Learn AI · Practical skills
Learn one useful AI skill at a time.
Short lessons for writing, research, images, and everyday work. Copy a prompt, follow the steps, and check the output.
The practical library
What do you want to do better?
53 lessons to explore
Updated
Veo 3.1: Write a Short Dialogue Beat That Has Time to Finish
Plan one speaker, one short line and a little silence before generating. Check audible wording, timing and unwanted voices rather than judging only the picture.
You’ll be able to: Create and review a short dialogue beat with explicit speech and silence windows.
Updated
Veo 3.1: Bridge Two Keyframes Without a Continuity Jump
Prepare compatible start and end frames before generating a short transition. Check object identity, motion direction and invented middle frames after rendering.
You’ll be able to: Prepare a two-frame transition and inspect its middle frames for continuity errors.
Updated
Nano Banana 2.1: Make Four Icons Follow One Style Anchor
Define a shared visual contract before generating an icon set. Compare silhouette, padding and stroke treatment at the size readers will actually see.
You’ll be able to: Judge a four-icon concept set against consistent visual rules and small-size readability.
Updated
Nano Banana 2.1: Keep One Product Consistent Across Angles
Use a product reference and a fixed detail checklist for multiple views. Reject invented seams, changed handles and misleading hidden surfaces before sharing concept images.
You’ll be able to: Review multiple product-angle concepts using the same visible-detail contract.
Updated
AI JSON: Separate Refusal, Missing Data and Incomplete Output
Design separate handling for refusals, incomplete responses and valid records with missing facts. Practise response-state decisions before passing model output downstream.
You’ll be able to: Choose a safe next step for each response state before validating and using its data.
Updated
Sol to Astra: Escalate Ambiguous Rows, Not the Whole File
Separate clear rows from unresolved ones before asking a stronger model for help. Keep the source evidence attached and measure useful outcomes instead of assuming savings.
You’ll be able to: Route unresolved records with enough evidence while keeping clear records unchanged.
Updated
Sonnet 5.5: Turn a UI Mockup into a Keyboard Test Plan
Use a mockup to draft focus-order checks, then test them in the real interface. Keep visible layout, proposed behavior and observed keyboard results separate.
You’ll be able to: Draft a focus-order test and record actual keyboard results without claiming a screenshot proves accessibility.
Updated
Claude Opus 5.5: Review an AI Draft in a Fresh Chat
Give a fresh reviewer the source facts and draft without the earlier persuasive discussion. Use a fixed rubric to catch unsupported claims and preserve uncertainty.
You’ll be able to: Find unsupported changes using a source-to-draft table in a separate review context.
Updated
Claude Fable: Check Every Requested Deliverable Is Present
Give each requested file or text asset an acceptance row. Use a completion manifest to spot missing outputs, blocked work and unsupported “done” claims.
You’ll be able to: Match every requested deliverable to a real output or an explicit blocked status.
Updated
Claude CSV Cleanup: Keep Leading Zeros and Text IDs
Tell Claude which columns are identifiers before cleaning a CSV. Check leading zeros, scientific-looking IDs and spreadsheet import types before accepting the file.
You’ll be able to: Preserve identifier strings exactly while normalizing only permitted numeric columns.
Updated
Make Claude Fable 5.1 Report Useful Milestones
Ask Claude for brief milestone updates during a longer task. Separate a prompt instruction from the API display configuration that makes updates visible to users.
You’ll be able to: Request informative milestone notes and distinguish absent updates from hidden interface output.
Updated
Claude Fable 5.1: Recover Context from a Verified Checkpoint
Keep confirmed decisions and unresolved questions in a short checkpoint. Re-anchor a long Claude task after context loss without treating an AI summary as the source of truth.
You’ll be able to: Restore a task’s confirmed state while preserving unresolved decisions after a context change.
A useful AI habit starts with one idea
Get new practical lessons and learning resources when we publish them.