agents2agentsai/hn-synthesis

HN Synthesis

Search Hacker News for practitioner discussions and synthesize them into structured knowledge base cards. This captures real-world signal that academic papers miss: deployment war stories, performance gotchas, community sentiment, alternative approaches practitioners actually use, and links to tools/repos that emerge in discussion threads.

Rules

1. Always use subagents — discovery and thread analysis are both done by subagents. The main agent does not call hn_search or hn_get_thread directly.
2. Start immediately. When the user asks to search HN, spawn the discovery subagent right away.
3. No KB references in prose. Never say "as found in your KB." Present findings as direct analysis.

When to Use This Skill

• User asks "what do people think about X?" or about real-world experience with a tool/library/technique
• User wants community sentiment or practitioner signal on a topic
• User asks to search Hacker News for something
• As a complement to paper-synthesis for practitioner perspective on research
• When investigating whether a technology has real-world traction beyond papers

Execution: All Work via Subagents

Two subagent types: $hn-discoverer (searches HN, deduplicates, ranks, returns top thread IDs) and $hn-synthesizer (retrieves and analyzes one HN thread, returns extracted signal).

Main Agent Flow

1. Spawn $hn-discoverer immediately (the subagent skill is hn-discoverer, NOT hn-synthesis):

   spawn_agent(agent_type="discoverer", message="$hn-discoverer\n\nTopic: [what the user is investigating]\n[If date range specified: Date range: YYYY-MM-DD to YYYY-MM-DD]\n[If specific keywords requested: Keywords: [list]]")
   

2. Wait for thread IDs.

3. Spawn $hn-synthesizer subagents in batches of 8 (the subagent skill is hn-synthesizer, NOT hn-synthesis). 20-thread system limit — spawning >10 at once causes silent failures. Wait once per batch, then spawn next. Cap at top 25 threads ranked by points x relevance.

   spawn_agent(agent_type="synthesizer", message="$hn-synthesizer\n\nThread ID: [HN item ID]\nHN URL: [discussion URL]\nArticle URL: [linked URL, if any]\nTopic context: [brief context from discovery results]")
   

4. Wait ONCE per batch — pass all subagent IDs in a single wait call. Do NOT call wait per subagent or poll in a loop.

5. Read all staging files via exec_command (e.g., cat ${CODEX_KB_PATH}/staging/hn-*.md).

6. Present a unified summary immediately (see Presentation). Do NOT write to KB before presenting.

7. If 3+ threads were analyzed, include cross-thread synthesis sections (see Cross-Thread Synthesis).

8. Spawn a $kb subagent (skip if KB is disabled) — fire-and-forget, do NOT call wait:

$kb

Process staged HN synthesis. Complete ALL steps: read staging files, write combined KB card, update journal+context, delete staging files. Card ID: [e.g. hn-ml-ai-agents-2026-02-20] Tags: [relevant tags] Staging files: [list all ${CODEX_KB_PATH}/staging/hn-<thread_id>.md files] User signals: [1-2 sentences about what the user asked for and interests revealed]

If KB is disabled: Skip step 8. Delete staging files with exec_command: for f in ${CODEX_KB_PATH}/staging/hn-*.md; do unlink "$f"; done. Use unlink, not rm -f — the sandbox blocks rm but allows unlink.

Pre-Synthesis Check (Optional)

Skip when the user explicitly asks to search. Only check KB when the request is ambiguous and KB is enabled: search cards for source_type: hackernews. If a matching recent card exists, return it and ask if the user wants a fresh search. If KB is disabled, skip entirely.

User provides a specific HN URL

Extract the item ID and skip discovery — spawn a single $hn-synthesizer subagent directly.

KB Card Persistence

Skip if KB is disabled. KB writes happen in the background via fire-and-forget $kb subagent, AFTER presenting. Each hn-synthesizer writes its analysis to ${CODEX_KB_PATH}/staging/hn-<thread_id>.md (with YAML frontmatter). The main agent reads staging files for presentation, then spawns $kb with card ID, tags, and staging file paths.

Card ID Convention

Kebab-case, prefixed with hn-: hn-{topic-slug}, hn-discuss-{slug}, or hn-survey-{slug}.

Cross-Thread Synthesis (3+ Threads)

When 3+ threads were analyzed, add before "Connections":

## Signal Evolution
<How has community opinion evolved across threads from different dates?>

## Consensus Map
<Broadly agreed (high confidence):>
- ...
<Contested (actively debated):>
- ...
<Acknowledged gaps:>
- ...

Presentation

Phase 1 (Outline): IMMEDIATELY call present_reading_view with document_id, title, and content containing ONLY ## section headings with empty bodies. This opens the reading view with "Generating..." placeholders.

Phase 2 (Fill): Call update_document_section(document_id, section_index=0, content="...") with full section content. Each tool result tells you the next section. Continue until all filled.

Markdown: Always put a blank line before list items (1., -, *). Without it, items after the first lose formatting.

Follow-ups: Use append_to_section with foldable=true for expansion requests (3-5 sentences per foldable block). Use patch_document_section for corrections. Use update_document_section only for full rewrites. Keep sections ≤30 visible (non-folded) lines. No editorial labels like "(expanded)" in headings.

Graceful Degradation

• No KB: Output full synthesis in chat; skip card storage.
• No threads found: Report and suggest alternative search terms.
• Few comments: Note that signal is thin and findings are preliminary.

Integration with Paper Synthesis

When used alongside paper-synthesis, cross-reference: mention existing paper cards in "Connections," note community-surfaced papers in "Resources Surfaced" as paper-synthesis candidates, and use HN cards to identify research directions with practitioner traction.