LearnByShortsAI Fluency Lab Build a skill →
← All learning paths

Free · zero to advanced

Don’t just prompt AI.
Direct it. Test it. Own it.

Learn the durable skills behind ChatGPT, Claude, Grok, Gemini, and coding agents—from your first useful prompt to a reusable, evaluated workflow that you can defend in an interview.

See the 7-level roadmap Choose my easiest start
  • No API key required
  • Tool-agnostic methods
  • Progress stays local
AI fluency stack Human judgment stays on top
  1. 06
    Evaluate agentsTests · traces · failure recovery
    Advanced
  2. 05
    Write reusable skillsTriggers · workflow · guardrails
    Builder
  3. 04
    Connect toolsStructured data · functions · APIs
    Engineer
  4. 03
    Code with evidencePlan · patch · test · review
    Practitioner
  5. 02
    Engineer contextFiles · examples · constraints
    Operator
  6. 01
    Specify outcomesIntent · rubric · format
    Starter
Core rule Never outsource the definition of “correct.”
Research-informed method Ask → inspect → test → explain → improve

Outcome validation is in progress.

Choose your easiest start

One curriculum. Eight useful paths.

Pick the work you actually need to do. We will start with a small win, then add independent practice, verification, and explanation.

Your four-session start

First-time AI user

Learn what AI can and cannot know, then complete one useful task without giving away private information.

    Continue into the full roadmap →

    Why tracks? “Easy” means reducing unnecessary load, not removing thinking. Each path begins with a worked example, fades support, and ends with independent recall.

    Seven-level curriculum

    From user to AI systems builder.

    Each level produces proof: a prompt, a verified artifact, a code change, a reusable skill, or an evaluated agent—not a passive completion badge.

    1. Level00

      Orientation · 20 minutes

      AI literacy and safe habits

      Understand models, chat history, hallucinations, privacy boundaries, and why confident output still needs evidence.

      • Model vs. product
      • Data boundaries
      • Verification ladder
      • When not to use AI
      Practice mission +

      Ask two tools the same factual question. Find the primary source, compare both answers, and write what evidence changed your confidence.

    2. Level01

      Foundation · 35 minutes

      Write a specification, not a wish

      Turn a vague request into an outcome with context, inputs, constraints, output format, examples, and a definition of done.

      • Outcome first
      • Relevant context
      • Constraints
      • Acceptance rubric
      Practice mission +

      Rewrite “make this code better” into a review request that names the file, audience, risks, allowed changes, required tests, and response format.

    3. Level02

      Operator · 45 minutes

      Engineer the context window

      Give the model the smallest trustworthy context that can solve the task: source files, schemas, examples, decisions, and explicit unknowns.

      • Context maps
      • Few-shot examples
      • Source grounding
      • Context compression
      Practice mission +

      Create a one-page context pack for a bug: expected behavior, actual behavior, reproduction, relevant code, logs, constraints, and facts that remain unknown.

    4. Level03

      Practitioner · 60 minutes

      Code with AI without losing the code

      Use AI as a pair programmer while you retain the mental model: inspect the repository, plan a small diff, run tests, review security, and explain every important decision.

      • Repository orientation
      • Small diffs
      • Test-first evidence
      • Diff review
      Practice mission +

      Ask AI to fix one bounded defect. Before accepting it, predict the changed files, inspect the diff, run targeted tests, add one missing edge case, and explain the fix without AI.

    5. Level04

      Engineer · 60 minutes

      Use structured outputs and tools

      Move beyond prose. Ask for schema-valid data, connect narrow tools, validate parameters, handle failures, and keep side effects behind human approval.

      • JSON schemas
      • Function calling
      • Tool permissions
      • Retry and fallback
      Practice mission +

      Design a research tool with a strict input schema, allowlisted domains, timeouts, citations, and a result validator. List what the model must never execute automatically.

    6. Level05

      Builder · 75 minutes

      Write a reusable AI skill

      Package expertise as a clear contract: when the skill should run, what inputs it needs, which workflow it follows, what it may change, and how it proves completion.

      • Trigger rules
      • Required context
      • Workflow steps
      • Safety boundaries
      • Validation
      Practice mission +

      Use the Skill Studio below to create a repository code-review skill. Test it on one real diff, record a false positive, then improve the instructions.

      Open Skill Studio →
    7. Level06

      Advanced · 90 minutes

      Build and evaluate agentic workflows

      Orchestrate planning, retrieval, tools, memory, and human checkpoints. Measure task success, cost, latency, safety, and recovery—not how impressive the demo looks.

      • Agent loops
      • State and memory
      • Evaluation sets
      • Tracing
      • Cost controls
      Practice mission +

      Build a ten-case evaluation set before building an agent. Include normal, ambiguous, adversarial, tool-failure, and “must refuse” cases. Track pass rate and cost per successful task.

    Roadmap practice0 of 7 levels practiced

    Why this learning loop

    Designed from evidence. Measured with humility.

    The research below supports specific learning mechanisms. It does not prove that LearnByShorts improves outcomes; that claim requires our own controlled, documented learner study.

    01 · Retrieve

    Answer before reveal

    Repeated retrieval can support delayed retention better than repeated study alone. Lessons ask learners to predict or attempt before seeing the answer.

    Karpicke & Roediger, 2008 ↗
    02 · Space

    Return after a delay

    The useful spacing interval depends on how long knowledge must be retained. Paths include next-day and later review rather than one long binge.

    Cepeda et al., 2008 ↗
    03 · Scaffold

    Example, completion, independence

    Worked examples can reduce unnecessary problem-solving load for novices. Support should then fade so the learner performs the work.

    Sweller & Cooper, 1985 ↗
    04 · Explain

    Make the mental model visible

    Self-explanation has been associated with deeper understanding of worked examples. Every artifact should be followed by “why this is correct.”

    Chi et al., 1989 ↗
    05 · Guardrail AI

    Protect productive struggle

    A large field experiment in high-school mathematics found that unrestricted answer access improved practice performance but could harm later unassisted performance; learning-oriented safeguards mitigated the effect.

    Bastani et al., 2025 ↗
    06 · Evaluate

    Test the learning design itself

    An AI-tutor trial reported strong results in one college course using a structured, research-based design. We treat it as promising evidence—not a universal guarantee.

    Kestin et al., 2025 ↗
    Claim standard

    Today: “research-informed learning design.” After a credible LearnByShorts study: report the method, sample, outcomes, uncertainty, and limitations—whether results are positive or negative.

    Interactive Skill Studio

    Turn expertise into a dependable instruction.

    A good skill is not “act like an expert.” It is an executable operating procedure with inputs, boundaries, checkpoints, and proof.

    Start from an example:
    Generated skill contract

    Untitled skill

    # Untitled skill

Complete the fields to generate a reusable skill contract.
    Contract strength Start with the trigger and outcome.

    The coding confidence gate

    Can you trust the AI-generated change?

    Select the evidence you would require before merging an unfamiliar AI-generated patch.

    0 / 5 gates Do not merge yet.

    AI speed does not replace engineering evidence.

    Failure-case coverage

    Practice more than the happy path.

    No finite course covers every future failure. This matrix covers the recurring categories learners must recognize, test, and recover from.

    Accuracy

    Confidently wrong answer

    Require a primary source, recompute the result, and record remaining uncertainty.

    Freshness

    Outdated information

    State the required date, verify publication and effective dates, and cite current sources.

    Privacy

    Sensitive data exposure

    Remove secrets and personal data before prompting; use approved enterprise boundaries.

    Security

    Prompt injection

    Treat retrieved content as untrusted data; keep system rules and tool permissions separate.

    Code

    Silent regression

    Inspect the complete diff, run focused tests, add an edge case, and plan rollback.

    Evidence

    Fabricated citation

    Open the source, confirm the claim is present, and remove citations you cannot verify.

    Fairness

    Biased recommendation

    Inspect assumptions, compare affected groups, add human review, and document limits.

    Control

    Over-automation

    Put approvals before external writes, messages, purchases, permissions, and deletion.

    Reliability

    Tool or API failure

    Use timeouts, bounded retries, idempotency, partial results, and a safe fallback.

    Cost

    Runaway agent loop

    Set token, time, and tool-call budgets; stop when progress or evidence stalls.

    Intent

    Ambiguous request

    Separate known facts, assumptions, and the one decision that needs clarification.

    Learning

    AI dependency

    Attempt first, request a hint before a solution, then repeat independently without AI.

    Multi-model fluency

    Learn principles once. Test tools honestly.

    Features and model names change quickly. Avoid permanent “best model” rankings; choose with a small evaluation set built from your actual work.

    O

    ChatGPT + OpenAI

    Prompting, tools, coding, and agents

    Practice clear instructions, structured outputs, tool use, coding workflows, and agent evaluation with official platform examples.

    • Run one task in chat
    • Repeat it with a structured output
    • Add a tool and an evaluation
    OpenAI Cookbook ↗
    A

    Claude + Anthropic

    Success criteria, context, and tool use

    Start with measurable success criteria, build strong context, compare prompt variants, and test tool-using workflows.

    • Define the evaluation first
    • Compare direct and example-led prompts
    • Record failures, not vibes
    Anthropic Courses ↗
    X

    Grok + xAI

    Reasoning, search, and API workflows

    Learn the API shape, prompt guidance, reasoning controls, and grounded tool patterns through the official xAI examples.

    • Reproduce a quickstart
    • Require cited, inspectable sources
    • Measure accuracy and latency
    xAI Cookbook ↗
    G

    Gemini + Google

    Prompt design and multimodal context

    Practice precise instructions, examples, iteration, long context, and multimodal inputs using official Gemini guidance.

    • Turn media into structured evidence
    • Check grounding and omissions
    • Compare against text-only input
    Gemini Prompt Guide ↗

    Interview-ready AI use

    Show judgment, not dependency.

    Some interviews allow AI tools and others assess independent thinking. Confirm the rules. When AI is allowed, make your intent, verification, privacy decisions, and final ownership observable.

    1. 01
      State the policy

      Confirm what tools and data are allowed before beginning.

    2. 02
      Narrate intent

      Explain why you are using AI for this step and what you expect back.

    3. 03
      Challenge the output

      Test assumptions, edge cases, security, and alternatives.

    4. 04
      Own the answer

      Be able to continue, debug, and explain without the tool.

    Open Modern Interview Lab →

    Official source library

    Learn from maintained examples.

    These are starting points, not copied courses. Rebuild the examples, test them on your own tasks, document failures, and cite the source you learned from.

    OpenAIOpenAI CookbookExamples and guides for the OpenAI API ↗ OpenAIAgents SDK for PythonAgent workflows, tools, handoffs, tracing ↗ AnthropicAnthropic CoursesPrompting, evaluation, tool use ↗ AnthropicAnthropic CookbookPractical Claude API patterns ↗ xAIxAI CookbookGrok conversations, reasoning, and media ↗ MicrosoftGenerative AI for BeginnersOpen-source multi-lesson curriculum ↗ GuideOpenAI Prompt EngineeringOfficial instruction patterns ↗ GuideAnthropic Prompt EngineeringSuccess criteria and prompt techniques ↗ GuidexAI Prompt EngineeringOfficial Grok prompting practices ↗

    External documentation changes. Verify model names, pricing, limits, and features on the provider’s current official documentation before building.

    Your next proof of skill

    Build one workflow you can explain.

    Create a reusable skill →