NyxID ICP Hunter Scoring
Use this when someone asks to run ICP Hunter — score likely NyxID ICP users, write new candidates to the ICP Hunter Lark Base, and post the run report.
You (the agent) run the WHOLE flow yourself with your own tools, then WRITE the results to Lark Bitable and post a summary to the chat. Do NOT refuse, do NOT just describe, do NOT claim "the skill only formats" — drive the flow. If a data source or the target base is unavailable, SKIP that part and still post what you have. Never ask the user for tokens — your NyxID-brokered tools broker all credentials.
How to run it
Resolve run context. Use today's date, compute
weekTagas ISO week (YYYY-Www), make abatchIdlikeicp-<timestamp>, defaultmax_resultsto5. Target Bitable: the ICP Hunter base baked into the builder (app_tokenBASE_APP_TOKEN_PLACEHOLDER_2,table_idTABLE_ID_PLACEHOLDER_3, chatoc_PLACEHOLDER_CHAT_ID_1); if the user supplies a realapp_token/table_id/chat_id, carry them into steps 6–8. Run the ICP query set: Claude Code/Codex/Cursor + MCP/API-key/credential pain,.mcp.jsonleaks/rotation, GitHub MCP config credentials, Home Assistant/Homelab credential pain, HA Forum agent/API-key threads, and n8n Forum credential-rotation threads.DISCOVER the search service, then COLLECT evidence. Do NOT hardcode a slug. First call
nyxid_services{action:"list"}and pick the caller's Tavily-compatible search service — match a service whose endpoint ishttps://api.tavily.com, or whoseslug/labelcontainstavilyorsearch. Use THAT service'sslugfor the search call:nyxid_proxy{slug:"<discovered-slug>", path:"/search", method:"POST", body:{ query, search_depth:"advanced", max_results:5, include_raw_content:true, include_domains }}. Keeptitle,url,content,raw_content, andscoreper result.- If no search service is found, do NOT fail. Include a one-line hint in the final run report
telling the caller to connect a Tavily search service in NyxID (endpoint
https://api.tavily.com, bearer auth; an org admin can share it team-wide vianyxid service add --custom --org <org>). NEVER ask the caller to paste an API key into chat — credentials are configured by the user in NyxID, never handled by this skill. Then continue withweb_fetchpublic-forum evidence only (degraded) and mark the search source as skipped.
- If no search service is found, do NOT fail. Include a one-line hint in the final run report
telling the caller to connect a Tavily search service in NyxID (endpoint
ENRICH platform evidence with what is actually bound.
- GitHub:
nyxid_proxyslugapi-githubforGET /users/{login}andGET /users/{login}/repos?sort=updated&per_page=5. - Public Discourse forums:
web_fetchthe public topic JSON, e.g.https://community.home-assistant.io/t/{slug}/{id}.json,https://community.n8n.io/t/{slug}/{id}.json,https://discourse.ros.org/t/{slug}/{id}.json. - ⚠
api-redditandapi-twitterare NOT bound — SKIP Reddit/X enrichment and note it in the report; score those candidates from the Tavilycontent/raw_contentand public web evidence instead. Do not block the run on them.
- GitHub:
ANALYZE/score with your own model. Score the enriched evidence against NyxID ICP fit and emit grounded candidate JSON with
users: username, priority, cohort, platform, profile/evidence URLs, scores, score breakdown, contact channel, language/timezone, raw evidence, recommended action, post date, thread role. Validate every username against the enriched author fields before keeping it; drop placeholders. For P0 users, draft outreach from the user's own raw evidence and keep it inoutreach_draft.READ existing Lark Base records (for dedupe) via
nyxid_proxy{slug:"api-lark-bot", path:"/open-apis/bitable/v1/apps/<app_token>/tables/<table_id>/records/search", method:"POST", body:{ page_size:500, field_names:["用户ID","平台","证据链接","Profile 链接"], automatic_fields:false }}. Use the returned records asexistingRecords(empty array if the base is unavailable).BUILD the exact payloads. Run
scripts/build_icp_hunter_payload.jsviacode_execute(language"javascript"; NO stdin — inline the data into the code):console.log(JSON.stringify(buildPayload({ mode: "build_payloads", users, existingRecords, runTime, weekTag, batchId, appToken, tableId, receiveId })))(runTime= ISO timestamp,receiveId= the target chat_id). If your model output needs parsing first, run the same script withmode: "parse_scoring". Use the output for every Base write and the report card. If code execution is unavailable, followreferences/icp-hunter-contract.mdfield-for-field.WRITE the Base records. For each
createRecordPayloads[], callnyxid_proxy{slug:"api-lark-bot", path:"/open-apis/bitable/v1/apps/<app_token>/tables/<table_id>/records", method:"POST", body:<that payload'sbody, i.e.{ fields: {…} }>}using the resolved app_token/table_id.POST the run report. Send the interactive card to the chat with
lark_messages_send(orreply_with_interaction) usingsendMessagePayload.contentfrom the builder. If no base target was resolved OR the writes fail, SKIP the Bitable writes and STILL post the built run-report card / summary to the chat — never fail the run for a missing base.REPORT one line: records written (or "summary-only"), P0/P1/P2 counts, and where the report was posted. If the builder returns
skip: trueor no source yields usable candidates, say so plainly.
Payload builder reference
Bundled script: scripts/build_icp_hunter_payload.js. Modes:
parse_scoring— accepts model output at root or underscoringResponse/geminiResponse/response; parses the first JSON object withusers; drops placeholder usernames.build_payloads— accepts scored users plus optional existing Base records; returns Base create-record payloads (createRecordPayloads),stats, and the Lark interactive report (sendMessagePayload). Required:usersandrunTime(plusfallbackPostDatewhen a kept user has no post date). Optional:existingRecords,weekTag,batchId,appToken,tableId,receiveId(base defaultsBASE_APP_TOKEN_PLACEHOLDER_2/TABLE_ID_PLACEHOLDER_3, chatoc_PLACEHOLDER_CHAT_ID_1).
Output may be skip: true, needs_more_information + missing, parsed users, or the full payload
object. Exact aliases, enums, dedupe rules, and examples are in references/icp-hunter-contract.md.
Guardrails
- Only use data the tools actually returned — never invent users, scores, outreach drafts, record ids, app tokens, table ids, chat ids, or records.
- Prefer running the bundled script; it is the deterministic source for field mapping, dedupe, stats, and the report card.
- All credentials and token exchange go through your NyxID-brokered tools; never handle raw secrets.