HN Distill

Fetch a Hacker News thread and distill it into an opinionated insight summary.

Dependencies

None beyond Node.js (uses curl for HTTP).

Analysis guidelines

Source critique

• Analyze the article before the thread. Identify core claims, evidence, and structural weaknesses.
• Test the source's evidence against its own thesis. Does the example actually demonstrate the claim, or reveal a confound? If the evidence undermines the thesis, lead with that.
• Separate claims when the source bundles them. Evaluate each independently — the thread may validate one and demolish the other.

Thread analysis

• Identify the important claims and push hard on them. Find the 2-3 that matter most and stress-test: what evidence supports them, what would falsify them, do the asserters have standing?
• Separate facts from judgment. Distinguish verifiable claims (data, direct experience, citations) from opinion, projection, or vibes.
• Be opinionated, not balanced. When two camps argue past each other, check if they're making the same mistake in opposite directions — name the shared fallacy.
• Insights must be non-obvious. Name the dynamic, not the state (what changed, what's accelerating, what mechanism appeared).
• Count signal, not volume. One well-argued contrarian comment outweighs 10 "+1" replies.
• Identify the framework if the thread converges on one.
• Note demographic splits. Flag astroturf/spam.

Quoting and attribution

• Quote with HN usernames.
• Interrogate celebrated comments. Ask why hasn't anyone done this already? — the answer usually reveals structural barriers the rant ignores.
• Elevate the star comment when one clearly outranks the rest. Don't force it.

Novelty extraction

Novelty is categorically different from insights. Insights are about thread dynamics (what people think, where they disagree). Novelty is about the world (someone built X, observed Y, and it's testable). Separate these explicitly.

• Hunt for empirical claims, not opinions. Novel = someone reports building, measuring, or observing something specific. "LLMs are bad at research" is opinion. "We built an arxiv MCP and saw a significant boost in method diversity" is a novel claim.
• Every novel claim must be specific and contextual. Not "AI is getting better" but "arxiv retrieval via MCP broke Claude out of weight-space recombination in our internal research workflow."
• Every novel claim must be falsifiable. State what would disprove it. If you can't, it's not novel — it's vibes.
• Every novel claim must cite its source exactly. HN username + direct quote. No paraphrasing that softens or inflates the claim.
• Develop implications. The claim itself is step one. Step two: what does it connect to? What compound system or second-order effect does it enable? This is where the analyst earns their keep — the commenter may not have drawn the full picture.
• "No novelty" is a valid finding. If no comment meets the bar (specific, falsifiable, cited), say so explicitly. That itself is signal — the thread was pure opinion.

Synthesis

• Verdict and "What the Thread Misses" must go beyond the thread. Name what the thread circles but never states. Identify blind spots. Flag meta-ironies (e.g., an AI article showing AI writing tells).

Scope

• 5-10 insights, 600-1500 words. Scale to thread complexity. Every sentence earns its place.

Step 1: Fetch the original article

Get the article URL from the HN API item endpoint, then fetch and read the article before the thread.

# Get the article URL from HN API (extract the "url" field)
curl -s "https://hacker-news.firebaseio.com/v0/item/<id>.json" | jq -r '.url // empty'

Fetch the article using the brave-search skill's content.js:

~/.pi/agent/skills/pi-skills/brave-search/content.js <article-url>

If that fails or is unavailable, search for the article:

~/.pi/agent/skills/pi-skills/brave-search/search.js "<article title>" --content -n 1

Skip this step for Ask HN / Show HN threads with no external link.

Step 2: Fetch the thread

{baseDir}/fetch.js <url-or-id>              # full thread
{baseDir}/fetch.js <url-or-id> --max 200    # cap at 200 comments

Accepts a full HN URL (https://news.ycombinator.com/item?id=12345) or a bare item ID.

Outputs a plain-text tree to stdout with metadata header and indented comments. Progress goes to stderr.

For very large threads (500+ comments), use --max to keep output within context limits.

Step 3: Fetch high-signal links from comments

Scan the thread for links cited as evidence for or against key claims — data, graphs, related threads, primary sources. Fetch these with content.js when they'd materially strengthen your analysis. Skip product plugs, tangential references, and anything that's just color. Typical threads have 0-3 links worth chasing; don't fetch more than 5.

Step 4: Analysis and Distill

Analyze the thread and produce a structured markdown distillation. Follow this template:

## HN Thread Distillation: "<thread title>"

**Article summary:** <1-3 sentence summary of the linked article, or "Ask HN" description>

### Dominant Sentiment: <characterize in 3-5 words>

<1-2 sentences on overall thread mood and whether it's notable>

### Key Insights

**1. <Insight title>**

<Analysis. Include the sharpest quote with HN username attribution.>

**2. <Insight title>**

...

### What's Novel

**<Claim title>** — [<HN username>]

> "<Exact quote from commenter>"

- **What it means:** <Specific, contextual description of the claim>
- **Why it matters:** <Implications, compound effects, what it connects to>
- **Falsifiable by:** <What would disprove this>

(Repeat for each novel claim. If no comment in the thread meets the bar — specific, falsifiable, empirically grounded — state: "No novel claims with evidence in this thread." and briefly note why, e.g. pure opinion thread, rehash of known positions.)

### Common Pushbacks

| Pushback | Quality | Note |
|----------|---------|------|
| ... | Strong / Medium / Weak / Misapplied | ... |

(Include when the thread has a clear debate axis. Omit for threads without structured disagreement.)

### What the Thread Misses

- <Blind spots, unasked questions, or angles nobody explored>

### Verdict

<2-4 sentences. Must name something the thread circles but never states.>

Step 5: Save to disk

Save the distillation as a markdown file. Default location depends on project context — follow the user's conventions for where research files go.