Tutorials, comparisons, and frameworks to write better prompts for ChatGPT, Claude, Gemini, Midjourney, Veo3 and Kling.
Page 21 of 27
OpenAI removes sora-2, sora-2-pro and the Videos API on September 24, 2026. Here is the prompt-syntax conversion to Veo 3.1, Kling 3.0 and Runway Gen-4.5, plus what does not carry over.
Read moreChatGPT forgets for five different reasons: a full context window, memory, project instructions, an empty new chat, or summarization. One fix each, with OpenAI's published limits.
Read moreChatGPT gives short answers when a setting, a stored instruction, or your own prompt lowers the default detail level. Five causes, a test for each, and the fix that adds length.
Read moreChatGPT apologises because preference training rewards agreeable answers. The mechanism, the four apology failure modes, and the instruction lines that reduce each one.
Read moreChatGPT repeats itself for four different reasons and only one is a sampling problem. How to tell them apart, and how to fix each from the prompt with no API access.
Read moreWhy AI ignores your word count: no running counter while it generates, just a soft style cue. The token math, verified at the source, plus four structural fixes that actually work.
Read moreChecked against Midjourney's own docs on August 26, 2026: --ar caps at 14:1 on V6, V7 and V8.2, but only 4:1 in HD. Syntax, every ratio, published pixel sizes, upscaling.
Read moreChatGPT slow today? Seven separate causes: long threads, reasoning effort, model and plan tier, live incidents, blocked websockets, extensions, and region. Plus a five-minute test.
Read moreImagen 4's Gemini API endpoints shut down on August 17, 2026. What replaced them, which Nano Banana model to pick, and a prompt-by-prompt conversion table.
Read moreHow to direct sound effects, ambience, score and silence in Veo 3.1, what Google actually documents, and why Gemini Omni Flash is now the default video model.
Read moreMCP server not connecting? 15 symptom-cause-fix pairs for Claude Desktop, Claude Code, Cursor and Codex, with the exact config path and log file for each client. Verified August 2026.
Read moreAI defaults to bullets from training bias, markdown-rendering chat interfaces, and stale custom instructions. The exact line that turns it off, and when lists are still right.
Read moreAI disclaimers come from safety tuning and topic classifiers, not law. Why ChatGPT and Claude hedge, which warnings are load-bearing, and instructions that cut the rest.
Read morePaste one instruction line and ChatGPT interviews you before it answers. The exact wording, where to make it permanent, and when asking questions makes the output worse.
Read moreAI rewrites untouched code for two separate reasons: no scope boundary, and a prompt that forces a full-file rewrite instead of a diff. The fix, and a reusable template, for each.
Read moreWhat xAI actually documents about Grok Imagine Video 1.5: modes, limits, the prompt upsampler, reference tags, and the parameters that do not exist. Checked at docs.x.ai.
Read moreThe model stops early because it decided the task was done, not because it hit a limit. Five distinct causes, five different fixes, and one habit that heads off all of them.
Read moreGarbled AI image text is almost always the model, not the prompt. Which models actually render text well in 2026, verified at vendor sources, and what prompting can still fix.
Read moreEvery Midjourney --v and --niji value in one place: which versions still render in 2026, what each one changed, and the parameter compatibility matrix across V6, V7 and V8.2.
Read moreAI images get extra fingers because models learn hands as a blurry pattern, not a fixed count. The 2026 state per model, what negative prompts actually fix, and the inpainting method that works.
Read moreA copy-paste Kling prompt generator that turns one line into a timed multi-shot list for Kling 3.0, with verified parameters, four worked examples and the specs Kling does not publish.
Read moreThere is no correct prompt length. What decides it: task complexity, missing context, whether examples help, and reuse. Plus why longer prompts can make output worse.
Read moreWhat Anthropic's docs actually say about XML tags in Claude prompts: which tags to use, how nesting works, how they sit with the system parameter, and why they are not a security boundary.
Read moreLuma Ray 3.2 prompting checked against Luma's own docs on 26 August 2026: 5s/10s clips, 360p to 1080p, six ratios, no native audio, and where Luma's numbers disagree.
Read morePage 21 of 27