What X is shipping with agents right now.

CyrilXBT@cyrilXBT·4h

A copy-paste settings.json plus two subagent files that split work by model: Opus plans, Sonnet codes, a Haiku scout searches files, and an Opus critic attacks the diff.

i can't comprehend why people run claude code on default settings one settings change and every model gets its own job: → opus plans every task and advises when stuck → sonnet 5.5 writes the actual code → haiku handles every file search → a critic agent tries to break your diff before you merge Opus thinks. The cheap models grind. paste this into claude code 👇 Add this to ~/.claude/settings.json. Keep every hook, permission and plugin I already have. Then create the two agent files below and show me the diff. ```json { "model": "opusplan", "advisorModel": "opus", "effortLevel": "medium", "modelSettings": { "claude-opus-5-5": { "effortLevel": "high" }, "claude-sonnet-5-5": { "effortLevel": "medium" } }, "env": { "CLAUDE_CODE_SUBAGENT_MODEL": "haiku" } } ``` ~/.claude/agents/scout.md ``` --- name: scout description: finds code fast. returns file paths and line numbers only. model: haiku tools: Read, Grep, Glob --- Locate exactly what was asked. Return file:line. No fixes. No opinions. ``` ~/.claude/agents/critic.md ``` --- name: critic description: reviews a diff or plan before any task is called done model: opus tools: Read, Grep, Bash --- Assume this is broken. Find how it fails. One line per issue, worst first. ```

OpenAI Developers@OpenAIDevs·12h

OpenAI's playbook for pointing Codex at maintenance: it finds dead code and stale experiments, checks them against production telemetry, and drafts cleanup PRs.

Here’s how our engineers use Codex at OpenAI to dig through our codebase, find what needs attention, and chip away at the maintenance backlog: https://developers.openai.com/blog/maintaining-software-with-the-gpt-6-family

ClaudeDevs@ClaudeDevs·15h

The Claude Code team's own write-up on reviewing agent-written PRs at volume, with numbers (2x merged PRs, 40% fewer reverts) and steps to adopt it.

Claude writes about 90% of Claude Code's code. From June to September, merged PRs more than doubled while the share reverted fell by more than 40%. We wrote about how we changed code review to get there, and how to adopt it: https://claude.dev/blog/solving-code-review-at-anthropic/

lauren@poteto·18h

The trust pyramid: put most effort into codebase design and static checks, use skills and AGENTS.md prose sparingly, and save human review for the highest-leverage spots.

how do you trust your agents' work more? copy this post/link and give it to your @bot or agent! 1. codebase improvements: this is just good software engineering - better architecture, stronger types and invariants, clear boundaries. agents love to copy what exists and take the shortest path, so make the easy thing the right thing. categorically eliminate problems through code and you won't need to think about it again. 2. static analysis: lint rules, tests, CI, verification (eg TLA+). very token efficient ways to catch issues ahead of time. 3. prose: comments, rules, AGENTS.md, skills - anything that can't be enforced and requires agent judgment. use it minimally, but do use /create-verification-skill. 4. human review: the hardest to scale. at the rate agents land code, it's pretty much impossible to thoroughly review every PR, so spend that effort on the highest leverage point, designing an environment where agents do the right thing by default. this forms an inverse pyramid: you want most of your effort at the top, least at the bottom. codebase improvements are deterministic, making it more token efficient, and lets you sleep more soundly at night. install https://t.co/u9xIlN3eN0 and create a new Project: 1. look through my slack/chat messages and code review comments where i've given feedback on PRs, code, architecture 2. use /recall on my past transcripts to learn how i steer agents 3. /poteto-mode use /correct and /architect on past commits to find the most common pitfalls agents fall into, then propose a plan to rearchitect the code i own so it's more agent friendly. answer open questions with prototyping rather than defer to me, and back the plan with real data. 4. before you begin, use /sharpen + /in-your-own-words and confirm this prompt with me

Ibragim@ibragim_bad·20h

Your old ~/.claude and ~/.codex sessions are an eval set: your corrections become test cases for benchmarking models, harnesses and plugins.

I had 1,270 agent sessions sitting in ~/.claude, ~/.codex and ~/.pi that’s an eval set: my corrections to the agent are the test cases turned it into the my-mini-benchmark. now I test any model, harness or plugin on my own tasks guide + the skill 👇 https://t.co/oaRtwFJTxT

Daniel San@dani_avila7·1d

/export file_name dumps a Claude Code session to plain text: hand it to another agent to continue, or mine it to write custom skills.

Claude Code lets you export your entire session with /export file_name This creates a plain text file containing your session history You can use it as context for another agent to continue your work, or analyze it to create custom skills that improve your Claude Code workflow https://t.co/cMZG0J01s9

Ackerman@Yarilo7brigada·1d

Shrink CLAUDE.md to core rules and move procedures into skills that load on demand, with descriptions written as 'Use when...' triggers.

THIS IS A F*CKING GOLDEN JACKPOT Your 800-line CLAUDE.md is killing Claude. Every session starts with the model reading instructions on deploys, releases, migrations, API style and ten other things. None of that matters when the task is fix this typo. The context is clogged from the first second, and the rules that actually matter get lost in the noise. The idea: move every procedure into its own skill. A skill is a folder with a SKILL.md file, plus optional scripts and reference docs. At the start of a session, Claude only sees each skill’s name and a one-line description. It loads the full text only when the task actually calls for it. Here’s the split: - CLAUDE.md holds only what’s needed every time: stack, structure, 10–20 core rules - skills hold procedures: deploys, releases, migrations, reviews, onboarding into a module - the description decides when a skill fires, so write it as a trigger: Use when... - scripts inside a skill get executed, not read, so deterministic logic costs zero tokens The key trick is progressive loading. Twenty skills in context cost about as much as twenty lines of description. You can pack your team’s entire wiki into the project, and Claude opens the right page exactly when it’s needed. Bonus: any skill can also be invoked manually as a slash command, like /deploy or /release. The same file works both automatically and on demand. CLAUDE.md remembers the essentials. Skills know the details. Context stays clean. full tree CLAUDE.md: stack, project structure, core rules, nothing else .claude/skills/deploy/SKILL.md: deploy steps, checks, rollback .claude/skills/release/SKILL.md: versioning, changelog, tags .claude/skills/migrations/SKILL.md: how to write and verify migrations + a check script .claude/skills/code-review/SKILL.md: the team’s review checklist every skill’s description starts with - Use when... paste the tree and this prompt into Claude Code ↓ Rebuild my Claude Code setup around this tree: Read CLAUDE.md and all linked docs. Split the content into two parts: what’s needed for every task, and procedures for specific situations. Keep only the first part in CLAUDE.md. Target: under 100 lines. Move each procedure into .claude/skills/<name>/SKILL.md with name and description in the frontmatter. Start each description with “Use when...” and list concrete triggers. If a procedure has deterministic steps (checks, generation, validation), move them into scripts inside the skill folder and call them from there. Find procedures I repeat that aren’t written down anywhere, and propose them as new skills. Show me every change as a diff first. No edits until I say go.” ↳ https://t.co/wK9W7spslN

Kaiju@kaijucode·1d

A full Claude Code setup using only built-ins: CLAUDE.md, opusplan, Haiku search subagents, critic subagent, hooks, skills, worktrees, /schedule.

claude code basically runs my repo now, i just review and press enter... here's why: almost all of it is free config you set up once on day 1. here's my whole setup and what it costs: required (ships with claude code) > memory → CLAUDE.md > models → opusplan (opus plans, sonnet builds) > search → haiku subagents for every file lookup > review → a critic subagent before every merge > safety → permissions in settings.json > undo → plan mode + /rewind > context → /clear between unrelated tasks optional (add when you need it) > automation → hooks that format on every write > repeatable → skills for the prompts you retype > parallel → git worktrees > research → /deep-research > escalation → an opus advisor when you're stuck > scheduling → /schedule routines the only thing on this list you need to pay for is your claude plan. everything else is included with claude code.

Theo - t3.gg@theo·2d

Theo flags that Codex still omits the current model name from its system prompt, so the agent can't tell which model it is.

It is so stupid that Codex STILL does not include the current model in the system prompt. Your agent has no idea what it is beyond "based on GPT-6". I've complained about this for a year straight and they won't fix it for no good reason. https://t.co/PepE5yeaZH

dex@dexhorthy·2d

Dex Horthy on idle compaction: compacting before the cache expires beats paying the cache miss when you return.

this is a good idea - always confused why claude would ask me if i wanna compact after 24h to save tokens since the cache expired…like i was gonna gonna eat the cache miss on the input to compaction anyway, right!? Anyways good ship imo https://t.co/UUHZ6YKepa

lauren@poteto·2d

pstack's most durable part is its tiny 10-20 line principle skills; expect most skills to compress into short principles as models improve.

perhaps one of the most durable parts of pstack are its engineering principles: very concise skills (~10-20 loc) that distill down good practices into its core. as models get smarter, i expect most of pstack skills to compress into short principles or workflows (a generic list https://t.co/bDIpHhE1cM

Lydia Hallie ✨@lydiahallie·2d

Claude Code now compacts long idle sessions right before the ~1h prompt cache expires, so resuming reprocesses only the summary.

We recently added idle compaction to Claude Code! It now compacts long idle sessions right before the prompt cache expires (~1h). If you come back to it later, it only has to reprocess the summary instead of the entire conversation, which saves you some usage. https://t.co/SG3QyTNFwr

Matt Pocock@mattpocockuk·2d

You don't strictly need skills if your process is great, but skills make it far easier to actually follow that process every time.

Do you need agent skills if you're following a great process? Of course not It's just that skills make it easier to follow that process

lauren@poteto·3d

Delete most code comments: agents copy what exists, so one stale workaround comment spreads through the codebase like a weed.

ban and delete most if not all code comments! agents love copying what already exists. a seemingly harmless comment explaining some temporary workaround is one of the worst viruses you can have in your codebase. like a weed it'll spread quickly as a tool for agents to quickly

lauren@poteto·3d

New /in-your-own-words skill in pstack makes the agent restate your goal so you confirm it understood intent before it starts.

this is now a new skill /in-your-own-words in pstack 0.15.23! extremely useful to ensure that you and your agent are aligned on the goal and that it understands your intent clearly https://t.co/3dIimxHJEt