Think-Reason-Learn
Research · Think-Reason-Learn · 9 November 2026

What gets a Show HN to 100 points in 2026

Across 85,270 Show HN posts, only 1.6% now reach 100 points, and the ones that do show real technical work, link to the project and sound like a person.

Score your own draft

Only 1.6% of Show HN posts reach 100 points now, down from 4.8% in 2024. The posts that get there tend to show deep technical work, link straight to the thing they built rather than a landing page, and read like a person talking about a project. Posts that read like marketing copy reach 100 points half as often as the rest. Posts built around an AI model reach it about a third less often.

Those findings come from 85,270 Show HN posts submitted between January 2024 and October 2026, and from a model we trained on them. Everything below is a correlation. Posts with a trait did better or worse, but that doesn't mean the trait caused it. Where a finding rests on few posts, we say so.

Show HN got harder

Share of Show HN posts reaching 100 points: 4.8% in 2024, 2.6% in 2025, 1.65% from January to May 2026, 1.6% from June to October 2026
PeriodPostsReached 100 pointsShare
202417,6218404.8%
202528,2767422.6%
Jan to May 202622,4853721.65%
Jun to Oct 202616,8882701.6%

More posts are competing for roughly the same front page. 2025 had 60% more Show HN posts than 2024, and fewer of them reached 100 points in absolute terms. We don't know how much of the rise comes from projects built quickly with AI tools, but the timing fits.

Fewer posts have body text now. In September 2025 about 73% of posts had text under the title. By October 2026 it was about 40%. That matters, because posts with text do better. In the last five months, 2.5% of posts with text reached 100 points, against 1.0% of posts without.

How we measured

We wrote 77 yes-or-no questions about a post, such as "Does the post read like marketing copy?" or "Is the project self-hostable?". A fast language model answered every question for every post. It read only what an HN reader sees before clicking: the title, the site the link goes to and the body text. A small scorer then learned from 4,885 posts submitted before June 2026 which answers go with reaching 100 points.

To test it, we scored 1,570 posts from June to October 2026 that the model had never seen. That's all 270 that reached 100 points plus a random sample of the rest, weighted back to the real mix. Pick one post that reached 100 points and one that didn't, and the model ranks the hit higher 82% of the time, an AUC of 0.82. A word-matching baseline that knows which words appear in hits gets 0.77.

The numbers below come from those later posts, not from the posts we wrote the questions on. We left out findings that only held on the older data. A question counts as true for a post when the model's answer is at least 50% yes.

What the posts that reach 100 points have in common

The post...Share of postsReached 100 when trueWhen falseRatio (95% CI)
Shows deep technical skill17%4.2%1.1%4.0x (3.0 to 5.3)
Credits its inspiration or prior art9%3.5%1.4%2.5x (1.8 to 3.4)
Links straight to the thing, not a landing page59%2.1%0.9%2.45x (1.8 to 3.4)
Has conversational asides9%3.45%1.4%2.4x (1.7 to 3.4)
Makes something visual or playable that non-programmers enjoy32%2.55%1.2%2.2x (1.7 to 2.8)
Names an existing tool and what it lacks10%3.05%1.45%2.1x (1.4 to 3.1)
Is self-hostable11%2.5%1.5%1.7x (1.2 to 2.4)

Deep technical work is the strongest signal. Posts that show real engineering reached 100 points four times as often as those that don't. "I implemented a neural network in SQL" (121 points) is the type. The title alone tells you someone did something hard.

Say where the idea came from. Posts that name their inspiration or the prior art they build on did 2.5 times as well. Posts that name an existing tool and say what it's missing did 2.1 times as well. Both tell readers you know the field, and both give them something to argue about in the comments.

Link to the thing itself. If the link goes to a demo, the code or the product, the post does 2.45 times as well as one that goes to a landing page or a waitlist. Readers want to try it, not sign up for it.

Sound like a person. Posts with asides, the small personal comments a builder makes about something they care about, did 2.4 times as well. "I spent 2 years designing a mechanical Magic Keyboard" (306 points) already sounds like a story.

What goes with fewer points

The post...Share of postsReached 100 when trueWhen falseRatio (95% CI)
Is built mainly around an AI model or LLM32%1.2%1.8%0.64x (0.47 to 0.86)
Has a coined, portmanteau-style name22%1.2%1.7%0.68x (0.47 to 0.93)
Reads like marketing copy*17%0.9%1.75%0.49x (0.31 to 0.74)
Is a business or SaaS product for companies*5%0 of 66 posts1.7%

* Fewer than 30 posts in this group reached 100 points, so treat these as weak signals.

A third of Show HN posts are now AI products, and as a group they reach 100 points about a third less often. That isn't a rule against them. The AI posts that made it have the traits in the first table. Of those that reached 100 points, 52% showed deep technical skill, against 15% of the AI posts that didn't.

Names coined to hint at what the tool does, like "SheetSQL", also did about a third worse.

Marketing copy (superlatives, slogans, lists of benefits, "join the waitlist") reached 100 points half as often. None of the 66 posts in our sample that pitched a business product to companies got there.

Four pieces of advice that don't hold up

"Make the title say what it does." It's good advice, but 90% of posts already follow it, so it doesn't separate the posts that do well (1.05x). What the scorer weighs most is whether the title makes you curious. Only 6% of titles fail that test, and they rarely reach 100 points. Just four of them did, which is too few to say how big the gap is.

"Link to GitHub." In 2024, posts linking to a code host reached 100 points 2.3 times as often as others. It was 1.9 times in 2025, 1.3 times in early 2026, and 0.86 times from June to October 2026. Whatever advantage a GitHub link had is gone.

"Ask for feedback." Posts that ask for specific feedback look better at first (1.9x). But only posts with body text ask, and posts with text do better anyway. Compared with other posts that have text, asking adds little (1.2x).

"Don't mention you used AI to build it." Before June 2026, saying you built the project with AI tools made no difference. From June to October 2026, the few posts that said so (4%) did 2.2 times as well as other posts with text. That rests on only 31 posts reaching 100 points, so hold it loosely. Still, it's the opposite of the usual worry. Building with AI and building an AI product are different things to HN readers.

Shorter titles do better

From June to October 2026, titles under 40 characters reached 100 points 3.2% of the time. Titles of 60 to 79 characters, the most common length, reached it 1.4% of the time. In 2024 length barely mattered: 4.6% against 4.7%.

How far a score can take you

We group posts into four levels by score. On the later posts:

LevelShare of postsReached 100 points
Weakbottom 50%0.2%
Averagenext 30%1.7%
Strongnext 15%3.5%
Front-page contendertop 5%9.2%

Even the top level misses nine times out of ten. Timing, who sees the post in its first hour and the second-chance pool decide a lot. "Show HN: Elevators" scored near the bottom and got 1,680 points. The model never opens the link, and sometimes the thing behind a plain title is just very good.

A score is most useful for comparing drafts. The top fifth of posts by score held 62% of the posts that reached 100 points, and the bottom half held 7%. If you have three titles and two versions of the text, score them all and post the best one.

Score your own draft

The model behind this post is Show HN Launch Predictor. Give it a title, the text and the link. It returns a level, the share of past posts at that level that reached 100 points, and the checks that pulled your score up or down.

curl https://api.thinkreasonlearn.com/v1/predict \
  -H "Authorization: Bearer $TRL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "think-reason-learn/show-hn-launch-predictor",
    "input": {
      "title": "Show HN: Pgline – a single-binary Postgres change-data-capture tool in Rust",
      "url": "https://github.com/danadev/pgline",
      "text": "Hi HN, I kept running Debezium + Kafka just to stream row changes from Postgres. Pgline is one 8 MB binary that reads the logical replication stream and delivers changes as JSON to webhooks, NATS or stdout."
    }
  }'

A call costs $0.01. /v1/improve goes one step further. It rewrites what pulled your score down, keeps your facts and voice, and returns the rewrite only if it scores higher.

Method and data

  • Posts. 85,270 Show HN submissions from 1 January 2024 to 13 October 2026. A hit is a post that reached 100 points.
  • Training. 4,885 posts from before 1 June 2026, sampled so that 40% were hits, with weights to restore the real rate. We wrote and chose the 77 questions on these posts only.
  • Test. 1,570 posts from 1 June to 13 October 2026: all 270 hits plus a random sample of the rest, weighted to the real mix. Hit rates are weighted. Counts are not.
  • Answers. A language model's probability of "yes" for each question. The tables count 50% or more as yes. The scorer itself uses the probability.
  • Intervals. 95% bootstrap confidence intervals on the ratios. We mark a finding weak when fewer than 30 hits sit on one side of it.
  • Model quality. AUC 0.82 on the test posts, against 0.77 for TF-IDF with logistic regression on the same text and 0.62 for rules on length, text and link domain.

Download all 77 checks as CSV, with the question we asked, the scorer's weight, and the hit rates on both the training and the test posts.