GitHub 产品调研与机会发现 · Product research
文档
market-recon
试用GitHub market recon when the user asks for AI money trends, hot niches, competitor signals, or "找风口/赚钱方向调研", especially when sources are restricted to GitHub only or a previous research run was interrupted. Produces an evidence-linked trend report.
它能做什么
GitHub market recon when the user asks for AI money trends, hot niches, competitor signals, or "找风口/赚钱方向调研", especially when sources are restricted to GitHub only or a previous research run was interrupted. Produces an evidence-linked trend report.
技能文档
Market Recon (GitHub-only)
Run a trend/opportunity recon where every claim traces to a github.com or api.github.com page fetched during this run. Never fill gaps from memory.
Steps
-
Fix constraints and scope. Default: only
github.comandapi.github.commay be fetched; noweb_search; no writes to user files; report lands in chat. State the constraint set (and any user-added ones) as the first line of the final report. Done when: the constraint set is written down for the report. -
Choose staffing. Recon needing ≤8 fetches: run inline. Larger: spawn 3–4 subagents, each task embedding the constraints verbatim plus its own URL list from
references/url-sets.md. Do not give subagents file access. Done when: every dispatched task carries the constraint block and its URL list. -
Fetch in capped batches. Max 5 parallel fetches (unauthenticated Search API allows ~10 req/min). Set
maxChars12000–20000 per fetch: Search API JSON runs 65–180KB and truncates otherwise. PreferextractMode: "text"for API JSON,"markdown"for HTML pages. Done when: trending pages and keyword queries returned HTTP 200, or the rate-limit fallback (step 3a) was used.3a. Rate-limit fallback. On 403/abuse signals: stop that batch, switch the remaining queries to
https://github.com/search?q=...&type=repositoriesHTML pages (still GitHub), and note the throttling in the report. Done when: throttling is either absent or disclosed. -
Salvage interrupted runs before re-running anything. If a subagent run was killed or a turn interrupted: pull each child session's
sessions_history(includeTools: true, small limit) and minetoolResultblocks for usable fetch payloads. Re-fetch only gaps the salvage cannot cover. Disclose killed runs and their recovered/lost data in the report. Done when: every usable payload from the dead run is accounted for and nothing was re-fetched needlessly. -
Synthesize the report. Group findings by opportunity (风口). Per repo: full name as link, stars, velocity (stars/day or /week), created/pushed dates, one-line monetization angle. Flag repos created <6 months ago with high velocity as "new". Close with 3–5 trend conclusions, each naming its evidence, plus the caveat: GitHub heat proves tool/content demand, not revenue. Done when: every claim has a link fetched this run and killed/missing data is disclosed.
-
Verify. Re-check the report: zero non-GitHub sources, no unfetched claims, rate-limit and truncation events disclosed,
stargazers_countvalues quoted only when actually seen (never "about"). Done when: the checklist passes; fix the report before sending if not.
Reference
references/url-sets.md— proven URL sets (trending pages, zh/en keyword API queries, vertical queries), URL-encoding for CJK queries, and batch/rate-limit notes.
相关技能
GitHub trend intelligence for AI agents — a daily scan of trending new GitHub repos, LLM-analyzed into a structured digest (top categories, language trends,...
Reconstruct open-source growth channels with evidence
GitHub & developer data for AI agents — repo metadata (stars, language, license), repo audits & risk scoring, user profiles, recent releases, and gist conten...
每小时聚合 Moltbook 社区摘要,提供免费轮询端点,搭配按次付费的爆款预测顾问。
Use when the user asks for GitHub Trending, GitHub hot repositories, daily/weekly/monthly trending repos, trending projects by programming language, spoken-l...