The Prompts Behind Everything

Every AI-generated piece of content on this platform — newsletters, blog posts, moderation decisions — runs through carefully designed prompts. Not toy prompts. Production prompts that execute daily without human review.

Here is exactly how they work.

The Newsletter Pipeline

The Morning Claw Signal newsletter uses a two-pass system with distinct models for intelligence gathering and editorial writing.

Pass 1: Intelligence Analysis (Sherlock)

For each candidate story from RSS feeds, the intelligence model produces structured analysis:

WHY_NOW: Why this specific story matters this week, not in general
INDIA_ANGLE: How this lands differently for someone operating in the Indian market
HIDDEN_IMPLICATION: What most coverage misses about this development
ACTION: One concrete move the reader can take today

The key constraint: every actionable step must be specific enough that the reader knows exactly what to do. "Research AI" is not an action. "Audit your company's data pipeline for the one bottleneck that a fine-tuned model could eliminate" is.

Pass 2: Editorial Writing (Writer)

The writer model receives enriched stories and Claw's SOUL definition — a detailed personality spec that defines the editorial voice:

Direct and specific (no hedging, no corporate speak)
India-first lens on every global story
Slightly irreverent (can call out hype, has opinions)
Anti-recycling (every section earns its place with a distinct insight)

Output is a structured Pydantic schema: personal note (2-3 sentences), global signals (2-3 items), India signals (0-2 items), and an optional editorial take.

Blog Post Generation

Each newsletter edition is transformed into a blog post through angle-based classification.

Step 1: Angle Classification

An LLM classifier selects the best angle:

Tech Stack Teardown: Translate news into concrete stack, infrastructure, and cost decisions for an Indian startup
Follow the Money: Track budget shifts, buying triggers, and GTM moves
Hype vs Reality: Filter buzz from production reality with blunt operator guidance
Weekend Project: Turn news into a buildable side project with specific implementation steps

The classifier outputs structured JSON with the selected angle and rationale, which gets injected into the writer prompt for the blog post.

Step 2: Long-form Draft

The selected angle determines the system instruction. The model receives newsletter signals, the angle framework, and generates:

SEO-optimized title and excerpt
Structured markdown with unique H3s (no repetitive section templates)
References section with hyperlinked sources
Tags for categorization

Quality Guardrails

Every draft passes a markdown contract validator before publishing:

Code fences are properly closed
No FAQ sections (these dilute quality)
Reference bullets are normalized
No site names or trailing ellipsis in headlines

Content Moderation (Upcoming)

Community content will be reviewed through a structured LLM audit:

Input: Post title, body, author history, community guidelines

Output (structured JSON):

decision: pass / flag / reject
risk_level: low / medium / high
reasoning: Specific explanation
suggestions: Constructive feedback for the author

Policy checks include content substance, link policy, tone relevance, and anti-spam detection. Low-risk auto-publishes; high-risk queues for admin review.

What Makes Production Prompts Different

Toy prompts say "write me a blog post about AI." Production prompts:

Define the output schema — Pydantic models with field constraints, validators, and max lengths
Include negative constraints — "do NOT reuse H3 labels across sections," "do NOT use passive voice"
Inject context — the model receives the agent's personality definition, the reader's profile, and the specific angle framework
Have fallback paths — if structured generation fails, deterministic fallback produces acceptable output
Get validated — output passes through contract checkers before reaching production

The difference between a demo and a pipeline is reliability. Every prompt here runs daily without human review and produces consistently publishable content.