We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Write less prompt and more verification
Addy Osmani says most people over-prompt Claude. For Opus 5.5, give one prompt that states the goal and how Claude should verify it reached that goal. Start with /effort medium and adjust from there . If the goal is still fuzzy, ask Claude to interview you until it can write a spec, then start a fresh session to build from that spec . Claude can run unattended when it has a pass/fail check or a number to beat, with reports it can read. For long background tasks, /goal has a separate model check a condition after every turn .
His other recommendations:
-
Trim your instructions for each new model. Over 80% of the Claude Code system prompt was removed for Opus 5 and Fable 5 with no measurable drop on coding evals. Cut your own instructions, add back only what the model misses, and use
/doctor prompt-auditto flag instructions written for older models . - Subagents: ask for them only for large standalone jobs such as audits or migrations. On small jobs they add cost and latency, and they don't make the agent smarter by default .
-
Cloud sessions run in a VM with a fresh clone of your repo, so push your commits before starting one (
claude --cloud). Parallel sessions only pay off once a single session can finish without you. Projects, in beta on Pro and Max, track which threads are done, which PRs are ready and what needs your input . -
Review: on the Claude Code team, Claude reviews and fixes most changes at least once before a human looks.
/code-reviewruns the review in a subagent that hasn't seen the conversation .
In a talk with Sonar, Osmani said Claude does the first-pass review and has to try the change the way a user would: browser checks, end-to-end tasks, linting. Hooks rerun on every edit . Write your team's standards in Claude.md/agents.md. Humans then check the agent's evidence, such as videos or screenshots of the UI working . The tests he relies on to merge automatically are end-to-end, property and visual-verification tests .
Theo's new video turns the same idea into a hill-climbing recipe. For TS Rust, he told the model to run real projects against a pinned TypeScript version and keep going until the outputs matched exactly. Once they matched, he told it: "Do not stop until you're two times faster", with existing tests still passing . Some goals are hard to check. For a Skate clone, he had the agent screenshot its builds and send them to a panel of LLM judges for comparison with real footage. He calls this crude, and says it burned at least $800 in tokens on the skybox . His main point is to keep asking the agent what is blocking it. Moving builds to snapshotted VMs on Boat, a company he has invested in, cut gate time from up to 70 minutes to 353 seconds . He warns against doing this at API prices .
Automate the PR "last mile"
Brent Traut's Codex setup:
- Install a PR-babysitting skill that addresses feedback, fixes merge conflicts and CI failures, and alerts you only when it needs you. He points to the one in the openai/codex repo.
- Create a "🚢 Outbound" section in the ChatGPT app.
- Create a "🤖 PR Shepherd" Codex thread scheduled every 15 minutes. It runs the skill on each Outbound thread until the PR is green, merges it, then archives the thread .
After requesting review, he drops the thread into Outbound and walks away .
Gemini 4 Carbon: strong coding scores, limited access
Artificial Analysis reports that Carbon (high), running in Antigravity CLI, scores 59 on its Coding Agent Index. That is 3 points behind Claude Code with Opus 5.5 (max, 62). It ranks second on Terminal-Bench 4.0 (64%) and FrontierCode v1.1 (51%) . It is only available to selected users so far, with a 50% launch discount and no confirmed end date . Access starts with cyber defenders, and Logan Kilpatrick says Google is working to make it more widely available soon . Kilpatrick says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, unsupported claims by about 2× versus Argon, and overreach by 26% . Matthew Berman reports pricing of $2/$10 per million input/output tokens during an introductory period of unspecified length, then $4/$20 .
T3 Code: one MCP server controls a fleet of machines
Theo showed T3 Code starting 8 threads on different machines through one MCP server, driven from Grok Bot . It works over the internet, not through computer use, so it still runs with his laptop off . To try it, enable it under Labs on the latest nightly. Theo calls it "VERY early" . Also new: cmd+k → "triage" clones T3 Code's source for your exact version, gives your agent a debugging playbook, and files an issue if the bug is T3 Code's fault .
Building games that are fun
Simon Willison says Opus 5.5's sense of fun beats GPT-6 Astra's . His recipe is to start in plain chat with Build a mobile-friendly three.js game where... and switch to Codex or Claude Code once the proof of concept works . Two prompts did a lot of work. One prompt that included "Think hard about how to make sure it is both fun and beautiful" produced a game in one shot . "Make it a LOT more fun" turned a dull 2D RTS into Battle Drones .
Memory and scheduled agents (LangChain)
- LangChain says over 70% of Managed Deep Agents runs come from schedules, not from someone typing a message . Its blueprint is a cron server that takes a prompt plus the skills or permissions the run needs, adds random jitter so runs don't fire all at once, and gives agents a controlled way to create and edit their own schedules .
- OpenWiki is an open-source CLI that builds a wiki of your codebase so coding agents don't rediscover the architecture on every task. Each claim in the wiki is tied to a hash of its source lines, so a claim gets flagged as stale without a model call when those lines change . LangChain then added an evidence graph and tested it on cross-document questions. It reports 27% higher accuracy, 2× citation consistency and 58% lower cost per correct answer .
Short notes
- Kent C. Dodds ties each Kody feature flag to metrics. A "Rerank" flag showed 7.6× more executions and a 5× lower error rate, so he will enable it. The same system led him to delete another feature .
- Riley Brown uses a Claude Project as a coordinator. It sends work that needs his computer to local sessions and research and artifacts to cloud sessions, and he runs it from his phone . He keeps 30 apps and 20 skills in Composio so any agent platform can use them. Shared memory across platforms is still unsolved for him .
- OpenAI: eligible ChatGPT Pro and Business subscribers can claim up to $500/month in API credits .
- GitHub Octoverse: merged PRs tripled in under a year to 220M per month. Willison reads this as the effect of agentic engineering .
- Riley Brown uses Claude Projects as an orchestrator, with separate projects for apps and editor/designer work; it delegates to local Claude Code sessions or cloud sessions, which he can launch and monitor from his phone. From an app-specific project, he can describe UI changes in natural language on his phone; the project routes the work to a local session, which updates the app and can deploy the result.
- Brown says Claude Opus 5.5 is strongest for his front-end-focused app work; he uses Claude for creating apps and assets, while Codex handles computer use and business coordination. His voice workflow uses Dot to launch a Codex session on his Mac to inspect Native Note, extend a background change to list and calendar pages, and push it to production; in the update shown, the session was running and the production push was still pending.
- He centralizes 30 connected apps and 20 skills in Composio, including a custom Native Note tool created by an agent so connected agent platforms can reuse it rather than repeat integrations. Cross-platform memory remains an unresolved gap: he says agent memories do not persist when switching platforms, and his Supermemory-through-Composio setup did not update frequently enough for seamless handoffs.
- For Twitter research, he has Claude Code call Grokbot through a configured webhook skill; Grokbot’s Twitter access lets him delegate the research while continuing to work through his preferred agent platforms.
- For long-running coding work, make progress machine-checkable: run real projects against a pinned reference compiler, inject errors, compare exact outputs, and set a speed target while keeping existing tests. Theo reports TS Rust began working in a day and improved over the following weeks; after setting a 2× speed target, the agent iterated until it met it.
- For harder-to-verify outputs, use an explicit proxy: the Skate 3 Bench agent captured screenshots of specific states, compared them with Skate references through an LLM judge or panel, and iterated. The result was a playable browser-based 3JS skate clone, but Theo calls the screenshot approach crude and says it spent at least $800 in tokens trying to improve the skybox.
- Treat agent orchestration as repeated unblock work: provide the tools and compute the task needs, then keep asking what is slowing the agent down instead of reviewing every code change. For TS Rust, Theo used Boat VMs and snapshots to reuse a prebuilt environment across multiple workers; he reports build-gate time dropping from 40–70 minutes to 353 seconds and daily code throughput more than doubling, closer to tripling. He also says he invested in Boat, an important caveat to his endorsement.
- This is an extreme setup, not a default recipe: Theo advises against replicating it and says not to run these experiments at API prices.
- Google presented Gemini 4 Carbon as its strongest model, claiming gains in coding and agentic tasks; it cited an 82.2% BenchPro result for finding vulnerabilities.
- The presenter questioned benchmark representativeness and noted that models appeared inconsistently across comparisons, raising cherry-picking as a possibility rather than a conclusion.
- General release was still pending: Google said Carbon was coming soon to developers, businesses, and consumers, and referred to an initial cohort of cyber defenders; the presenter had not tried it and planned to test it when released.
- The quoted price matched Argon: $2 per million prefill tokens and $10 per million generated tokens.
-
For a low-friction game proof of concept, start in regular ChatGPT or Claude chat with
Build a mobile-friendly three.js game where...; Willison says a full coding agent is not needed at this stage. His games used HTML/JavaScript and static hosting, and he recommends moving to Codex or Claude Code once the project is past proof of concept. - Make the desired experience explicit in the prompt: for Tree Appreciator, he asked Claude chat for two algorithmically generated trees in a spot-the-difference game and to think hard about making it fun and beautiful; it produced a mobile-friendly game in one shot, with no follow-up prompts.
- When a prototype works but lacks appeal, iterate on the experience directly: after a workable-but-not-great 2D RTS made with Claude Fable 5, Willison prompted Opus 5.5 to “Make it a LOT more fun” and then “Port this game to three.js.” Battle Drones later supported 2–4-player multiplayer and model-versus-model play, where the game advances five seconds before pausing for models to inspect the state and issue commands.
- Plan to playtest complex generated games and correct state-related problems: Emberwick needed substantial playtesting and follow-up prompting because story changes could leave a character in the wrong location to complete a quest; the finished story takes about 15 minutes to play.
- In Willison’s comparisons, Opus 5.5 had a stronger sense of fun than GPT-6 Astra; a GPT-6 Pro version of Emberwick looked great to him but played less well.
-
Supabase’s CLI exposes an instance as a local
supabasefolder and virtual filesystem. The announced PG Delta library provides a declarative file representation of a Postgres database’s tables, policies, and comments, and works with any Postgres database; the aim is for agents to read schema files rather than reconstruct the database by replaying large migration histories. - Supabase Compute was presented as long-running serverless container compute on the database’s private network, supporting Rust, Python, agents, and direct Postgres connections. The team described it as the missing primitive for durable workflows, which community members could not run on Edge Functions because of their timeout.
- Lovable’s database-specific agent skills prevented mistakes such as putting image data in Postgres or disabling RLS; the interview notes that models had since improved on those behaviors.
- Supabase largely lets teams choose their own AI coding tools and policies, while consolidating some cost decisions; it reported spending about $300,000 per month across 450 employees. Its stated effectiveness focus is end-to-end processes solved and reducing human-in-the-loop work, not spend alone; it uses its async company documentation as an AI harness for surfacing company principles and historical decisions.
- DHH says he ran “Kimmy K25” as his primary driver for over a month in spring and liked its speed: about 200 tokens/second versus 30–40 tokens/second for frontier models at the time.
- His Linux distro project integrates agents for making themes, apps, and plugins. For the China-oriented setup, he aims for a preconfigured developer machine ready in 30 seconds rather than requiring half an hour or half a day of setup; this is a stated target.
Factory describes a multi-agent workflow for larger projects: Missions assign separate parts of the work to droids, a validator clicks through the working application and sends failures back to builders, and Factory Router selects a model as work progresses. Factory claims its Router cut costs 63% versus frontier models in production, with 84% of routed sessions saving at least half.
Google announced Gemini 4 Carbon High but had not made it generally available: rollout was to start with cybersecurity partners, expand to trusted partners, and reach developers, businesses, and consumers later; the presenter says he had not tried it. In the announced benchmarks, Carbon scored 65 on Terminal Bench 4.0 versus Argon’s 57; Google also reported 0% successful cheating and 4% attempts to cheat on that benchmark, and about half as many unsupported claims as Argon. Announced pricing was $2 per million input tokens and $10 per million output tokens during an introductory period, then $4/$20; Google did not specify the introductory period’s length.
- For low-stakes browser-game prototypes, start in regular Claude or ChatGPT chat rather than a full coding agent: prompt for a “mobile-friendly three.js game,” build in HTML/JavaScript for static hosting, then move to Codex or Claude Code after the proof of concept; if the game is not fun enough, try “Make it a LOT more fun.”
- A concrete one-shot approach: ask for a specific core loop and aesthetic, then explicitly ask the model to think about fun and beauty. A prompt for a three.js spot-the-difference game using procedurally generated trees produced a mobile-friendly game in regular Claude chat with no follow-up prompts.
- For a larger game, build in stages—prototype the engine and world, add requested features, then expand into mechanics and plot—and playtest between stages. Emberwick needed repeated playtesting and prompts to fix quest problems caused by characters moving to the wrong location; the completed story took about 15 minutes to play.
- In Willison’s comparison, Claude Fable 5 made Battle Drones a workable but unfun 2D RTS; Opus 5.5’s “Make it a LOT more fun” prompt and subsequent port to three.js transformed it. The project grew to support 2–4-player multiplayer through a vibe-coded WebSockets relay, and its Apache 2.0 release was followed by Panache, a French steampunk-themed fork.
- Model caveat: Willison found GPT-6 Astra could build games but did not match Opus 5.5’s sense of fun; a GPT-6 Pro version of Emberwick looked good but, in his view, played less well.
-
For Opus 5.5, use one prompt stating the goal and how Claude should verify success; start with
/effort mediumand adjust. If the task is unclear, ask Claude to interview you until it can write a spec, then start a new session to build from it. For unattended work, provide a pass/fail check or numeric target with readable reports;/goalcan have a separate model check a condition after each turn on long background tasks. -
For models like Opus 5 and Fable 5, over 80% of the Claude Code system prompt was removed without measurable degradation on coding evals. For a new model, try trimming your own instructions and add back only what it misses;
/doctor prompt-auditflags instructions written for older models. - Let Claude delegate to subagents by default; explicitly request them for substantial standalone work such as audits or migrations, not small jobs where they add cost and latency. Subagents can parallelize work or provide a second opinion, but do not inherently make the agent smarter. For custom agents, a single agent with skills can often outperform a subagent design at lower cost and latency, based on the author's experience.
- Use Claude Code cloud sessions for work you want to continue away from your laptop: they run in a separate VM with a fresh repo clone, so push commits first; check in, answer questions, then review the diff and open a PR. Keep work local if it needs machine-only resources. Parallel sessions pay off only once a single session can finish without you; Projects can organize issues into parallel threads, apply shared instructions, and track completed tasks, ready PRs, and pending input. Projects were in public beta on Pro and Max when the article was written.
-
Keep human review in the loop:
/code-reviewreviews a diff in a subagent that has not seen the conversation, or use a new session with the spec and instructions to ignore stylistic issues. The author reports that Claude Code team changes are usually reviewed and fixed by Claude at least once before a human looks.
- At Anthropic, Claude Code performs a first-pass PR review and can test changes more like a real user: run end-to-end tasks, open the experience in a browser, and run tools, linting, and other checks. Hooks rerun on edits and inspect diffs to help correct PRs quickly.
-
Put team standards and “taste” in shared
Claude.mdoragents.mdfiles for the review agent to consult; have people check the agent’s evidence, such as browser-tested UI and generated videos or screenshots. Addy says end-to-end, property, and visual-verification tests are useful for building confidence in changes and enabling automatic merges.
- For functional reimplementations of familiar software, Sanfilippo says decompilation is often unnecessary: use the working program as an oracle, probe it with functional tests, turn observed behavior into a description, and reimplement from that description; he also describes this as a route to clean-room implementations.
- Reserve decompilation for cases where exact behavior matters, such as proprietary file formats, peculiar algorithms, or hardware drivers. For a narrow reverse-engineering question, a frontier model may be able to inspect assembly directly; in one example, Sanfilippo used binaries from a 1990s system to reconstruct its simulator and says it worked well.
- In a decompilation test, Sanfilippo gave Opus a binary and compared its reconstruction with Ghidra output using another agent. He reports that Opus recovered runtime exceptions and meaningful function and variable names, going beyond mechanical decompilation.
GitHub activity surged in 2026: merged pull requests tripled in under a year to 220 million per month, commits more than tripled to 7.38 billion per month (with most of the increase in the second half), and new repositories doubled to 30.3 million per month. Willison presents these figures as evidence of agentic engineering’s impact.
Swyx said Gemini’s Artificial Analysis lead prompted him to reopen Antigravity and that he was getting an early review. On the Artificial Analysis Coding Agent Index, Gemini 4 Carbon (high) via Antigravity CLI scored 59, three points behind Claude Code with Claude Opus 5.5 (62); it ranked second on Terminal-Bench 4.0 and FrontierCode v1.1. Carbon was rolling out to selected users and was not yet publicly available.
In a sponsored Hyper Agent demo, Fireship’s presenter used a Slack-like “room” for work on a V4 date-library release, bringing agents and human co-maintainers together. Separate agents triaged incoming issues, drafted feature PRs, and wrote the changelog in parallel, handed work off, and contacted a human when approval was needed; teammates could also ask questions or make changes directly. The presenter said users could sign in with an existing ChatGPT plan rather than set up another AI subscription.
Kent C. Dodds puts Kody features behind flags linked to metrics, using the results to decide whether to enable them for everyone or delete them; he says the same system led him to reject and delete another feature. For Kody’s “Rerank” feature, he planned to enable it after the comparison reported 7.6× more executions and a 5× lower error rate among enabled users.
Terk Terkowitz’s guidance for engineers: if current AI tools do not help with your work, document why, share the finding with peers and your manager, and work to find solutions; being able to explain where AI falls short demonstrates the judgment expected of engineers .
For choosing models in coding-agent workflows, Ted Sanders says capability is multidimensional, but larger models generally offer better performance and reliability and better recall of esoteric facts, at higher per-token cost; this is a broad tendency, not a coding-specific benchmark. Model version numbers can signal release order, relationships between releases, and the scale of improvement—for example, 6.1 could follow 5.6 without representing a leap large enough to warrant a new major number.
- Riley uses Claude Projects as a phone-accessible coordinator that routes work to local computer sessions or cloud sessions; he runs computer-dependent editing locally while delegating research and artifact creation to the cloud, including while the local session is working. Local sessions require his computer to be on.
- His reported tool split is Claude Code/Projects for building apps and creating designs, slides, and video edits; he prefers Opus 5.5 for front-end work and visualizing information. He uses Codex through Dot for computer-use and business workflows: by voice, it can inspect a video transcript, prepare a YouTube upload with details and sponsor links, and save it as a draft for him to review before publishing.
- Riley uses Composio to centralize integrations and skills—including custom tools—so they can be reused across agent platforms. Cross-platform memory remains an unresolved gap for him: he has not found a simple way to keep agent memory current when switching platforms, and his Supermemory setup did not update it frequently enough.
- For Twitter research, he delegates from Claude Code to Grokbot through a webhook skill, using Grokbot’s Twitter access without working in its UI.
Simon Willison identifies October as another inflection point for coding-agent capabilities, saying Claude Opus 5.5 can build video games that are “actually fun.”
Addy Osmani argues that, as execution becomes cheaper, engineers should demonstrate judgment by prototyping an idea that moves the business internally or building a portfolio of small prototypes externally; both should show taste and judgment. He cites Claude Code Desktop as an example: two newer engineers started it as a weekend prototype.
