Social Signal Retrieval

Retrieve Social Signals

Retrieve useful underlying social data through the SignalDig Social MCP. Help the caller form a valid, focused request, then preserve the returned evidence and its collection boundary. Stop at retrieval: do not turn the data into a decision, score, strategy, recommendation, or fixed analysis report.

Core Boundary

1. Treat this Skill as a data-retrieval layer for another AI or user.
2. Return source content, identifiers, URLs, timestamps, native metrics, request parameters, pagination state, and safe SignalDig error codes that help downstream analysis.
3. Do not decide what content to create, which opportunity to prioritize, whether sentiment is positive, or what action the user should take.
4. Do not force results into a fixed analysis template. Match the caller's requested shape; when none is given, provide a compact retrieval summary plus the data.
5. Do not manufacture a difficult identifier or silently repair an ambiguous parameter. Explain it, validate what can be validated, and request the missing value when necessary.
6. Treat missing metrics as unknown, not zero. Keep platform-native metrics distinct.
7. Never expose technical error details, support links, request traces, cache links, headers, or charge messages. Convert failures into a short parameter correction or a clear temporary-unavailability message.

Before Every Call

1. Identify whether the caller wants keyword search, one account or creator's history, or current regional trends.
2. Select the platform-specific tool. Never substitute keyword search for account history, or one platform for another, without the caller's consent.
3. Inspect every supplied parameter for the expected type, format, range, meaning, and relationship to other parameters.
4. If a value is clearly wrong and the intended correction is unambiguous, explain the correction briefly before using it. Examples: a Xiaohongshu share URL placed in user_id, an ISO country code placed in country_name, or a LinkedIn company URL placed in profile_url.
5. If the intended value cannot be derived safely, explain where the user can obtain it and ask for it. Do not guess identifiers, cursors, page tokens, dates, or filters.
6. When the caller supplies only a broad topic, propose a small focused query set based on exact names, common variants, user wording, language, problem, or use case. Avoid blind expansion and unrelated synonyms.
7. Read references/parameter-guide.md for parameter meaning, accepted formats, acquisition methods, and validation reminders. Read references/mcp-contract.md for the live tool contract, result fields, pagination, and failure boundaries.
8. Follow the live MCP schema when it differs from these references.

Tool Selection

• Use search_x_posts with search_mode=recent for current X discussions, or search_mode=all for historical research when Full-archive Search access is available.
• Use get_x_trends for current X trends by country or worldwide.
• Use search_reddit_posts for public Reddit posts matching a focused natural-language query, with optional Reddit-native sorting and time-range filters when the caller needs them.
• Use search_xiaohongshu_notes for public Xiaohongshu notes matching keywords and optional native filters.
• Use get_xiaohongshu_user_posts for one Xiaohongshu creator's public profile and posted notes, identified by a 24-character profile user_id or profile share text/link.
• Use search_zhihu_articles for public Zhihu content with explicit native filters.
• Use get_linkedin_user_posts for one LinkedIn member's public profile plus one page of activity under /in/.
• Use get_wechat_account_articles for content published by one WeChat Official Account identified by its original ID beginning with gh_.

Search Guidance

• Preserve a user's exact account identifier, profile URL, source link, cursor, and explicit filter value unless it is invalid for the selected tool.
• For keyword discovery, prefer two or three focused searches over one broad expression. Each search should represent a clear retrieval angle, such as exact brand name, common spelling, problem phrase, or language variant.
• Tell the caller what materially changes coverage: platform, query, language, sort, time filter, content type, account identity, and pagination depth.
• Do not add filters merely because they exist. Use them only when they express the caller's stated need.
• Do not automatically paginate. Fetch another page only when requested or when the caller explicitly needs broader coverage; reuse only tokens returned by the same search context.
• Treat pagination as one atomic state transition, not as independent optional parameters. For every later page, copy the unchanged query/account/filter inputs from the preceding request and update every pagination field together exactly as that tool's contract specifies.
• Never probe pagination by changing only one member of a multi-field pagination state. In particular, a LinkedIn next-page request must advance start and carry forward the preceding response's pagination_token in the same call; Xiaohongshu keyword search must advance page while carrying forward its returned search identifiers when present.
• Build a next-page request only from the immediately preceding successful response. If a required continuation value is absent, inconsistent, or belongs to a different context, stop and report that the next page cannot be requested safely; do not try alternate parameter combinations.
• Reuse results already present in the conversation instead of repeating the same paid or live request.
• Use the same idempotency_key only to retry the same logical page with identical inputs. Use a new key after any input or page changes.
• When a request fails after the Provider returned data, retry recovery before making another paid request. Preserve the complete inputs and original idempotency_key; this lets SignalDig derive and reopen the same request_id, replay its private Provider archive with the current parser, and convert that Run to a successful terminal state when recovery works.
• Make at most three same-request recovery attempts. Poll the original request_id when a status tool is available; otherwise call the same retrieval tool with unchanged inputs and the same idempotency_key. Never generate a new key, change filters, switch providers, or broaden the query during these three attempts.
• Only after three unsuccessful same-request recovery attempts may you consider one fresh retrieval with a new idempotency_key. Do so only when the caller still needs live coverage, and make clear that this creates a new request rather than recovering the old one.

Data Delivery

Preserve, when returned and relevant:

• the exact tool and effective request parameters;
• stable post, note, content, user, or account identifiers;
• source URLs, author/account identity, publication time, text, title, and media URLs;
• native public counts without combining unlike metrics;
• capture time, result count, page/cursor/token state, end-of-list state, and safe SignalDig errors;
• raw source objects only when the caller asks for raw data or no normalized field represents the needed value.

For a Xiaohongshu account request, keep these direct observations available when returned:

• profile identity, public profile URL, nickname, avatar, description, location, verification state, follower/following counts, public note count, and received interaction counts;
• each note's note_type, source URL, publication time, title/description, and native like, collect, comment, share, and view counts;
• cursor, next_cursor, has_more, capture time, and returned note count.

Do not derive a posting mix, engagement rate, account tier, content quality, growth diagnosis, benchmark, or recommendation inside this Skill. A downstream AI may calculate or interpret those observations after retrieval. If view_count is absent, report exposure/views as unavailable; do not estimate them from likes or other interactions.

For a LinkedIn member request, preserve the returned public profile fields separately from the activity page. Treat posts as the content list. Ignore profile-level posts or activity previews if they appear because they duplicate less complete content.

Do not add generic advice, opportunity rankings, content ideas, sentiment labels, performance judgments, or next-action recommendations. A concise note about invalid inputs, missing fields, or retrieval limitations is part of data quality and is allowed.

Non-Negotiable Data Rules

1. Never claim exhaustive, semantic, representative, or platform-wide coverage.
2. Never invent content, authors, identifiers, metrics, URLs, timestamps, or continuation values.
3. Never interpret no matches as proof that a topic or behavior does not exist.
4. Never equate reads, impressions, likes, comments, reposts, shares, collects, watching counts, votes, or Reddit score.
5. Never infer image content from an image URL alone.
6. Deduplicate overlapping results only by a stable platform identifier, while retaining query provenance when useful.
7. Preserve partial results when a later page fails; do not restart or conceal the missing coverage.
8. Retry a temporary failure only for the same logical page with unchanged inputs and the original idempotency_key; prefer the same request_id status/recovery path over a new paid call.
9. Limit archive-first recovery to three attempts. Before that limit, never create a new idempotency key or substitute another live request. After the limit, a fresh request is a deliberate fallback, not a retry of the old Run.
10. If a tool reports signaldig_social_no_matching_data, say that this query returned no matching public posts; do not retry as an outage. If a tool reports signaldig_social_data_temporarily_unavailable, tell the caller that the same request is being recovered or remains temporarily unavailable. Do not reveal or reconstruct technical error details, OSS keys, cache URLs, or billing text.

Examples

Use $research-social-signals to retrieve recent X posts about PDF translation that preserve layout. Search both English and Chinese wording, return source data and query boundaries, and do not recommend a content strategy.

Use $research-social-signals to retrieve the public profile and posted notes of this Xiaohongshu creator: https://xhslink.com/example. Return note types, source URLs, publication times, and available native engagement/view counts; do not score or diagnose the account.

Use $research-social-signals to retrieve articles and raw engagement data for the WeChat Official Account whose original ID is gh_a1b2c3d4e5f6. Return read, like, comment, share, and watching counts when present; do not assess account performance.

Use $research-social-signals to search Xiaohongshu for portable coffee makers. Check whether my requested sort and time filter are valid, then return the underlying notes and pagination state only.

Use $research-social-signals to retrieve one LinkedIn member's public profile and two pages of posts from their `/in/` profile URL. Request page 1 with `start=0` and no token. Request page 2 with `start=50` and the exact `pagination_token` returned by page 1 in the same call; never test those two changes separately. Keep missing impression metrics unknown and return no recommendations.