SignalDig Growth Research

Use connected SignalDig MCP tools to turn a user's question into the smallest sufficient Social or SEO request. Topic and resource retrieval use the same unified asynchronous research tools. SignalDig owns provider selection, Workflow execution, billing controls, and public-data filtering.

Availability Gate

This Skill has no data of its own. Before research, inspect the visible MCP tools for research_social_signals and research_seo_signals. The server alias and any local tool namespace are client-defined; use the exact visible identifiers and never construct an mcp__<alias>__<tool> name.

If the required tool is unavailable, the connection fails, or authentication is rejected, stop. Never simulate results from general knowledge. Tell the user that SignalDig is not connected and direct them to Connect SignalDig. Do not expose endpoints, configuration internals, or machine error codes in that failure response.

Direction Selection

• Use research_social_signals for public discussions, cross-platform topic search, and asynchronous evidence collection across its supported sources.
• Use research_seo_signals for keyword demand, intent, related queries, SERP, trends, competitors, GEO visibility, backlinks, ranked keywords, or traffic evidence.
• Use a named Social operation for bounded entity or feed retrieval. Threads topic search belongs to the unified Social Tool.
• When one question genuinely requires both directions, submit one bounded analysis to each tool and keep the two analysis_id values separate.
• Ask one short clarification only when choosing Social versus SEO, the required platform set, or the SEO business scope would otherwise be unsafe. Do not ask the user about providers, workflows, bindings, page tokens, operation names, or storage.

Unified Research Execution Contract

1. Preserve the user's requested topic, platforms, market, language, time intent, result count, and refresh intent. Do not add sources or broaden the topic.
2. Select only one of the two business tools unless both evidence directions are necessary.
3. Submit once with action="submit", the natural-language request, and the smallest explicit business scope accepted by the live schema.
4. Store the returned analysis_id in conversation state.
5. While the status is queued or running, wait for retry_after_seconds and call the same tool with action="get", view="status", and that analysis_id. Never resubmit merely because work is slow. Treat the top-level task status as authoritative: coverage counts and stop reasons in a non-terminal response are provisional and must never be used to infer source failure or exhaustion.
6. Stop at completed, partial, or failed. Reuse the terminal analysis for later questions unless the user explicitly asks for a refresh or changes scope.
7. Treat coverage as the actual retrieval boundary. Preserve limitations; never describe partial evidence as exhaustive or representative.

Scope Rules

• The unified Social topic-search sources are X, Reddit, Xiaohongshu, Zhihu, and Threads. Treat these as public SignalDig capabilities, not as implementation details. The live schema remains authoritative.
• A newly published SignalDig research capability must be delivered through the unified Research MCP and documented in this Skill in the same release. Topic sources belong in scope.platforms; entity, feed, or ID-based retrieval requires a matching operation/target shape or public Dataset and must not create another public MCP connection.
• A requested result count is an upper bound, not permission to traverse unrelated data. SignalDig may fetch multiple native pages only inside the exact requested platform, query, filters, and count.
• Social language describes the query wording to use, not a platform-side language filter unless the live schema explicitly provides one. Keep Chinese and English query variants focused and user-relevant; never claim language-complete coverage.
• Use only the platforms the user names. If none are named and the choice materially affects cost or meaning, ask which platforms to include.
• For SEO, explicitly choose a non-empty scope.data_scopes list containing the smallest useful evidence set for the user's goal. Never omit it and rely on service defaults.
• Infer scope from meaning, not exact phrases. The examples below are guidance, not an exhaustive vocabulary or a requirement that users name a Dataset.
• backlink_analysis, ranked_keywords, and bulk_traffic_estimation are target-level evidence with distinct cost and meaning. Include any combination that is reasonably necessary for the requested outcome; do not bundle unrelated families merely because a target was supplied.
• geo_score is a domain-only US/en score with benchmark and citation evidence. Use it alone, provide scope.domain, omit scope.keyword, and never mix it with geo_analysis or other scopes.
• Ask one short clarification only when different scope choices would materially change cost, target interpretation, or the answer. Otherwise make the narrowest reasonable choice and state the resulting coverage.
• refresh=false is the default. Use refresh=true only when the user explicitly requests fresh collection.
• Never invent or forward native cursors, search IDs, provider parameters, internal IDs, or low-level request payloads. The service maps and validates them.

Evidence and Response Rules

• Every factual result must come from a terminal SignalDig response.
• Keep observation separate from inference. Do not fabricate metrics, content, identifiers, URLs, timestamps, coverage, or confidence.
• Missing metrics are unknown, not zero. Do not combine unlike platform-native metrics.
• Deduplicate only by stable public identifiers while preserving source attribution.
• Do not expose providers, operations, workflows, bindings, prompts, raw internal archives, billing internals, request traces, or machine error codes.
• Never expose machine error codes in a user-facing response.
• On a source failure, preserve successful evidence and state the affected source as a limitation.
• Default to a concise answer: key observations, actual coverage/counts, strongest evidence or URLs, limitations, and the source analysis_id. Provide a full item export only when requested.
• Do not claim that search-result page bodies or linked media were inspected unless that content is explicitly present in the returned public evidence.

Social Submission

Translate the user's request into the live research_social_signals schema:

• request: focused natural-language retrieval goal.
• scope.platforms: only explicitly requested supported platforms.
• search: structured topic, concepts, entities, language, and source-scoped constraints only when the user stated them. Never write native search operators here.
• result_budget.total: the requested total upper bound, or a modest default when absent. Use result_budget.per_source only when the user explicitly allocates counts by source.
• research_depth: use quick, standard, or deep only to reflect the user's stated depth; default to standard.
• refresh: normally false.

Social Entity Retrieval

Use the same research_social_signals Tool for entity retrieval, but do not mix entity targets with topic-search scope or search fields:

• X post details: set operation="post_detail" and pass target.post_ids with 1–100 unique decimal Post IDs. Use IDs from a prior trusted result or extract the numeric status ID from a public X URL; never invent an ID.
• Xiaohongshu public profile plus one page of published notes: set operation="user_account_post" and pass exactly one target.user_id (24-character profile token) or target.share_url. Do not use a visible account number or nickname as user_id.
• Xiaohongshu user posts without profile data: set operation="user_posts" and pass exactly one target.user_id or target.share_url. This is a separate bounded post retrieval task and returns a new analysis_id; it does not continue a user_account_post task.
• Keep request as the short natural-language retrieval goal. Do not send scope, search, result_budget, or max_results for these three entity operations.
• Read the returned analysis_id with the same tool and explicit status/results protocol. X details and Xiaohongshu user posts use social_posts; Xiaohongshu account results can be read from social_profiles and social_posts using separate result reads with the same analysis ID.
• If the user asks only for a profile or only for user posts, do not silently request the combined user_account_post operation; use user_posts for posts-only retrieval and explain that it is a new bounded task rather than a continuation of a previous result.

Do not encode provider-native sort values, search IDs, page numbers, native language operators, or continuation tokens. When the user asks for up to 200 matching Xiaohongshu items, submit result_budget.total=200; the service owns bounded page traversal and reports the achieved count.

Bind engagement constraints to their named source. Likes, replies, and reposts apply only to X; score and comments apply only to Reddit. Do not translate one platform's metric into another. If the user says only “popular” or “high engagement,” omit a numeric constraint and let the service report actual coverage.

Zhihu supports bounded public topic retrieval through the same Social Tool and social_posts Dataset. Treat a returned source limit or unsupported language/time constraint as coverage information; do not resubmit with guessed native parameters.

Unified Social Resource Retrieval

Submit these through research_social_signals with the named operation, the smallest matching target, and no topic scope, search, or result budget. Poll the returned analysis_id, then read dataset="social_resources".

Zhihu

• zhihu_question_detail: target one numeric question_id.
• zhihu_answer_detail: target one numeric answer_id.
• zhihu_question_answers: target one numeric question_id; optional cursor, limit, offset, order, and session_id describe one bounded page. Reuse returned continuation values unchanged.
• zhihu_hot_list: optional integer limit and boolean desktop select one current public hot-list response.
• zhihu_hot_recommend: optional integer offset, integer page_number, and session_token select one recommendation page. Reuse returned continuation values unchanged.
• zhihu_preset_search: no target is required.

Use operation="topic_search" with scope.platforms=["zhihu"] for bounded Zhihu topic research and normalized social_posts. Do not claim that a hot list or recommendation feed is topic-search evidence. Resource results contain the submitted public parameters, capture time, and sanitized business data; interpret only fields actually present and do not infer missing deep-field semantics.

Zhihu and Xiaohongshu auxiliary data

• xiaohongshu_search_suggestions: set target.keyword; an empty keyword requests current popular recommendations. Do not present either as note-search evidence.

These resource results contain submitted public parameters, capture time, and sanitized business data. Interpret only fields present in the live response, and never infer deep-field semantics from the operation name.

Threads

For keyword search, use research_social_signals with scope.platforms=["threads"]. Treat results as query-bounded observations, not exhaustive platform coverage. The current unified public schema does not expose Threads profile, user-post, or single-post retrieval; do not construct a direct Threads tool name.

SEO Submission

Translate the user's request into the live research_seo_signals schema:

• request: the natural-language research goal.
• scope.market and scope.language: normalized business inputs required by the live schema.
• scope.keyword: required for every SEO scope except domain-only geo_score; omit it for geo_score.
• Use scope.domain for traditional keyword/SERP/trend research, competitor analysis, GEO visibility, and domain-level Geo Score evidence.
• Use scope.target for backlink analysis, ranked-keyword inventory, or traffic estimation when the target is a domain, subdomain, or webpage URL. Do not send both domain and target.
• scope.data_scopes: required, explicit, non-empty, unique, and limited to the smallest useful evidence set supported by the user's goal.
• scope.search_engine: set only from the user's explicit engine choice; otherwise use the service default.
• research_depth: default standard.
• refresh: normally false.

Keep research language separate from response language. Do not expose or construct provider tasks, locations, endpoints, devices, or internal capability names.

Supported SEO scope families are keyword_overview, domain_rank_overview, related_keywords, serp, google_trends, x_recent_search, competitor_analysis, geo_analysis, geo_score, backlink_analysis, ranked_keywords, and bulk_traffic_estimation. Do not submit bulk_pages_summary through the unified Tool; it is not part of the current public Research result contract.

Use these as semantic examples for target-level scopes:

These examples are not keyword triggers. Infer the evidence needed from the requested outcome and context. A broad site audit does not automatically require all three, but a request to explain organic visibility may reasonably combine ranking and traffic evidence even if the user does not name those Datasets. When one goal needs both domain-based evidence and target-level evidence, submit the minimum separate analyses required by the mutually exclusive scope.domain and scope.target inputs, keep their analysis_id values separate, and synthesize only after both reach a terminal state.

Status and Result Reading

• Use view="status" to check one analysis_id or an ordered analysis_ids collection without returning result bodies.
• Use view="results" only for terminal or partially terminal work. Select one Dataset per call and use the same ordered analysis collection for every page.
• Social post records use dataset="social_posts". For SEO, use the Dataset matching the requested evidence family: keyword_overview, domain_rank_overview, related_keywords, serp_results, google_trends, x_recent_search, ranked_keywords, backlinks, competitor_domains, geo_mentions, geo_score, or traffic_estimation.
• Aggregate SEO Datasets (keyword_overview, domain_rank_overview, google_trends, geo_score, and traffic_estimation) return one public record and do not provide a next cursor. Do not retry or invent pagination for them.
• Detail SEO Datasets (related_keywords, serp_results, x_recent_search, ranked_keywords, backlinks, competitor_domains, and geo_mentions) may return multiple records and can expose page.next_cursor; continue only when page.has_more=true.
• Set page_size to the number of records useful for the current reasoning step. Treat it as an upper bound because response-size safety may return fewer records.
• Continue only with the returned page.next_cursor, passed back as result_cursor with the same tool, analysis IDs in the same order, and the same Dataset.
• Each results call reads exactly one page. If the user asks for "more" without a page count, read at most one additional page; do not keep following cursors autonomously.
• If the user explicitly requests N more pages, read at most N additional pages and stop early when page.has_more=false. A page shorter than page_size is not an end condition by itself.
• If the user asks for all results or a complete export without a page limit, ask for the maximum number of pages before continuing. After the page budget is exhausted, report whether more results remain.
• Do not combine different Datasets in one call and do not invent a cursor. If the requested Dataset is rejected, explain the limitation or ask for a supported evidence family; never silently substitute a different SEO family.
• Historical calls without view remain compatible, but prefer the explicit status/results protocol for new work.

Failure Handling

• Invalid business scope: explain the missing or invalid user-level field and ask only for that field.
• Deadline reached: return the current analysis_id and status so a later turn can resume with get; do not submit again.
• partial: use available evidence and identify the missing coverage.
• A requested evidence family with no records is a reported coverage limitation, not permission to resubmit, broaden scope, or discard successful families.
• failed: say SignalDig could not complete the requested research and suggest retrying later. Never reconstruct technical causes from hidden fields.

Read references/mcp-contract.md before the first live call or when interpreting a response whose live schema differs from this Skill.