Research Dispatcher

Research Dispatcher

Single-command research across HN, Reddit, GitHub, and web search. Each source has its own query syntax, rate limits, and "what to put in the search box" pitfalls. This skill encodes those once so the agent doesn't have to remember.

When to use this skill

Activate when the user expresses any of:

• "Research X across HN, Reddit, GitHub"
• "What's the community saying about Y?"
• "Find me recent activity on Z"
• "Survey OSS projects related to W"
• "Compare what different platforms think of V"
• Any time multi-source synthesis would beat single-source search

Do not activate for:

• A single specific URL fetch (use WebFetch directly).
• Searches that require authentication beyond GitHub (e.g., LinkedIn, X/Twitter, internal Slack — those need different tools).

Sources covered

Bluesky, YouTube, TikTok, X/Twitter and other "Route B"-style sources are not directly supported by this skill but can be wired up via a host queue (out of scope here).

Procedure

Default flow

1. Parse user intent to determine sources to hit. If the user says "broad survey", default to HN + Reddit + GitHub. If they say "what's trending in <community>", lean Reddit. If "recent OSS work on X", lean GitHub.
2. Run the dispatcher with one command:

   ./scripts/research.sh "<topic>" --sources hn,reddit,github
   

3. Aggregate results into a single Markdown summary, deduplicating identical URLs across sources.
4. Surface signals: top stories by score, recent commits, recurring themes.
5. Recommend follow-up sources or queries if the initial sweep was thin.

Source-specific query tips (highlights)

See references/SOURCES.md for the full guide. Key tips:

• HN: spaces are AND. Use &numericFilters=created_at_i>UNIX_TS for date range. Topics like "AI safety" hit better than "AI regulation" on HN.
• Reddit: full-text search is weak; subreddit + /hot.json or /top.json?t=week gives better signal than /search. Always send a User-Agent header.
• GitHub: use qualifiers — topic:llm stars:>200 pushed:>2026-04-01. topic: is far stronger than free-text.
• Brave: freshness=pm for last month. Site-specific narrowing with site: operator. Watch the 2K/month free quota.

Choosing sources

+-------------------+-------------------+-------------------+
|     User wants    | Primary sources   |   Skip            |
+-------------------+-------------------+-------------------+
| Tech trends       | HN, Reddit        | GitHub (too noisy)|
| OSS activity      | GitHub, HN        | Reddit            |
| Community vibes   | Reddit            | HN, GitHub        |
| Breaking news     | HN, Brave         | GitHub            |
| Niche topic       | Reddit, Brave     | HN (sparse)       |
+-------------------+-------------------+-------------------+

Edge cases

• Zero results: surface the actual query_sent to each source so the user can see why. Suggest: shorter query, OR-joined synonyms, longer time window.
• Rate-limited: Reddit may return 429 with no User-Agent. The script always sends one, but if you see 429, wait 60s and retry.
• GitHub anon hit 60/hr: switch to authenticated mode by exporting GITHUB_TOKEN (a fine-grained PAT with public_repo is enough).
• Brave quota exhausted: skill should warn at 80% (1600/2000) and refuse at 100%. Fall back to other sources.
• API JSON shape changes: each parser is a few lines of jq. If a source breaks, update scripts/research.sh minimally.

Output format

Return a single Markdown response with:

1. One-line query summary
2. Per-source bullet list of top 5 hits (title, score, URL, 1-line preview)
3. Cross-source themes (if any obvious overlap)
4. A "what's missing" note if a source returned 0 hits or was skipped

Differentiation

• Generic web search (Brave, Google) misses community context. This skill brings HN/Reddit ranks into one view.
• Single-source agents (e.g. an HN-only summarizer) don't compare. This skill uses the contrast between sources as a signal.
• Manual multi-tab research: the goal of this skill is to replace 30 minutes of tab-switching with a single command.

Limitations

• No paywall content, no logged-in views (e.g. private subreddits, X/Twitter feeds).
• No real-time / WebSocket subscriptions.
• Score interpretation is naive — high HN score doesn't mean correct, just popular.
• Reddit /search is genuinely poor; subreddit + sort is the workaround but requires knowing relevant subreddits.