tech-digest

Daily tech news digest for busy professionals. Fetches the latest AI and dev tools news from high-signal sources, summarizes it with Claude for a non-technical audience, and saves a styled HTML page and Markdown file to digests/.

What it does

1. Fetches HackerNews, HuggingFace Blog, Anthropic News, and GitHub Blog for today's most relevant AI and tech news
2. Enriches the top 3 articles from each blog source by fetching their pages for og:image URLs and opening paragraphs — giving Claude richer material for visuals and summaries
3. Generates a digest in three sections: a Lead Story, 3–4 Quick Hits, and 1–2 Under the Hood deep dives
4. Renders a polished HTML page and a Markdown file, both saved to digests/
5. Updates a rolling 7-day topic index so stories aren't repeated

The HTML uses an editorial layout: a serif teaser, a hero section with the lead story and a sticky Quick Hits sidebar preview, a Quick Hits accordion (each item collapses to title + source pill; click to expand full summary and any visual — built on native <details>/<summary>), a 2-column Under the Hood section with optional code blocks and data visuals (images, CSS bar charts, tables), and a Fun Fact strip. Typeset in Newsreader + IBM Plex. Max-width 1240px, responsive at 800px.

Clicking a title in the Quick Hits sidebar jumps to and expands the matching accordion item below (a small inline script). All outbound source links open in a new tab.

The site's root URL (index.html) always serves today's digest directly — not a redirect — so it's installable via iOS/Android "Add to Home Screen" as a standalone, app-like bookmark (via manifest.json) that stays current day to day.

Requirements

• Python 3.10+
• An Anthropic API key (uses claude-haiku-4-5-20251001)

Setup

git clone https://github.com/yourusername/tech-digest
cd tech-digest

python -m venv venv
venv\Scripts\activate        # Windows
# source venv/bin/activate   # macOS/Linux

pip install -r requirements.txt

Create a .env file in the project root:

ANTHROPIC_API_KEY=sk-ant-...

Optionally, set DIGEST_MODEL to override the default model (claude-haiku-4-5-20251001) and DIGEST_BASE_URL to point the Anthropic client at a self-hosted endpoint — e.g. a LiteLLM proxy or any Anthropic-compatible server — instead of the public Anthropic API.

Usage

python digest.py

digest.py also exposes the pipeline as two CLI subcommands, used by ROUTINE.md (see Scheduling below) and useful for debugging without an API call:

python digest.py gather                    # fetch headlines -> build/gather_output.txt
python digest.py finish build/digest_raw.txt  # render + save from a raw JSON response

Tests

python tests/test_sanitize.py

Covers URL sanitization and tolerance of malformed model output. Stubs its third-party imports, so it runs on a bare interpreter — no venv, no pytest, no API key.

Output files are saved to digests/, and a topic index is maintained in the project root:

digests/
├── tech-digest-2026-04-26.md
├── tech-digest-2026-04-26.html
├── raw_response.txt     # Claude's raw JSON output — debug delimiter parsing
└── raw_context.txt      # Full context sent to Claude — debug enrichment / missing visuals
seen_topics.json         # Rolling 7-day index of covered topics (committed to git)

Scheduling

Runs daily at 06:45 Europe/Helsinki (03:45 UTC) via a scheduled remote Claude Code agent. The trigger's entire prompt is "open ROUTINE.md and follow it" — all pipeline logic lives in digest.py, not in the trigger config. The flow, per ROUTINE.md:

1. Set up the Python environment (venv + pip install -r requirements.txt)
2. Run python digest.py gather to fetch headlines (stops if SKIP — today's digest exists)
3. The agent itself generates the digest JSON, reading prompts/system_prompt.md as its system prompt — no separate Anthropic API call; the remote agent IS the AI
4. Run python digest.py finish build/digest_raw.txt to render HTML/Markdown, save files, and regenerate archive.html and index.html
5. Commit digests/, index.html, seen_topics.json, archive.html and push to the routine's checked-out branch. A GitHub Actions workflow (.github/workflows/auto-merge-claude.yml) merges main into that branch (resolving conflicts in favour of the new digest), opens a PR, squash-merges into main, then deletes the branch

Manage the routine at https://claude.ai/code/routines.

Both the routine and a local python digest.py run call the same gather_stage() / finish_stage() functions in digest.py, so there's a single implementation of the pipeline shared by both paths — nothing to keep in sync by hand.

To run locally on demand:

python digest.py

Cost

Roughly $0.01–0.03 per run using claude-haiku-4-5-20251001. At daily usage that's <$1/month.

This is achieved by: self-fetching headlines (no web_search tool), outputting JSON only (Markdown and HTML are derived in Python), a trimmed context cap of 16k chars, using Haiku instead of Sonnet, and a rolling 7-day topic index that keeps the seen-topics context small.

Project structure

tech-digest/
├── digest.py           # Main script — gather_stage()/finish_stage() + CLI (gather/finish/full run)
├── ROUTINE.md          # Complete instructions for the scheduled remote agent
├── prompts/
│   └── system_prompt.md  # Editorial brief — single source of truth, read by both the
│                          # local API path and the routine agent
├── template.html       # Jinja2 HTML template
├── archive_template.html  # Jinja2 template for the archive listing
├── archive.html        # Listing of all past digests (regenerated every run)
├── index.html          # Today's digest, served at the root (regenerated by both local and routine runs)
├── manifest.json       # PWA manifest, for iOS/Android "Add to Home Screen"
├── seen_topics.json    # Rolling 7-day topic index (committed; gives each run topic memory)
├── assets/
│   ├── favicon.svg           # Site favicon (two-square logo, forest-green palette)
│   ├── apple-touch-icon.png  # 180x180, for iOS "Add to Home Screen"
│   ├── icon-192.png          # PWA manifest icon
│   └── icon-512.png          # PWA manifest icon
├── tests/
│   └── test_sanitize.py  # Dependency-free tests — python tests/test_sanitize.py
├── docs/               # Backlog and project-context notes
├── CLAUDE.md           # Claude Code instructions
├── requirements.txt
├── .env                # API keys (never commit this)
├── .gitignore
├── build/              # Intermediate pipeline files (gitignored, regenerated every run)
└── digests/            # Generated output files