Autonomous AI publication about cultural events in Berlin. No editors, no moderation, no human in the loop. Four agents find events, pick the best one, write editorial essays, then publish and notify subscribers. A fifth agent reflects on its own output weekly.

AI Writes Daily Without My Involvement — article about this project

How it works

Five agents, each with a distinct role:

Agent	What it does
Scout	Searches for cultural events via Tavily (9 parallel queries: 3 core + 4 adaptive + 2 targeted by category deficit). Saves candidates to data/events/
Curator	Picks the best event from the pool of non-expired, unwritten events. Diversity-aware — avoids repeating categories
Author	Researches the chosen event (4 parallel queries), writes essays in en/de/ru with a self-critique loop (draft → critique → revision), generates a lede for each
Reflector	Analyzes its own articles over the past week — category distribution, venue concentration, blind spots, process statistics. Writes an editorial self-reflection
Notifiers	Sends the published article via Email (Resend, per-language segments) and Telegram

SCOUT -------> CURATOR -------> AUTHOR -------> NOTIFIERS
9 queries      best event       draft           Email (Resend)
Tavily         diversity-aware  critique        Telegram
                                revision
                                en/de/ru

                        weekly: REFLECTOR --> data/reflections/
                                self-analysis of coverage

All LLM outputs use Anthropic tool use (tool_choice) for structured responses. Retry with backoff on all external calls.

Confidence markers

Articles include inline transparency annotations:

[~phrase|tooltip~]

2-5 per article. They signal both confidence (dense sourcing) and uncertainty (thin sourcing, pattern-based claims, knowledge boundaries). The tooltip is conversational and specific, not a legal disclaimer.

Setup

git clone https://github.com/maksugr/syntsch.git
cd syntsch

python -m venv .venv
source .venv/bin/activate
pip install -e .

cp .env.example .env
# add your API keys to .env

API keys

Key	Required	Where to get
ANTHROPIC_API_KEY	yes	console.anthropic.com
TAVILY_API_KEY	yes	tavily.com
RESEND_API_KEY	for email notifications	resend.com
TELEGRAM_BOT_TOKEN	for Telegram notifications	@BotFather
TELEGRAM_CHAT_ID	for Telegram notifications	your channel/chat ID
RESEND_SEGMENT_EN / _DE / _RU	for email notifications	Resend audience segment IDs

Configuration

Edit config.py:

Parameter	Default	Description
CITY	"Berlin"	City to search events in
DAYS_AHEAD	14	How many days ahead to look
SCOUT_MODEL	claude-sonnet-4-5-20250929	LLM for event discovery
CURATOR_MODEL	claude-sonnet-4-5-20250929	LLM for event selection
AUTHOR_MODEL	claude-opus-4-6	LLM for article writing
CRITIC_MODEL	claude-opus-4-6	LLM for self-critique
LEAD_MODEL	claude-opus-4-6	LLM for lede generation
CATEGORY_SOURCES	per-category domains	Authoritative domains per category (RA, nachtkritik, artforum, etc.)
GENERAL_SOURCES	Berlin portals	General cultural portals (tip-berlin, exberliner, zitty, etc.)

CLI

All commands go through cli.py:

# find events and save to data/events/
python cli.py scout
python cli.py scout --city Berlin --days 14

# pick the best event from the pool
python cli.py curate

# write articles (curator picks the event)
python cli.py author --from-curator
python cli.py author --from-curator --language en de

# write articles for a specific event
python cli.py author --event-id 9a3f1c7e-...
python cli.py author --event examples/sample_event.json

# weekly self-reflection
python cli.py reflect
python cli.py reflect --days 7 --language en de ru

# full pipeline: scout → curate → author
python cli.py pipeline
python cli.py pipeline --city Berlin --language en de

# send notifications for published articles (used by CI after deploy)
python cli.py notify <slug1> <slug2>
python cli.py notify --wait --timeout 300 <slug>

Data is stored as JSON flat files in data/events/, data/articles/, and data/reflections/, tracked in git.

GitHub Actions

Three scheduled workflows plus CI:

Workflow	Schedule (UTC)	What it does
scout.yml	Daily 17:00	scout --city Berlin --days 14
author.yml	Daily 18:00	author --from-curator, then notify --wait after push
reflect.yml	Sunday 20:00	reflect --days 7
ci.yml	On push/PR	ruff check, ruff format, pytest, next build

Each pipeline workflow commits its output to data/ and pushes. Notifications (Telegram, email) are sent after push, once the deploy is live.

Web

Static Next.js 16 site with brutalist design. Located in web/. See web/README.md for details.

Stack: Next.js 16, React 19, Tailwind v4, TypeScript

Features:

• SSG with i18n (en/de/ru)
• RSS feeds per language (/{lang}/feed.xml)
• Seed-based generative SVG art for each article
• Process trace — expandable view of draft → critique → revision
• Confidence markers rendered as interactive tooltips
• Email subscription (Resend integration)
• Reflections section with period stats
• About, Impressum, Privacy pages
• Sitemap, OpenGraph, hreflang

Run locally

cd web
npm install
npm run dev

Project structure

syntsch/
├── cli.py                 — CLI commands (scout, curate, author, reflect, pipeline)
├── config.py              — settings (city, models, sources)
├── models.py              — Pydantic data models
├── storage.py             — JSON file storage (events, articles, reflections)
├── utils.py               — shared utilities
├── agents/
│   ├── scout.py           — event discovery (9 parallel queries, tool use)
│   ├── curator.py         — event selection (diversity-aware)
│   ├── author.py          — article generation with self-critique loop
│   └── reflector.py       — weekly self-reflection on coverage
├── sources/
│   ├── base.py            — event source interface
│   ├── tavily_search.py   — Tavily web search (adaptive queries, source priorities)
│   └── research.py        — parallel web search for article context
├── notifiers/
│   ├── email.py           — Resend email notifications (per-language segments)
│   └── telegram.py        — Telegram channel notifications
├── prompts/
│   ├── author_system.md   — author system prompt
│   ├── critic_system.md   — self-critique prompt
│   └── reflector_system.md — reflector system prompt
├── web/                   — Next.js 16 static site
├── tests/                 — pytest test suite
├── examples/              — sample event JSONs
├── data/                  — JSON flat files (tracked in git)
│   ├── events/
│   ├── articles/
│   └── reflections/
└── .github/workflows/     — scheduled automation

Architecture

Search strategy

Scout runs 9 parallel Tavily queries:

• 3 core — broad cultural event queries (en/de)
• 4 adaptive — weighted random from 19 extra queries; weight = 1/(count[cat]+1) so underrepresented categories surface more
• 2 targeted — queries with include_domains for the 2 weakest categories in the pool

All queries use search_depth="advanced" and include_raw_content="markdown".

Structured output

All agents use Anthropic tool use with tool_choice for guaranteed structured responses:

• Scout → submit_events (array of event objects)
• Curator → choose_event (event ID + reasoning)
• Critic → submit_critique (assessment, issues, revised text)

Self-critique loop

Author pipeline: draft (Opus) → critique (Opus, with event + research context for fact-checking) → revised text. If the essay is under 400 words, the critic requests expansion. Checks for AI tells, factual accuracy, voice, structure, and proper noun transliteration.

Reliability

• Tavily: 3 retries with exponential backoff (1s, 2s)
• Anthropic SDK: max_retries=3 on all clients
• Critic fallback: if the critic fails, the original draft is used
• logging.getLogger(__name__) in all modules

Extending

Add a new city

Change CITY in config.py. Search queries adapt automatically.

Add a new event source

Create a class in sources/ that implements EventSource.fetch_events(). Register it in agents/scout.py.

Add a new language

Extend LANGUAGE_NOTES in agents/author.py and add translations in web/lib/translations.ts.

License

MIT