Everything published here, newest first. Essays, notes, and experiment write-ups.
Short answer: only on short files. Once a PDF crosses a few dozen pages, ChatGPT switches from reading the entire document to chunked File Search. Here is how to catch it in 10 seconds, and 5 ways to get full-page reads.
NotebookLM (now Gemini Notebook) decks look great but are hard to edit in PowerPoint. Here is the 6-step workflow I use to rebuild them as a native, editable .pptx.
Between strict company AI portals, Poe's 300-point cap, and Gemini or DeepSeek tabs, everyone I met in Hong Kong was juggling four apps. Here is the workspace I showed my friend, and why it replaces that patchwork.
Swapping words never fixed it. The AI tone lives in structure and stance.
Can ChatGPT use draw.io directly? I tested 5 AI draw.io diagram workflows: Mermaid import, draw.io Generate, Eraser, BioRender, and direct .drawio file output.
Against Claude Opus 5.5 and GPT-6.1 Sol. One got the math wrong, one forgot the handlebar, one added text nobody asked for.
Gemini Gems transition to Skills around Nov 17, and OpenAI Custom GPTs retire Dec 11, 2026. A practical comparison of Gems vs Skills, plugin migration, and where each workflow goes.
GPT Image 2.5 supports native 4K in the API, but the ChatGPT web app still hits a 1024px resolution limit. Here are 5 tested ways to get full 4K output.
I'm a working PM who hasn't picked an idea yet. This is the workspace, the three-round process, and the kill list I use to figure out what I could run alone in 2026.
Looking for a Claude text watermark detector or Gemini SynthID checker? No public tool can verify provider text watermarks yet. Here is what 5 popular checkers actually do.
Search first. Open the page. Summarize last, or never. Long PDFs on a cheap model: index, page jumps, one artifact.
Study mode is back, but it still can't remember what you learned last week. Here's the PLAN.md method I use to actually finish learning something with AI.
I built a careful personal prompt library. It still died. What travels now is role, residue, and a short ticket.
Gemini has quietly marked your text since 2024, and it has quirks Claude doesn’t. Five methods, one that actually improved the writing.
Manus data deletion, one-time restore, websites still down, and how to choose a cloud or local agent so your next eight months don't live only in a vendor sandbox.
Five methods people try. Only one improved the writing instead of just hiding the fingerprint.
OpenClaw, Hermes, DeepSeek Harness. Different names, same "local = private" illusion.
The problem is not finding another model. It is keeping research, decisions, drafts, and context connected as the work changes.
What I actually packed when the knowledge base went away. Not a ranking of bot stores.
Export is the handoff now, not where thinking starts.
You can generate a polished report in minutes. Can you explain how you got there?
Same dilemma, four models, anonymized peer review. Self-preference and four blind spots.
57% hide AI use from managers. Same people brag about workflows online.
AI can improve average quality while collapsing diversity. What mitigates it.
Family medical prep in one project. Patterns, not diagnoses.
Workflow first, tools second.
Persona and memory are the asset. The model is just the engine.
A living voice document built by rejecting bad output, not "write like X".
One real project is rarely one chat. Who holds the files between conversations.
On-brand creatives stop being a slot machine when strategy stays in a hub and generators only get a finished prompt.