Experiences & Lessons
What went wrong, what to avoid and what actually worked. Written after the fact, from real use, not from the docs.·Page 1 of 2
Meta Muse needs you to personally confirm every public action, like posting to Instagram or replying to comments, and there's no setting to loosen that. The easier approach is to have it write a batch of posts and stop before publishing, then approve them all at once.
Seedance 2.5 looks the best, but 10 seconds costs about $1, so a two-minute video easily runs over $14. To save money we switched to H3 on jqsp, and the same explainer with a host costs under $0.50.
Reinstalling is easy. What's hard to get back on a new computer is your MCP servers, Skills, rules files, env vars and project dependencies. Ask your AI agent to inventory them, set up a private backup repo, list your keys, and write a restore script. Copy-paste prompt included.
Let Meta Muse find recent viral TikToks in your niche, sort them into patterns, and break down their scripts. Then swap in your own product and generate a remake, with video generation free for now. Includes the prompt for studying viral videos.
The open source reverse engineering tool REA lets an AI agent take apart an app without its source code. Point it at a niche Android app and have it map only the user roles, flows, core data, and pricing points. Then build your own version from that brief. Includes a ready prompt and a note on install steps and infringement risk.
Over two days I moved my Gemini subscription from Ultra down to Pro, then back up to the top tier because of Gemini 4 Carbon. What hit me hardest was how it writes copy without the AI feel. Includes a quick test for deciding whether to upgrade.
Working with Fable 5.2, the session gets compacted into a summary after almost every task, and the app shows how many tokens each compaction saved. Put rules in files and give each task in one message so nothing gets lost.
Subscription quota often beats pay-per-use API calls by several times. Give your own non-urgent, one-off back-office jobs to your AI agent, and use the API only for features that serve users live at high volume. Includes four questions to decide and the limits you must not cross.
Learning something new used to mean getting through a steep entry phase first. Now you can have an Agent take you straight to a result. Fixed steps like commands and configuration can be written into a Skill and handed off, so you don't have to memorize them. But checking the result is something you always have to do with your own head.
The main session uses Sonnet, subagents go to Haiku, and Opus serves as an advisor that only reviews plans, checks repeated errors, and does final checks. The --advisor / --subagents startup flags circulating online throw errors in the current version, so this article gives working alternatives.
Agents get less obedient the more you use them, usually because AGENTS.md has piled up too much. Write only what the agent can't see for itself and that really affects results, prune it regularly, and include a cleanup prompt.
A prompt I have refined over a long period on ChatGPT and Gemini. It makes the model give the answer first, attach conditions to its judgments, search first for time-sensitive information, and end with a next step. Includes a simplified full text and usage instructions.
When the base-tier Claude quota runs out, it's usually because bulk output work was also handed to it. Keep judgment tasks for Claude, send bulk generation to Codex or a cheaper tool, and write the division of labor into CLAUDE.md.
A breakdown by use of what Google AI Ultra bundles: Colab compute, monthly Cloud credits, Flow image and video quotas, Antigravity coding quotas, 30TB of storage and various memberships, plus how the quotas are calculated separately. The list comes from materials provided by users; the official plan page is the reference.
A list of the models we currently find most useful, by use case: general, coding, complex reasoning, deep research, web search, long documents, slides, image drawing and editing, AI video, digital humans, voice and music, value for money, and open weights. Personal experience; models change fast, for reference only.
Strata has to support NVIDIA, AMD, Intel, Apple and many other kinds of hardware, so it can only take a general-purpose approach and can't squeeze out every bit of performance for one particular GPU generation. We think a version aimed specifically at the 40 and 30 series may come later and could be faster; you can already use Strata on 30 and 40 series cards today.
A learner asked how to keep faces consistent in multi-person face swaps. The group's conclusions: swapping multiple people at once in wan3 rarely succeeds, so try Kling; viral videos mostly feature 1 to 3 people, and 1 to 2 is the most stable; if quality is poor, switch models, adjust prompts, provide context, and supply reference images; cheap models get what you pay for. On compliance, non-consensual intimate deepfakes are illegal, payment processors will ban accounts, so only use faces you have permission to use.
Zoho International has a CLI, so an Agent can easily take over sending and receiving mail. Zoho China has no CLI and only a limited API. The workaround: CC the mail to Gmail, have the Agent read it with gws and write a draft, then you review it and send it to the client yourself.
X's built-in trending tab has few entries and monotonous content, and it is personalized to your interests, so it does not show the real network-wide hot topics. For occasional checks, just ask Grok directly; for high-volume needs, set up a scheduled task with the Grok bot and pair it with a subscription.
AI is advancing quickly in CAD work: reading drawings, producing drawings, revising drawings and annotation have largely been solved. We expect repetitive drafting roles to be replaced first. What remains are people who understand the industry, know how to use AI, and can communicate with clients.