灵感捕手

概述

Hotspot Scout

Use this skill to discover fresh cross-platform content signals and convert them into account-fit ideas, creation references, and publishable content angles.

This skill is positioning-driven. Do not assume the account is AI Agent unless the user says so or the workspace account context clearly applies. First identify the creator's niche, audience, content promise, platform priority, proof assets, and risk boundaries.

Intake Protocol

Proceed when the user gives at least:

• field: professional field, topic category, business domain, or account track
• audience: who the content serves
• content promise: what the account helps the audience understand, decide, do, avoid, compare, or prove

If the direction is vague, ask 3-5 concise questions:

1. Who is the audience?
2. What field, niche, product, tool, or scenario should the account focus on?
3. What value should the account provide: tutorial, teardown, opinion, case, review, decision aid, trust building, or conversion support?
4. Which formats matter most: Xiaohongshu carousel/video, Douyin short video, Bilibili/YouTube, long article, or multi-platform reuse?
5. What are the boundaries: medical, legal, finance, education, advertising, traffic diversion, privacy, or unverifiable claims?

If the user is unsure, infer 2-4 possible account positions from their clues, compare them briefly, and ask them to choose one before a full scouting run.

Scouting Flow

1. Set the scouting window. Default to the last 7 days. Use 24-72 hours for "today", "latest", "fast-rising", or "recently rising". Treat this as a hard signal window, not a soft preference.
2. Write a compact Account Profile: field, audience, content promise, platform priority, proof assets, and risk boundaries.
3. Build a query matrix from:
   • Core terms: exact nouns, tools, products, roles, exams, destinations, diseases, methods, workflows, or scenarios.
   • Audience terms: who searches, saves, shares, comments, or buys.
   • Pain terms: anxiety, mistake, objection, misconception, before/after state.
   • Outcome terms: get, avoid, choose, buy, learn, pass, fix, compare, explain, prove.
   • Format terms: 教程, 避坑, 清单, 测评, 复盘, 案例, 对比, 模板, 攻略, 经验.
   • Adjacent terms: neighboring pains that reveal audience language even when not direct competitors.
4. Run the source preflight:
   • If agent-reach is installed, run agent-reach doctor --json first. Read references/agent-reach-adapters.md and choose only sources whose channel status and active backend are usable enough for the task.
   • For Xiaohongshu and Douyin on this machine, prefer OpenCLI browser-session adapters for read-only scouting.
   • Use the older local Xiaohongshu DPAPI wrapper and Douyin MediaCrawler wrapper only as fallback paths when OpenCLI is unavailable or insufficient.
   • If Agent Reach/OpenCLI is unavailable, continue with the local adapters in references/source-adapters.md.
5. Collect candidates and creation references:
   • Xiaohongshu: platform-native demand, saves, comments, title patterns, visual templates, using OpenCLI first.
   • Douyin: short-video hooks, pacing, emotional peaks, mass-market topic language, using OpenCLI first.
   • Agent Reach/OpenCLI sources: Twitter/X, Reddit, Xiaohongshu, Douyin, Bilibili, YouTube, GitHub, V2EX, RSS, web/Jina, Exa search, LinkedIn, Xueqiu, Xiaoyuzhou when relevant and configured.
   • AI HOT: macro AI/product/news context for AI, software, tools, productivity, or business innovation tracks.
6. Normalize candidates to:

type, platform, title, url, author, published_at, captured_at, likes, comments, collects, shares, views, tags, source_query, summary, source_role, source_confidence. Set type=signal only when the item has a reliable recent timestamp. Items without publish time are reference, not rising signals.

1. Score candidates. Prefer repeated-snapshot delta velocity; otherwise use freshness-weighted engagement inside the active time window. Read references/scoring.md or run scripts/score_hotspots.py. Do not let old high-like posts outrank recent rising posts as signals.
2. Split candidates into:
   • signal: recent rising evidence inside the active time window that explains why the topic matters now.
   • reference: reusable hook, structure, visual evidence, proof style, comment insight, or creator format.
   • idea: account-fit publishable adaptation derived from signals and references.
3. Translate references into account-fit ideas. Do not copy another creator's topic mechanically. Identify the audience job, creator permission, proof layer, platform-native hook, and risk boundary.
4. Write the three standard output files to hotspot-scout-reports/<YYYY-MM-DD>_<niche-slug>/.

Source Rules

Use fresh data. Do not answer latest/recent/today/fast-rising from model memory.

Read these references as needed:

• references/source-adapters.md: platform roles, local Xiaohongshu/Douyin paths, AI HOT, manual exports.
• references/agent-reach-adapters.md: Agent Reach doctor preflight, source routing, platform command families, and fallback rules.
• references/creation-reference.md: what counts as a reusable creation reference.
• references/scoring.md: cross-platform ranking, source confidence, and score interpretation.

Do not silently fall back from a platform-native source to weak web snippets. If a source is unavailable, report it as unavailable and lower confidence for any replacement evidence.

Freshness is mandatory for rising claims:

• A signal must have a timestamp inside the active scouting window.
• Content without publish time can be a reference, but not proof of recent heat.
• Old evergreen content with high lifetime engagement can support packaging, but must not be described as fast-rising unless repeated snapshots prove current growth.

For login or cookie-backed platforms, respect the user's account safety:

• Do not ask for primary account cookies unless the user explicitly accepts the risk.
• Prefer dedicated low-risk accounts for Twitter/X, Reddit, Xiaohongshu, Xueqiu, and LinkedIn.
• Do not post, comment, like, follow, DM, or operate pages; this skill is for reading and research.

Output And Delivery

Every scouting run must produce three files on disk under:

hotspot-scout-reports/<YYYY-MM-DD>_<niche-slug>/

The niche-slug is a short URL-safe label derived from the account field, e.g. ai-agent, insurance-broker, study-abroad.

File 1: Hotspot Brief (brief-<niche>.md)

Keep it under 80 lines. Include:

• Executive summary: strongest current signal and why it matters now.
• Account positioning input: field, audience, content promise, platform priority, proof assets, risk boundaries.
• Quick conclusions: top 3 actionable directions, first platform to publish on, directions to skip with reasons.
• Source coverage: which channels were used, which were unavailable, and confidence implications.
• Platform adaptations: Xiaohongshu title/carousel/caption, Douyin opening/visual proof/pace, Bilibili/YouTube/search/community adaptations when relevant.
• Risk boundaries: specific compliance constraints, not generic warnings.

File 2: Ranking Table (ranking-<niche>.md)

Include:

• One-line scoring method and window.
• Table columns: Rank | Signal | Platform | Why Rising | Evidence | Score | Account Fit | Confidence.
• At least one evidence URL or source reference per row.
• Rows sorted by score descending.

File 3: Merged Candidate Pool (pool-<niche>.md)

Use this schema:

Rules:

• Type is signal, reference, or idea.
• The Why Rising / Borrow This column must never be empty.
• Every row must have confidence: high, medium, or low.
• Ideas must include: audience job, creator angle, proof layer, hook, why now, borrow from references, and do not say.

Account-Fit Filters

Strong fits:

• The topic serves the stated audience and content promise.
• The creator can show proof instead of only opinion.
• The topic maps to a decision, mistake, workflow, transformation, comparison, or emotional tension.
• The execution can be adapted to the user's target platform format.
• The claim stays inside expertise and compliance boundaries.
• The same tension appears across two or more relevant source roles, such as social discussion plus platform-native saves, or GitHub issues plus Twitter debate.

Weak fits:

• Generic hot news with no account permission.
• Old evergreen content with high totals but no recent signal.
• Viral formats that require copying another creator's identity.
• Topics requiring guaranteed outcomes, medical/legal/financial conclusions, private-domain traffic diversion, or unverifiable claims.

Original AI Agent Account Mode

When the user's account is the original AI Agent / WorkBuddy account, use this positioning:

Use AI / Agent to redo human activities, then explain the process so ordinary users can replicate it.

Treat WorkBuddy as a concrete tool anchor, not the account boundary. Expand to Codex, Cursor, Claude Code, DeepSeek, MCP, browser automation, Agent Reach, and other agent tools when the activity requires them.

Strong fit means a concrete workflow demonstration with screen evidence, tool calls, grounded data, deliverable handoff, or pressure-testing. Good source roles include GitHub issues, Twitter/X discussion, Reddit pain, Bilibili/YouTube tutorials, V2EX threads, and Xiaohongshu/Douyin packaging references.

Delivery Checklist

Before handing off:

• Confirm all three output files exist.
• Confirm the pool table has no empty Why Rising / Borrow This cells.
• Confirm every row has a confidence label.
• Confirm source coverage and unavailable channels are documented.
• Confirm no latest/recent claim relies on model memory.
• Confirm risk boundaries match the account's actual compliance constraints.