP — it does the same resolution internally.)

Recipe 4 — Add real-user reviews for "usage tips"

curl -s "$TRUEAI_API_BASE/api/tools/cursor/external-reviews?page_size=5"

Surface 2–3 of the most upvoted reviews as quotes when the user asks _"how do people actually use it?"_ or _"is it any good?"_.

Recipe 5 — Compare a shortlist

curl -s -X POST "$TRUEAI_API_BASE/api/tools/compare" \
     -H "content-type: application/json" \
     -d '[9, 16, 23]'

Returns each tool's five sub-scores side by side — render as a markdown table.

Output guidelines for the agent

• Always include `official_url` when recommending — let the user click

through.

• Quote sub-scores honestly (`overall_score / usability_score /

effect_score / price_score / speed_score, all 0–10). A 0` means "not yet rated", not "bad".

• Respect `is_free`, `need_vpn` and `form_factor` when the user has

expressed a constraint about money, location or platform.

• Don't fabricate fields. If a field is missing from the JSON

response, say so or just omit it.

• Prefer 1–3 recommendations when the user is undecided; offer to

drill into any of them.

Field reference (excerpt)

Full schema (34 fields + nested Category / ExternalReview): docs/ai_tool_schema.json.

License

MIT. See the TrueAI repository root.