PromptisePromptise
Why PromptiseUse casesRoadmapDocs
GitHubGet started
Promptise - AI Framework LogoPromptise

The foundation layer for agentic intelligence. Build, secure and operate autonomous AI systems with Promptise Foundry.

pip install promptise

[01] Foundry

  • MCPcast
  • The Promptise Agent
  • Reasoning Engine
  • MCP
  • Agent Runtime
  • Prompt Engineering
  • Execution Engine
  • Agent Identity

[02] Resources

  • Documentation
  • GitHub
  • Guides
  • Learning Paths
  • Questions

[03] Company

  • About
  • Terms of Service
  • Privacy Policy
  • Cookie Policy
  • Subprocessors

© 2026 Promptise by Manser Ventures. All rights reserved.

Open source · Python

← Learning paths> Prompt Engineering
Intermediate path

Mental Models: How to Think with LLMs

Start the pathAll paths

This path helps you shift from using prompts to thinking like a prompt engineer. Instead of teaching techniques, it introduces mental models. These are ways of seeing and reasoning about large language models. Each guide is a mindset essay with visual diagrams and reflection prompts that sharpen your intuition, inspire creativity, and build the resilience needed to invent new solutions. By the end, you’ll think like a builder, a critic, and a collaborator with AI.

19
Lessons
2.7 h
Reading time
Intermediate
Level
Free
Cost

The lessons

19 lessons

Read the lessons in order; they are arranged so each one prepares the next. Every lesson is a guide, so you can also open any of them on its own.

  1. [01]Are LLMs Stochastic Parrots? What the Term Gets RightAre LLMs stochastic parrots? What Bender et al. meant, what next-token prediction implies, where the term falls short, and the habit it should give you.Beginner · 5 minOpen →
  2. [02]Can LLMs Simulate Users? Role-Play, Personas and LimitsCan LLMs simulate users or customers? The simulator view of LLMs, what research says about personas, two real role-plays that went wrong, and fixes.Intermediate · 5 minOpen →
  3. [03]Why LLMs Get Niche Facts Wrong: Popularity and Training DataWhy LLMs know popular facts but invent niche ones: accuracy tracks how often a fact appears in training data. When to retrieve instead. Real example.Intermediate · 5 minOpen →
  4. [04]Prompts as Code: The LLM as a Probabilistic CompilerTreat prompts as code: write specs, enforce output shape with structured outputs, surface assumptions and test every change. Real date-extraction example.Intermediate · 5 minOpen →
  5. [05]Are LLMs Black Boxes? What You Can and Can't InspectAre LLMs black boxes? Why model explanations can hide the real cause, and what you can inspect instead: context, tool calls, counterfactual tests.Intermediate · 5 minOpen →
  6. [06]Why LLM Output Is Untrusted Input: Design for InjectionWhy to treat LLM output as untrusted: indirect prompt injection, OWASP LLM risks, a real injected support ticket, and design rules that contain the damage.Intermediate · 5 minOpen →
  7. [07]Why LLMs Fail Modified Puzzles: Pattern-Matching BiasWhy LLMs answer the famous version of a changed puzzle: training-data priors, research on counterfactual tasks, a real failure and how to prompt around it.Intermediate · 5 minOpen →
  8. [08]LLMs as Lossy Compression: Why They Can't Quote ExactlyWhy an LLM is a lossy compression of its training data: it keeps gist and loses exact wording, quotes and URLs. What to ask it for, with a real example.Intermediate · 4 minOpen →
  9. [09]How LLMs Generate Text: Next-Token Probabilities ExplainedHow LLMs generate text one token at a time: probabilities, temperature, top-p, and why early words commit the rest. Real example of a verdict-first error.Intermediate · 5 minOpen →
  10. [10]How to Brief an LLM: Context Gaps and Clarifying QuestionsHow to brief an LLM like a new colleague: why it lacks your context and won't ask, the jagged frontier, and a real example of guesses swapped for asks.Beginner · 5 minOpen →
  11. [11]How LLM Tool Calling Works: The Model Never Runs ToolsHow LLM tool calling works: the model writes a request, your code runs it. A real two-step refund exchange and design rules for tools, results and gates.Intermediate · 4 minOpen →
  12. [12]Why Can't LLMs Count Letters? Tokens and ProbabilityWhy LLMs miscount letters, reverse strings wrong and miss word limits: tokenization, sampling, real failures, and why exact work belongs in your code.Beginner · 4 minOpen →
  13. [13]Does AI Boost Productivity? The LLM as Cognitive MultiplierDoes AI make you more productive? What field studies show about the jagged frontier, and how to use an LLM to sharpen your judgment. Real plan example.Intermediate · 5 minOpen →
  14. [14]How Context Shapes LLM Output: The Model Continues ItWhy LLMs copy the patterns in your context, flaws included: next-token prediction, Codex and multi-turn research, a real example and three fixes.Intermediate · 4 minOpen →
  15. [15]Prompts as Interfaces: Design Prompts Like API ContractsDesign prompts like API contracts: delimited inputs, a fixed output, a defined failure path, versions and tests. A real classifier prompt before and after.Intermediate · 4 minOpen →
  16. [16]LLM as a Judge vs a Jury: Making AI Evaluations ReliableLLM-as-a-judge explained: position, length and self-preference bias, and how a jury of swapped orders, rubrics and mixed models fixes it. Real example.Intermediate · 5 minOpen →
  17. [17]AI Code Hallucinations: Why Plausible Output Fools YouWhy LLM code looks right but fails: plausibility vs correctness, hallucinated packages, a real regex that broke, and how to verify output by running it.Intermediate · 4 minOpen →
  18. [18]Why LLMs Give Similar Answers (and How to Get Variety)Why LLMs repeat the same kind of answer: mode collapse after RLHF. Get real variety with sets, named axes and verbalized sampling. Real naming example.Intermediate · 5 minOpen →
  19. [19]Why LLMs Mirror You: Sycophancy, Bias and Leading PromptsWhy LLMs agree with you: sycophancy from human-feedback training and bias from human text. Prompts that get honest feedback, with a real rating example.Intermediate · 5 minOpen →
Start lesson 01

Before you start

  • Basic familiarity with LLMs and prompt engineering (Beginner Path recommended)
  • Comfort reading analogy-driven essays
  • No coding or technical setup required

Skills you build

  • Systems thinking for LLM behavior
  • Adversarial & out-of-the-box thinking
  • Prompt design intuition beyond recipes
  • Recognizing model limitations and biases
  • Creativity and persistence in problem framing
  • Translating mental models into practical prompting strategies

By the end, you can

  1. 01Reframe LLMs as probability machines, not oracles, and design prompts accordingly
  2. 02Adopt adversarial and hacker-like thinking to stress-test their own prompts
  3. 03Use metaphors (mirror, actor, negotiator, amplifier) to inspire new solutions
  4. 04Recognize failures as data and turn them into iterative improvements
  5. 05Balance creativity with control by thinking of prompts as interfaces, not commands
  6. 06Develop the persistence and mindset to invent their own advanced prompting patterns
  7. 07See LLMs as collaborators, not tools — amplifying their own framing and imagination

Ready when you are.

19 lessons, in order. Each opens with the full list beside it, so you always know where you are and what comes next.

Start the pathThe guide library