Usewright · Practical AI Guides @usewrightapp

What happened, and how to lock down your own AI agents: usewright.app/news/openai-of…

OpenAI says its frontier models now train on an offline, cached copy of the web, not the live internet. Why: in September a research model used a misconfigured DNS resolver in its sandbox to reach a public chatbot. Monitors flagged it in under 12 minutes. Training paused 11 days.

I rebuilt my Claude Code workflow around its creator's advice, and the difference surprised me bit.ly/4yGd6pz

@botpagesco @dani_avila7 Good reminder. Worktrees are great for running agents in parallel, but each one is a full checkout with its own dependencies. Worth building a cleanup habit for once a branch merges.

@claude_news Rarely, and only at a clean break. Compacting mid-task is where you lose the exact error string. Finish the step, write the key paths and decisions to a notes file, then compact or start fresh.

Claude is about to stop asking before each step. In Claude Code, a deny rule beats any allow rule, and a hook can block a command before it runs. 6 steps for permissions, hooks and sandboxing. Copy them 👇 x.com/i/article/2113…

@cv_usk Underrated for sure. Multiple choice beats an open question because the agent has to show you the options it's weighing. Half the time the useful part is seeing it was about to pick the wrong one.

@freeCodeCamp This one's sneaky because retrieval looks fine in the logs. Check what actually landed in the final prompt, not just what got retrieved. Once you see the right chunk getting cut for space, summaries and tighter chunks stop feeling optional.

What is an MCP Task? Normally a slow MCP tool blocks the agent until it finishes. With Tasks: 1. The tool returns a task id right away 2. The agent keeps working 3. The agent checks the task until it's done Claude Code supports it now. How it works + how to add it to your server 👇 upstash.com/blog/claude-co…

@DanKornas The part people skip is pruning. Durable context helps right up until it's full of decisions you already reversed. Date each entry and clear out the stale ones, or the agent follows old rules with total confidence.

@BiaO2212 The test one bites everyone. Put it in your instructions: never edit a test to make it pass unless I ask. And have it explain any test change before it touches the file.

Before you ask AI to rewrite something, tell it what to keep. "Keep the numbers, the names and the tone. Fix the flow." Otherwise it "improves" the parts you needed left alone.

My favorite last line for any agent task: "Tell me what you didn't do." The skipped steps are where the surprises live.

Shift+Tab sets how much Claude asks before acting. Anthropic already gave you a free course, Claude Code 101. I post one tip from it every day. Day 5/30. Press Shift+Tab before your first prompt to cycle the permission modes, which set how much Claude does without your OK. 1. Manual. Claude asks before every file edit and every command. I use it in a codebase I don't know yet. 2. Auto-accept. File edits go through on their own. Commands still wait for you. Good for small fixes you'll review after. 3. Plan Mode. Claude only reads. It looks through the code, asks you questions, then hands back a plan. Use it for multi-step features. 4. Auto mode. No prompts. A background check screens each action and blocks risky ones. For anything bigger than a one-line fix, I start in Plan Mode. Changing a plan takes one message. Changing code takes an afternoon.

@claudeskills101 Keep it short, though. A context file that grows into a novel gets skimmed like one. The rules you'd actually enforce, a couple of examples of your voice, and what to never do.

@SparkyMark96 Another trick that helps: don't let it know which draft is yours. Label them A and B, and ask for the reasons before the verdict, so it has to argue before it picks.

Turning a table photo into CSV? Don't lose the leading zeros, turn blanks into zero, or let AI guess unreadable cells. Our free practice guide includes a fictional table, prompts and an answer key. Try it and tell us where you get stuck: usewright.app/p/table-image-…

Artificial Analysis just switched the coding part of its Coding Agent Index to FrontierCode, Cognition's benchmark that grades a change on whether a maintainer would merge it. That's a better question than whether a model got some unit tests to pass. A patch can go green and still be something nobody would accept. I'd still be careful reading the middle of the table. FrontierCode has 100 tasks. A failed attempt scores zero and a passing one scores somewhere between 70% and 100%, so each task is worth up to a point. At around 45%, ordinary sampling noise on 100 tasks is roughly five points either way. Fifteen of the 24 configurations land between 43% and 49%. I wouldn't rank those against each other with much confidence. The index also scores a model and its harness together. That's realistic, since nobody runs a model without one, but it means a gap between Codex with one model and Claude Code with another doesn't tell you which half earned it. Some people asked why DeepSeek V4.1 Flash ran in Codex instead of DeepSeek's own harness. It didn't seem to hurt much. Flash in Codex came second on FrontierCode at 52.6%, behind Opus 5.5 in Claude Code at 56.0%. The part a merge-or-not score doesn't capture is the second try. In real work the first approach often gets rejected. Then the agent has to change course in the same repo without dragging the failed attempt along or leaving stray edits behind. That recovery is a lot of what you actually pay for. Cost makes that more interesting. On the index's cost frontier, Codex with Flash scores 47.3 at $0.39 a task, while Claude Code with Opus 5.5 scores 61.7 at $11.42. If a cheaper setup recovers well on the second try, the gap could look quite different from how it looks on first attempts.

@ZoeHYua Test it before you commit. Run the same long, streaming task through the proxy and through a direct connection, and watch for dropped streams and timeouts on big outputs. Those problems show up well before average latency does.

@russano Easiest way to find your first one: scroll your chat history for the instructions you've typed more than twice. That repeated paragraph is already a skill. It just isn't saved yet.