We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Long runs: give a blunt goal, then a structured brief
Simon Willison built Battle Drones, a browser RTS, mostly with Claude Code. Two of his prompts are worth reusing.
- A vague goal plus orchestration. He sent Opus 5.5 the prompt "This game isn't fun enough yet. Make it a LOT more fun." in Ultracode mode, where Claude Code runs "dynamic workflows": JavaScript scripts that orchestrate many subagents at once . The run took from evening until late the next morning . Claude ran more than a thousand simulated playtests and used the results to rebalance units and add features. In 1,200 AI-vs-AI matches, games still undecided after 25 minutes fell from 14% to 0.9% .
- Research and critique before building. When the game's AI ignored infantry and troop carriers, he wrote a detailed design brief: the incentives, the map constraints and the win condition. It ended with "First research and then critique this design. Only build after we have agreed on any modifications" . The result was the Archipelago map, which he calls the most fun map in the game .
Other lessons from the project:
- Static HTML/JS with Three.js is a strong medium for agent-built games because the models know those platforms well .
- Giving Claude an OpenAI API key lets it generate images, look at them and refine its prompts .
-
Replays and multiplayer need a deterministic simulation. Claude had to replace
Math.sin()/Math.cos()with custom versions because Chrome and Safari returned slightly different numbers .
Peter Steinberger started a loop to move OpenClaw's Lit frontend to Solid 2.0. Removing Lit took about 6 days, and making the code idiomatic Solid took another week. He checked in every other day, and the work took about 600 PRs . The agent missed a few small regressions, mostly button shapes and minor UX details . His setup:
- Astra plans and Sol 6.1 codes. Prompting Astra to run Codex CLI in worktrees worked "FAR better than relying on subagents" .
-
A second agent challenges the first. Instead of one agent checking its own progress against a text file, as with
/goal, a separate agent pushes back on Astra. He built this as an OpenClaw plugin he calls "smart goal" .
He also moved OpenClaw to tsrs (npx tsrs). Its announcement describes a Rust port of the TypeScript compiler, built in collaboration with the TypeScript team. It claims up to 8x the speed of TS 7.0 and about 70% less memory, runs in WASM, and is due to join Oxc . Steinberger calls it a drop-in replacement and saw about a 5x speedup , or 9x against stable releases, since they were already on 7.1-dev .
Give agents a paved path to verification
Kent C. Dodds says the fix for low confidence in legacy codebases is verification. Agents should do what humans did: run with mock data, use the UI, seed varied database data and deploy to a production-like environment. They will do this reliably if you give them a paved path . His agents maintain a control-kody CLI for exactly this (SKILL.md) . Separately, a team that was "all in" on agents turned out not to be running tests in CI. Dodds calls that a prompting problem: an agent asked to audit the codebase for obvious problems would have caught it .
Theo reported something similar after a rough night. Most breakages came from legacy vendors, GitHub Actions and flaky Chromium-driver tests. The "slop" parts, such as ts-rust, were stable . In a new video, he describes building his machine images for the agents rather than for himself. The image is a customized, headless Ubuntu with Bash instead of Zsh. His audit found over 30% of Claude's errors came from Zsh's newline handling breaking ripgrep commands, and agents also did worse on Arch and CachyOS dev boxes .
AGENTS.md as a scheduler for follow-up work
After the eMudhra CA key leak, Geoffrey Huntley asked an agent whether his NixOS fleet was affected. Because Nix has a single source of truth, one patch covers transitive dependencies: patch, commit, autodeploy . His new trick is to have the agent write a dated reminder into AGENTS.md so the temporary workaround gets backed out later . The reminder names the nixpkgs commit to check for, the command to run, a recheck every two days, which files to delete, and which unrelated changes to keep .
Overalls also got "hive mode." Agents in the same profile zone (oss, standard or customer) share a message board for coordinating work. The customer zone has no outbound connectivity, and a machine's entries are wiped when it suspends . Huntley points to Claude Code's crossSessionInbound=accept setting as the apparent primitive behind hive swarms , and says it has him "back to loving claude" . He is moving 95% of his usage to Anthropic because OpenAI "falls over" after every usage reset .
OpenAI: longer cache retention and a usage-limit fix
The API now offers 12-hour prompt caching, so persistent agents pay the cached-input rate even when requests are hours apart . Tibo says logged-in Codex and Work users get cache kept "forever," with guaranteed hits when they resume old conversations . OpenAI also says about 40,000 accounts had usage limits that didn't match their plan. It has corrected the issue and will add credits for affected users .
T3 Code Stable
The new T3 Code Stable release has more than 1,400 changes . Highlights:
- Add a project to another machine from the "Run on" menu.
- PR diffs that separate code, tests and docs.
- Idle Claude threads compact early so the cache stays warm (beta).
-
t3 connectreachability checks. - Webhook-task filters and dedupe.
- More reliable skill discovery .
Theo suggests moving back to Stable because Nightly is getting more experimental .
Other notes
- Tenex's "meta harness" is structured repo docs plus a CLI that agents use to drive plans and readouts. Agents log their decisions, and everything is checked into Git, so the reasons behind changes live in the repo . Tenex has loosened the plan template as models improved . Its engineers mostly drive with Opus 5.5 and use Codex to review, which they say gives the best results . They report output roughly doubled this year .
- Claude Code Projects tip: tell the coordinator to save a project-memory rule to post a one-line update linking the thread whenever a thread is blocked or finishes a milestone. Before this, it wouldn't even report blocked threads .
- Kody lets agents file feedback about their experience, which often becomes a fix within an hour or two .
- Simon Willison built Battle Drones through 12 major iterations and 321 agent commits, mainly by Claude Code; although he knew little TypeScript and reviewed almost none of the code, he playtested it extensively . For a major gameplay revision, he prompted Claude Opus 5.5 with “Make it a LOT more fun” in Ultracode, a dynamic workflow that orchestrates many subagents; the overnight run included more than 1,000 simulated playtests . In 1,200 Normal AI-vs-AI matches, games still undecided after 25 minutes fell from 14% to 0.9%, while wins by sides that had been 800+ scrap behind rose from 24% to 46% .
- When the game’s AI ignored infantry and troop-carrying Condors, Willison turned that observation into a design brief: specify the strategic incentives and map constraints, then have the agent research and critique the design before building, with changes agreed first . The resulting Archipelago map had seven uplinks capturable only by flying infantry in a Condor, and Willison described it as the game’s most fun map .
- To add audio multiplayer, he pointed Claude Code at his working Chirp Pong proof of concept and asked it to adapt the approach; Claude built a Radio Probe to measure bandwidth before implementing an audio modem . He had not extensively tested the feature, though it seemed to work across laptops, iPhones, and iPads; after a player got a headache from the high-pitched sounds, he added a warning and moved toward online multiplayer, using a stateless WebSocket relay after WebRTC failed between his devices .
- The Claude Code team at Anthropic describes a review workflow in which Claude reviews and fixes most changes before a human reviewer is asked to focus on higher-level questions; it frames review around what could go wrong and which automated or human check can catch it, rather than whether a human read every line . As a complementary enforcement technique, Geoffrey Huntley described a pre-commit hook that checks diffs for direct calls bypassing a diagnostics library, fails the commit, and injects a prompt telling the agent to use the library—avoiding reliance on instructions that can drift .
-
DeepSeek open-sourced DeepInfer, the production inference framework used for V4.1 Flash, with distributed inference, speculative decoding, SWA Bounded Replay and structured output; it runs on NVIDIA and Huawei Ascend and includes the
deepseek-recipeAPI/token conversion component. KVCache management and load balancing are not fully open-sourced. - Agent-serving cache pattern: retain reusable Block KVCache prefixes in shared 3FS storage. DeepSeek reports a 3-day TTL for agent traffic; shared storage lets any prefill machine hit any prefix without prefix-affinity routing, reducing repeated prefill work. SWA KV remains in machine memory (about 180 seconds’ TTL and over 98% hit rate); on a miss, SWA Bounded Replay recalculates only the latest 128 tokens to rebuild it.
-
deepseek-recipemaps prompt-builder-inserted special tokens to token IDs and treats user/tool text that merely resembles them as ordinary text, preventing token injection; it also parallel-encodes ordinary-text segments between known special-token boundaries to reduce long-context tokenization latency. DSpark allocates speculative-token validation according to per-token confidence; DeepSeek reports 6% higher online throughput than fixed five-token validation.
-
For an agent-visible personal workspace, he used a Markdown
to-do's.mdfile as the shared data layer and synced it to a web app every 10 minutes; he and his agent could edit it from different devices, and access was restricted to approved email addresses. - To keep build requests from interrupting his personal-agent conversations, he set up a separate builder, Bill, as a new thread using workspace instructions; the personal agent can pass build requests to Bill, and he finds agents better at prompting one another than he is. For review, he asked subagents to test desktop and mobile, check for bugs, and perform a thorough review, then had the agent fix the reported issues.
- In BB, he uses Codex, Claude, Factory, and Pi; for UI exploration, a prototyping scale generated three alternatives that he mixed and refined with a simple request for a Linear-like feel, producing features such as a Kanban board and alternate detail views. He also uses Executor to connect services and tools once, then gives its setup prompt to an agent so the connected tools can be used across workflows.
- Omarchy is designed as an agent-ready Linux desktop: agents are preconfigured so users can start making a theme, app, or plugin from an invitation to “make something cool.” DHH’s stated setup target is a computer ready to use in 30 seconds, rather than spending half an hour or half a day configuring a developer machine.
- He says he used a model he names “Kim E K25” as his primary driver for over a month, praising its 200 tokens per second versus 30–40 tokens per second for frontier models at the time.
- Simon Willison built Battle Drones through 12 major iterations and 321 coding-agent commits, mainly with Claude Code; he reviewed almost none of the TypeScript code but playtested extensively. He found static HTML/JavaScript with WebGL, Phaser, and Three.js effective platforms for agent-built games.
- For a stalled gameplay problem, he gave Opus 5.5 the blunt prompt “Make it a LOT more fun” using Claude Code’s Ultracode dynamic workflows, which orchestrate multiple subagents. The overnight run included more than 1,000 simulated playtests; in 1,200 AI-vs-AI matches on Normal, games still undecided after 25 minutes fell from 14% to 0.9%, while wins by a side that had been 800+ scrap behind rose from 24% to 46%.
- When infantry and troop-carrying Condors lacked a strategic role, he prompted Claude to design a map around that problem and explicitly requested research and critique before implementation, with agreement on modifications before building. The resulting Archipelago map made seven island uplinks capturable only by flying infantry there in a Condor.
- To add audio multiplayer, he pointed Claude Code at his working Chirp Pong proof of concept, then had it build a Battle Drones version. Claude created a Radio Probe to measure device bandwidth before implementing a full audio modem; Willison reported successful trials between laptops, iPhones, and iPads, while cautioning that testing was still limited.
-
For replay and multiplayer, use a deterministic simulation: tiny differences in browser implementations of
Math.sin()andMath.cos()caused desynchronization, so Claude replaced them with custom versions.
- For this add-on, the host treated model choice as secondary: a capable Codex setup was likely enough, while Opus was described as more correct but much more expensive. Cursor was recommended as an approachable interface for building without relying on the terminal.
- Keeping the project folder pointed at the WoW add-on folder made changes testable with a game reload; the host also recommended initializing Git. Local-only version control was considered workable, but a remote copy would help if switching computers or losing the machine.
- Check platform capabilities before settling on a feature: in this WoW add-on exploration, the investigation found no general HTTP/socket or arbitrary file access, and detailed combat-log events were restricted.
- Use a review gate before implementation: the host avoids formal plan mode but has agents discuss their approach and answer questions before he authorizes coding, because their eventual implementation can differ from what he expected.
- Build and verify in small increments: the first generated legacy-menu dropdown reused an existing screen, so the pair tested it in-game and refined the request; the host recommends starting small and working one step at a time.
- The host said dictating ideas on his phone while pacing worked better for him than typing, which could not keep up with his thoughts. He also said agent-assisted coding had become better than his hand coding for completing tasks, while manual coding still produced better quality and thoughtfulness; with implementation less of a barrier, idea and design quality remain limiting factors.
- Tenex’s “meta harness” pairs structured repository documentation with a CLI that agents use for plans, visualizations, client readouts, and other parts of the engineering process; its canonical plan has become less prescriptive as models improved. Agents log decisions and activity, check the artifacts into Git, and keep the context for why decisions were made alongside the code as a durable record.
- The team’s current model preference is Opus 5.5, with Codex still useful for computer-use tasks and for reviewing Opus’s work; engineers report the back-and-forth produces the best results. One engineer uses AMP’s cloud “orbs” to run work off his computer and have another orb review Opus’s output; he says a newly added subscription hookup cuts costs by about 80%.
- Tenex reports roughly doubled engineer output over the period discussed, tracks average story points per engineer per month, and re-evaluates story-point criteria and project estimates at least quarterly as models improve.
- DeepSeek open-sourced DeepInfer, its large-scale inference framework for V4.1 Flash. It supports distributed inference, speculative decoding and structured output, and runs on NVIDIA and Huawei Ascend hardware.
- DeepSeek Recipe handles encoding and decoding between mainstream LLM API protocols and model tokens, with processing aligned to DeepSeek’s official API for third-party deployments. For agent-serving systems, DeepSeek identifies prefix KVCache reuse as a way to cut repeated prefill as contexts grow; its stated agent-traffic cache TTL is 3 days. Its persistent Block KVCache is stored on 3FS, while SWA KVCache stays in machine memory and has a reported ~180-second TTL and over 98% hit rate.
- Caveat: DeepInfer’s KVCache-management and load-balancing components are coupled to DeepSeek’s internal infrastructure and are not yet fully open source.
Sanfilippo describes DeepSeek Infer as a cloud inference engine for DeepSeek V4.1 Flash, built from a vLLM fork and ideas from other engines; its key design choice is specializing kernels for a particular model rather than maintaining a general-purpose engine. The release targets SM100/B200-class Nvidia and Huawei accelerators, but is incomplete for some serving needs: KV-cache management is withheld as a proprietary advantage, and prefill/generation separation for microbatching is not included. For local deployment, he cautions that kernels tuned for SM100 do not transfer directly to DGX Spark’s SM121 architecture and are optimized for heavy batching; he may adapt selected parts only if they offer a benefit.
- After a CA-key leak, Huntley prompted an agent to check whether his NixOS fleet was affected. He used Nix’s single source of truth to patch transitive dependencies fleet-wide, then commit and autodeploy the fix.
-
For a temporary workaround, he had the agent add an
AGENTS.mdreminder to reassess and back out the patch later. The reminder specified checking on or after Oct. 22 whethernixos-26.05included nixpkgs commit7816d39, checking again two days later if not, and removing the workaround once it was included—while retaining separate TLS andrustls-webpkichanges.
Simon Willison built browser-based RTS Battle Drones mainly with Claude Opus 5.5; Claude also generated all of its Three.js 3D models, and a model explorer lists their animation states.
The game can pit two LLMs against each other by pausing every five seconds for them to assess the game state and queue commands; a replay pairs Claude Haiku 5.5 with GPT-6 Luna. The device-to-device audio multiplayer produced a persistent high-pitched noise, prompting a warning that it might cause a headache; Willison added online multiplayer using a small Fly.io-hosted WebSockets relay, so players can share a link to join.
- Tenex’s MetaHarness combines shared skills, structured project documentation and a CLI. Agents work from a canonical plan, log decisions and activity, and check the records into Git alongside the code—preserving the rationale and context needed to resume long-running work. Tenex cautions that this level of structure can be overkill for a small team on one project.
- As models can get to roughly 90% of a solution when given the right context, Tenex’s CJ says his work has shifted from coding toward drawing out customer context and providing it to the model.
- Tenex describes Opus 5.5 as the team’s usual model, with Codex used for computer-use tasks and review; the discussion also describes using AMP’s remote Orbs to switch models and review another agent’s work.
- Tenex reports engineering output roughly doubled over the period as models improved at long-running tasks, and says it reevaluates story-point criteria at least quarterly.
A TODO-board build was shared with a video and a post containing copy/paste instructions for agents, giving developers a starting point for trying an agent-assisted project.
- Macroscope describes a PR-review workflow in which language-specific code “walkers” provide change context and team rules run as native GitHub checks that can block failing PRs. The workflow auto-approves a PR when it passes those checks and a code review and has a small blast radius; Macroscope says it now auto-approves 40% of customer PRs.
- For motion graphics, Claude can generate the code that renders a sequence instead of producing video with diffusion; having the code makes the result easier to customize and analyze.
- The video’s self-described “Trust Me Bro” benchmark roundup gives Gemini 4 Argon a mixed coding profile: it reports a lead on deep software engineering, but Claude Opus 5.5 scores 66 to 57 on Terminal Bench 4.0, GPT-6 Astra leads by more than 10 points on Frontier S, and Argon ranks ninth in webdev Code Arena. These are the video’s benchmark-specific claims, not a settled overall ranking.
For AINews reporting, swyx made Opus 5.5 the default after a side-by-side evaluation, praising its concise, tasteful writing and reduced “slopese” versus Opus 5 . He also cautioned that Opus’s writing quality can be undermined by unspecified “no-no words” in prompts, making it unsuitable for the critical path of his AI media organization .
- Theo built a custom Ubuntu-based environment for agents rather than his personal preferences: it is headless but can spin up XFCE-fork GUIs for agent computer use, and uses Bash instead of Zsh. He says over 30% of Claude’s errors across his machines came from Zsh newline formatting that led Claude to misformat ripgrep commands; he also found agents less capable on Arch and CachyOS dev boxes.
- Choose the stack around agent throughput, not developer preference: Theo kept the React Native mobile app instead of pursuing his preferred SwiftUI rewrite because agents could work faster in the existing app; he says team performance work made it comparable to, or in some ways faster than, the alternative.
- Theo reports that an AI-assisted Rust rewrite of the TypeScript compiler was working in under 20 hours, ran 2–3 times faster than TSGO, supported WASM, and achieved 99.99% accuracy against other TypeScript versions on the real projects he cared about.
Peter Steinberger says his “claw” automatically mutes most replies he characterizes as “slop.”
Ben Tossell describes a custom workspace combining a Superhuman-like email experience with a Linear-style todo board; the email is built on Rhys Sullivan’s Executor. His bot uses the workspace and can be tagged in comments; it also tracks domains, expirations, hosts, and linked local repositories, with plans to add links to mini-apps. He shared a video and copy/paste instructions for agents explaining how he built it.
ThePrimeagen says he dislikes Claude’s apparent preference for legacy versions “just in case.”
The quoted post claims DeepSeek V4.1 Flash 1017 beat GPT 6 Sol on a multi-step offensive-cybersecurity task and says it can be run locally . Primeagen disputed the local-running claim, saying that even his “incredible setup” is far from sufficient to run it locally .
