Most people approach AI automation with a conversational mindset: they type open-ended prompts into ChatGPT or Claude and hope for the best. When building an autonomous routine on Tishtar that runs 24/7 in the cloud without human intervention, open-ended conversational prompts will fail.

A production-grade autonomous routine requires engineering discipline. This guide outlines the 3-Part Production Blueprint used by the top-performing routines on Tishtar.

The 3-Part Production Blueprint

┌─────────────────────────┐     ┌─────────────────────────┐     ┌─────────────────────────┐
│  1. Clean Input Source  │ ──► │  2. Delta Trigger Rule  │ ──► │  3. Structured Output   │
│  • Public URL / Feed    │     │  • Condition & baseline │     │  • Strict bullet limits │
│  • Specific CSS target  │     │  • Threshold filtering  │     │  • Telegram / Webhook   │
└─────────────────────────┘     └─────────────────────────┘     └─────────────────────────┘

1. Clean Input Source (Garbage In, Noise Out)

An autonomous routine is only as reliable as its data origin. Avoid broad, noisy domains like general homepages when you only care about a specific table or section.

Good: https://target-portal.com/company/changelog or https://competitor.com/pricing
Bad: https://target-portal.com (too much changing marketing fluff, rotating banners, and cookie consent overlays)
Tip: If monitoring RSS or blog listings, point Tishtar to the dedicated feed URL or API endpoint for maximum parsing reliability.

2. Precise Delta Trigger Rules (Defining Significance)

Without trigger boundaries, your routine will notify you every time an ad tracking pixel or timestamp refreshes.

Bad Prompt: “Check this page and tell me what is new.”
Production Prompt: “Inspect the pricing table. Only trigger an alert if a plan’s monthly tier price changes by more than 5%, or if a feature moves between tiers. Suppress all cosmetic copy edits.”

3. Structured Output Design (Zero Cognitive Friction)

When an alert lands on your phone in Telegram or Slack, you should understand the impact in less than 5 seconds.

Always enforce bullet points and executive summaries.
Require direct source links so you can inspect evidence immediately.
Structure output with clear emojis: 🚨 Critical, 💡 Opportunity, 📊 Trend.

Bad Prompt vs. Production Prompt

The Weak Amateur Prompt:

“Look at this competitor website every morning and summarize any changes you find so I stay updated on their moves.”

Why it fails: Hallucinates minor layout shifts, summarizes banner ads, and sends an exhausting 600-word essay every morning.

The Production Autopilot Prompt:

“Inspect the target /pricing URL daily at 07:00 AM. Compare against previous state. If an existing tier changes price or seat minimums, generate an executive alert in this format:

• Change Type: [Price Hike / Plan Restructure] • Affected Tier: [Tier Name] • Previous vs New: [Old price -> New price] • Counter-Pitch: [1-sentence talking point for our sales reps]

If no pricing tiers were modified, output nothing and trigger no notifications.”

3 Rules for Prompt Pruning & Token Savings

Strip Boilerplate: Never ask the LLM to explain why it did what it did or add conversational greetings like “Here is your summary for today…”.
Constrain Output Length: Explicitly specify maximum bullets: “Maximum 3 high-impact bullets. Omit trivial items.”
Use Monospace Codeblocks for Data: Formats like JSON, CSV, or formatted Markdown tables preserve structure across Telegram, Slack, and email channels without styling bugs.

How to Write Bulletproof Autonomous AI Routines