Blog

ChatGPT is adding watermarks to text for EU users, and it's opt-in in the API. Who gets it, what it can and can't show, and how Claude and Gemini compare.

AI detectors guess from how predictable your text is, so plain, formal or heavily polished writing gets flagged. Why it happens on Turnitin and GPTZero, who gets flagged more, and what to do if you're accused.

Karpathy's ASD-STE100 tip, set up in Claude Code: one rule in CLAUDE.md or AGENTS.md that uses STE for explanations and your own voice for drafts, an output style if you want it on for every reply, a skill for rewriting one text. Why 80% is safer, test results, and where STE helps and hurts.

ChatGPT Space turns chats into shared Pages with comments and paragraph-level edits. Which writing problems in ChatGPT and Codex it fixes, when you don't need it, and what to use if you want Claude and Codex on the same draft.

Say what shipped and who it helps in the first line, then tell the one thing you learned building it. A four-line template, a real example, and what to cut.

AI LinkedIn posts sound alike because they start from a topic, not from something that happened to you. Give the model your specifics, pick a voice on purpose, and edit out the formula.

Claude Opus 5.5 writes shorter, clearer prose than Opus 5 and follows style rules better. Some AI tells remain. Here's what changed, plus prompts to copy.

A humanizer edits the AI sound out of a finished draft. A model trained to write like a person starts closer. What each one fixes, from our tests, and when to use both.

The top of the writing leaderboards is open models too big for any desk. Eleven open-weight models sorted by what they are built for, what is independently measured, and what fits on one GPU or a Mac.

Claude is good at sources, structure and long drafts. Hemmingway-1 is trained to sound like a person. How to split one article between them without losing your facts.

Is Hemmingway-1 better than Claude or ChatGPT at writing? The makers' test scores, what they leave out, and five prompts to run on all three yourself.

Yes. Download a 17 GB Q4_K_M file and run it with llama.cpp or LM Studio, or use an MLX build on a Mac. Setup steps, memory needs, settings and fixes.

Hemmingway-1 has an OpenAI-compatible API, so it should work wherever you can paste a base URL. It works in Obsidian and VS Code. Cursor is shaky, and Codex and Claude Code can't use it.

Hemmingway-1 is a small open model that only writes. What its makers' benchmark claims rest on, where it loses, what it costs, and three places to use it.

Install blader/humanizer, paste a draft, and add a few paragraphs of your own writing. Then compare before and after to check your facts and opinions survived.

For fiction, poetry and roleplay, base models write livelier prose than the chat models built on them. Here is why, and the few places that still host one in 2026.

Thomas Ptacek's two rules for writing with an LLM, plus what only appears in the Hacker News thread: his copyediting system prompt, how many suggestions he takes, the strongest objection, and when the rules apply.

Founder-led SEO means writing your product's content yourself, with AI helping with research and drafting. Choose a reader's question, check the answer, and keep your sources for the next article.

AI drafts read as empty because none of your thinking is in them. Interview yourself through the agent, edit out the AI habits, then turn your own edits into writing preferences you can reuse.

Understand AI writing traces, preserve your voice, and compare Humanizer with similar open-source skills for editing blogs, newsletters and product announcements.

Compare Humanizer and Stop Slop for English AI writing, with the same draft, a no-skill control, and a writing-sample test. See what actually changed.

We fed the 43k-star skill a paragraph built to trip its own pattern list. It flagged 7 of its 35 patterns, cut 67 words to 45, and kept every fact.

Yes, in the ChatGPT desktop app: Work and Codex read and edit files in folders you open. On the web and phone, ChatGPT only sees files you upload or connect.

Yes. Uploads give Claude the text of a Word file, not its images. Claude for Word also reads comments and tracked changes. Claude Code needs a skill or pandoc.

For plain chat, the ChatGPT app and chatgpt.com are the same. The app adds folder access, Computer Use, Appshots and local Codex. When the browser is enough.

Claude Desktop includes Claude Code. Use Cowork to hand off file tasks that now run in the cloud, and Claude Code to work on your computer and review each edit.

Claude Desktop works in folders on your computer, runs local MCP tools and uses your apps. For chat, uploads and cloud connectors, claude.ai is enough.

Find old notes with Perplexity Computer on Mac. Learn what local mode supports, what needs internet, and how to turn source passages into a draft.

The Gemini web and mobile apps only read files you upload. The Mac app can open and edit real folders through Gemini Spark, and Antigravity CLI works on files in a folder you choose. What each one can do, the limits, and how to set it up.

Claude described my 943-page PDF after reading just 4 pages. Ask what it read, check its tool calls, and prepare the full document for repeated use.

Upload a few files or connect a folder through Cowork. Understand the search gaps, then use StashBase to prepare documents and scans as context for Claude.

You don't need to upload PDFs one at a time to a chat site. Index the folder they already live in — scans included — and ask questions across all of them from Claude, Codex, or any MCP client.

Claude Desktop reads and edits folders on your computer through Cowork, included in every paid plan. How to connect a folder, what the permission modes mean, and what to add when plain folder access stops being enough.

Codex reads and writes in the folder you launch it in. How to add more folders, give Codex full access (and when not to), and reference a specific file with @.

Cursor's agent reads only what's inside your open workspace. Three real fixes: add a folder to your workspace (multi-root, saved as a .code-workspace file), open a parent folder, or connect an MCP server that searches the folder for you.

Upload-a-PDF tools break down at forty papers. Keep the library as a local folder instead: extracted text, hybrid search, an agent that cites which paper a claim comes from.

How to give Claude access to a folder with Cowork — plus Projects uploads, Claude Code --add-dir, filesystem MCP servers, and an indexed library. What each method actually does, and when each is the right choice.

Find files by what they mean, not what they're named: how semantic search over a local folder works, what stays on your machine, and where keyword search still wins.

Google Docs now edits .md files without converting them. Sync Claude's folder with Google Drive, and your team edits the same files Claude reads.

Notion has agents now, on its Business plan and its own credits. If you want to write with the Claude Code or Codex plan you already pay for, compare Palette, StashBase and Obsidian.

One command and a Notion sign-in connect Claude Code to your workspace through MCP. What it can do there, where it falls short, and why long pages use so many tokens.

Claude Code can write in Notion through MCP, but it edits a Markdown copy of the page. When to keep Notion, when to draft in files, and how to use both.

Can you use Claude Code or Codex? Sync between computers and phones? Compare five Obsidian AI options by agent support, devices, sync, and cost.

Both make you upload documents into their own store — the real choice is whose machine and whose model. A fact-checked comparison as of September 2026, and what neither one does.

Karpathy's LLM Wiki pattern on your own folder: one prompt builds an index and topic pages that link back to your files. A second brain in plain Markdown, with or without Obsidian.

The agent-maintained second brain, Codex edition — point OpenAI's coding agent at the folder where your notes already live, give it a wiki contract in AGENTS.md, and know where grep runs out.

A local AI knowledge base is files on your machine plus an index an AI can search. The three architecture decisions that matter more than the tool list — and which tool fits each answer. Checked September 2026.

Every open-source, self-hosted, or offline NotebookLM (now Gemini Notebook) alternative still worth running in September 2026 — checked repo by repo — plus the graveyard nobody mentions.

Measured, categorized repositories where the knowledge lives inside the repo — books, courses, guides, references, and notes worth keeping locally.

Explore five ways to use Claude Code Projects, from feature launches and competitor tracking to newsletters, user feedback, and webinars.

Claude subagents return results to one session. Agent teams let agents message and challenge each other. When each fits, and what to check first.

Custom GPTs stop working on December 11, 2026. Whether your GPT becomes a Claude Skill or a Claude Project, and how to move its instructions and files over.

Claude Code mods add panes, prompt rewrites and safety checks. Fifteen worth a look, grouped by what they help with, plus how to install one, check it before it runs, and ask Claude to write your own.

Gemini Gems become Skills on November 17, 2026. Why the Claude equivalent is now a Claude Skill, when a Claude Project still fits better, and how to move a Gem over.

People are letting Meta's Muse call the cable company and cut their bills. What it does, the cases people posted, where it gets blocked, and how to write the message yourself when it does.

Claude can't generate images, but Claude Code can call a model that does. One command and a free Hugging Face account connect it to Qwen-Image-2.1. No GPU needed.

A post on X says Qwen-Image-2.1 was uncensored overnight. The files it points to are the original weights, unmodified. What the facts are, and two ways to run the model on your own computer.

Jev can't write a sentence, so what is it for? Eighteen projects from its first week, each explained by the idea behind it, with the repository when there is one.

Yes for cloning, dubbing, transcription, and audiobooks on your own GPU; no for real-time agents, fine-tuned clones, and the voice library. A fact table from both products' own docs, checked September 8, 2026, plus the license catch that decides commercial use.