CCSwitch made the Dots entry vanish from Codex? cute.

Ask Codex to make a desktop shortcut. Not a fix, but at least you can get back in. Someone’s already filed an issue with CCSwitch.

Fits in VRAM is not the same as fun to use. Diffusers' sequential CPU offload saves memory by shuttling submodules between devices. The docs call it extremely slow. Put latency next to the memory chart, please. huggingface.co/docs/diffusers…

Midjourney's Style Reference carries colors, texture and lighting, not the same person. If the face keeps changing, I'd check the reference type before adding another paragraph of adjectives. Give the poor prompt a break. docs.midjourney.com/hc/en-us/artic…

Making tiny whoosh noises at your laptop is apparently a legitimate editing workflow. Firefly's video editor beta lets your voice guide SFX timing and intensity. I'd rather perform the beat than describe it in six adjectives. helpx.adobe.com/firefly/web/fi…

I don't want to regenerate a whole poster because the date changed. Ideogram's Layerize Text makes detected copy editable. Keep the composition, fix the words. That's the unglamorous feature I'd reach for.
ideogram.ai/features/text-…

For agent UIs, '12 files checked' beats a made-up 87%. MCP allows progress updates without a known total. Show completed work and the current step; save the percentage for when its denominator means something.
modelcontextprotocol.io/specification/…

A judge's agreement score needs a 'not judged' count beside it. Promptfoo's calibration guide excludes grader errors and missing verdicts from that calculation. Keep both counts: agreement alone doesn't tell you how much got judged.
promptfoo.dev/docs/guides/ll…

Caption editors: check word timestamps before swapping transcription models. OpenAI's timestamp_granularities still requires Whisper-1, due to shut down Feb 26, 2027. Make alignment a migration acceptance test.
developers.openai.com/api/docs/guide…

A 60fps deliverable needs a decision before the AI edit. Runway's Edit Studio downsamples inputs above 30fps. Plan the conversion before the Aleph 2.0 pass; don't discover the mismatch after approving the shot.
help.runwayml.com/hc/en-us/artic…

For a Stable Audio 3 inpaint that barely changes, try a wider mask before a louder prompt. The surrounding audio can dominate a tiny edit. Save the version you like, then narrow the region in later passes.
github.com/Stability-AI/s…

An agent that gets one attempt needs a first-try metric. pass@10 asks a different question: did any of ten tries work? Keep that number if retries are real, but don't use it to promise reliable one-shot answers.
anthropic.com/research/demys…

Maintaining a Gemini 3/4 tool loop on legacy Generate Content? Keep the returned response parts intact. Rebuilding history from just tool names and arguments drops thought signatures and can turn the next step into a 400.
ai.google.dev/gemini-api/doc…

readOnlyHint: true is easy to over-trust. MCP's annotations describe a tool; the runtime still has to enforce its limits. I'd rather see a clear permission boundary than a reassuring label on a tool with unrestricted access.
blog.modelcontextprotocol.io/posts/2026-03-…

Meshy's Remesh doesn't support T2 models or split parts. That changes the order of work: choose T2's face budget at generation, or remesh other models before splitting. 'We'll optimize it later' needs a plan.
docs.meshy.ai/en/webapp/guid…

HeyGen's reference-to-video reads only the first 5 seconds of each clip. Trim to the gesture you want before uploading. The good bit at 0:12 never makes it in.
developers.heygen.com/docs/models/he…

A glossary field belongs before the upload button. Granite Speech 5.0 2B can bias transcription with up to 2,000 keywords. For interview tools, that's a useful place for guest names and weird acronyms, before anyone has to fix the transcript.
huggingface.co/blog/ibm-grani…

In Anthropic's sample agent, a merge cut latency but skipped a delivery-date lookup. The failure appeared in only one of two runs. Keep your old passing cases and rerun them before calling the simpler design a win.
claude.com/resources/arti…

On OpenAI's Responses API (GPT-5.6+), 12h cache writes cost 2.5× ordinary input vs 1.25× for 30m. Reads cost the same. Check the gaps between prefix reuses before making 12h your default.
developers.openai.com/api/docs/guide…

OpenAI's Decisions API now accepts up to 400 questions per request. Useful for independent checks on one input. If B needs A's answer, the docs say to split the requests. Sketch the dependencies before filling that array.
developers.openai.com/api/docs/guide…

ElevenLabs' Instant Soundtracks puts generated sound effects on separate editor layers. Moving one late footstep while keeping the music is the kind of control AI video tools need.
elevenlabs.io/blog/introduci…