Hacker News-simulator – it predicted its own Show HN reception

# hackernews-simulator > Predict how Hacker News will react to your post before you submit it. [](https://www.python.org/downloads/) [](LICENSE) An ML-powered tool that predicts your Hacker News score, generates realistic simulated comments, and explains *why* your post will perform the way it will. Uses LightGBM for scoring, LanceDB for finding similar historical posts, and Claude for comment simulation.  ## The Meta Test We ran hackernews-simulator on its own Show HN post. Here's what it predicted: ``` ╭─────────────────────── HN Reaction Simulator ────────────────────────╮ │ Show HN: hackernews-simulator – predict how HN will react to your │ │ post before you submit it │ ╰──────────────────────────────────────────────────────────────────────╯ Predicted Score: ~4 points Reception: LOW (78% confidence) Percentile: Top 68.8% of HN stories Expected Score: ~10 points (multiclass) Why This Score: ↓ title_word_count (-0.01) ↓ has_url (-0.01) Simulated Comments: xkcd_fan_42: So you trained a model to predict HN reactions, and your own tool predicted a score of 4 for this post. That's either refreshingly honest or a sign you should have iterated more. throwaway_ml: This is a Claude wrapper that calls itself ML. LightGBM for the score prediction is fine, but the 'realistic simulated comments' part is just prompting an LLM. That's not a prediction, that's fanfiction. old_hn_lurker: Someone builds one of these every couple of years. The fundamental problem is that HN engagement is a fat-tailed distribution — most posts get 0-5 points regardless of quality, and virality is driven by a handful of power users and algorithmic timing. vecdb_curious: The LanceDB RAG piece is actually the most interesting part. Using similar historical posts as context for comment generation is a clever approach. ``` The tool is brutally honest about itself. We ship it anyway. --- ## Quick Start ```bash # Install pip install git+https://github.com/malakhov/hackernews-simulator.git # Download pre-trained models (~400MB) hn-sim init # Predict! hn-sim predict --title "Show HN: My Cool Project" ``` Or with [uv](https://docs.astral.sh/uv/): ```bash uv pip install git+https://github.com/malakhov/hackernews-simulator.git ``` ## Features ### Predict Score + Simulated Comments ```bash hn-sim predict --title "Show HN: I made a database in 500 lines of C" ``` Generates: - Predicted score and comment count - Reception label (flop / low / moderate / hot / viral) with confidence - Percentile ranking ("Top X% of HN stories") - 3-5 simulated HN-style comments with realistic usernames and tones - SHAP-based explanation of *why* this score - Best posting time recommendation ### Compare Variants Test multiple titles against each other: ```bash hn-sim predict --title "Why Rust is the future" --no-comments hn-sim predict --title "Show HN: I rewrote our API in Rust, here are the benchmarks" --no-comments ``` Or use YAML for batch comparison: ```yaml # variants.yaml variants: - title: "Show HN: My project" description: "A cool project" - title: "Ask HN: Would you use this?" description: "Thinking about building..." ``` ```bash hn-sim compare --variants variants.yaml ``` ### Suggest Loop (AI Title Optimization) Let Claude iteratively improve your title: ```bash hn-sim suggest-loop --title "My project" --max-iterations 5 ``` Each iteration generates new variants, scores them, and keeps improving until convergence. ### Web UI ```bash pip install hackernews-simulator[ui] hn-sim ui ``` Full Streamlit interface with predict, compare, and suggest tabs. ### Backtest Evaluate model accuracy on held-out data: ```bash hn-sim backtest ``` ## How It Works ``` Your Title + Description │ ├──→ [Feature Engineering] │ 15 structural features (title length, domain, time, etc.) │ + 384-dim sentence embeddings (MiniLM-L6-v2) │ + domain reputation (Bayesian smoothed) │ = 399-dimensional feature vector │ ├──→ [LightGBM Ensemble] │ Regression model → predicted score │ 5-class classifier → reception label + confidence │ SHAP TreeExplainer → "why this score" │ ├──→ [LanceDB RAG] │ Vector search → 5 most similar historical posts │ + their real comment threads │ └──→ [Claude CLI] System prompt with HN culture rules + similar posts as few-shot examples → 5 simulated comments with realistic tones Training data: 144K stories + 319K comments from open-index/hacker-news (2023-2025) ``` ## Requirements - **Python 3.11+** - **Claude Code CLI** — for comment generation and title suggestions ([install](https://docs.anthropic.com/en/docs/claude-code)). No API key needed — uses your Claude Code subscription. - ~400MB disk for pre-trained models Without Claude CLI, score prediction and SHAP explanations still work — only comment generation and suggest-loop require it. ## Train Your Own Model ```bash # Full pipeline from scratch (downloads ~2GB, takes ~30min) hn-sim init --from-scratch # Or step by step: hn-sim fetch --sample-size 200000 hn-sim train hn-sim build-index ``` ## Limitations We believe in honesty: - **~60-70% accuracy ceiling** — HN engagement is dominated by social cascade effects (who sees it first, early vote momentum) that are fundamentally unpredictable from content alone - **Score prediction is relative, not absolute** — the model is better at ranking ("A will outperform B") than exact numbers - **Training data is 2023-2025** — cultural drift will degrade accuracy over time - **Comment generation requires Claude Code CLI** — without it, you get scores but no simulated comments - **"Show HN" posts have different dynamics** than regular submissions The tool is most useful as a **probability advisor and creative testing tool**, not an oracle. ## Contributing See [CONTRIBUTING.md](CONTRIBUTING.md). ## License MIT — see [LICENSE](LICENSE).