AI Decode Wire
-
Can Google and AI Assistants Read Your Restaurant’s PDF Menu?
Google says it converts text PDFs into HTML, so a PDF menu can be found and used in AI answers. The catch is scanned menus and phones. A five-minute check.
-
How to Use AI on a Spreadsheet and Still Trust the Numbers
Make the AI calculate rather than read, then check its work against one number you already trust. A five-step routine, a prompt, and a worked example.
-
Why AI Gives Different Answers to the Same Question
AI picks each word by weighted chance, and the dial that used to calm that down is being removed. Why answers vary, and how to use the variation as a check.
-
Multimodal AI Explained: What Happens to the Photo You Upload
An AI model never sees your photo, only a copy cut into squares and often shrunk first. What that costs, what the vendors say it gets wrong, and what to do.
-
How to Summarise a Long Document With AI Without Losing the Point
AI summaries of long documents tend to drop what sits in the middle. Here are five prompts, each with a one-minute check, to find what the summary left out.
-
Six AI Myths That Refuse to Die
Six beliefs about AI that the vendors’ own documentation contradicts: hallucinations, context windows, fine-tuning, detectors, training and reasoning.
-
The Three Questions That Decide What AI Should Touch First
The usual advice points you at the work a script does better. Three questions that screen for what AI really helps with, plus the first month in order.
-
“Open Source” AI Usually Isn’t. Here’s the Three-Part Test.
Most models called open source only ship the weights. The three things the Open Source AI Definition asks for, and the Llama clauses that actually bite.
-
Nine AI Prompts for Work That Check Their Own Answers
Nine copy-and-adapt prompts for summaries, emails, policies, tables and meeting notes, each built so you can see at a glance when the answer has gone wrong.
-
What “Free” Actually Gets You on ChatGPT, Claude and Gemini
Listicles quote free-plan message caps. ChatGPT, Claude and Gemini publish none. What each free plan really includes, and where to see your own limit.
-
Fine-Tuning vs RAG vs Prompting: Which One You Actually Need
Want an AI that knows your business? You probably need a better prompt, not training. A symptom-to-fix table, plus OpenAI’s fine-tuning wind-down dates.
-
MCP Explained: What You’re Actually Connecting
An MCP connector links your AI app to another service through the Model Context Protocol. What it is, remote vs local, and what to check before you connect.