David Wilson

@nbaschez Would love to try. Made leafmill.net, which is free and online so OpenClaw/agents can send you links on the go. Use it multiple times a day.

Do you spend a lot of time reviewing markdown docs written by AI? Wish it were a better experience? Say hi if you wanna try a new (free, open source) thing

@nikunj @gauravmc @claudeai @OpenAI @conductor_build @Railway Love the conductor tabs too - great for reviewing new work without context/bias. But the plugin approach is nice for iterating hands-off. Planning hard things for 10-30 mins is useful. Particularly on an older/larger codebase.

@daviddbwilson @gauravmc @claudeai @OpenAI @conductor_build @Railway Ah! Yes

Even with all the model releases, this “stack” continues to be undefeated (for me).. - @claudeai Opus for planning and frontend design - @OpenAI codex for engineering - @conductor_build to orchestrate - @Railway to deploy It’s never been more easy and fun to build!

@nikunj @gauravmc @claudeai @OpenAI @conductor_build @Railway I do it IN conductor :)

@daviddbwilson @gauravmc @claudeai @OpenAI @conductor_build @Railway I just use conductor so don’t even need to do that. It’s just a new tab

@nikunj @gauravmc @claudeai @OpenAI @conductor_build @Railway Have you tried the /codex plugin for claude code? Best planning flow today imho is smth like: "Plan carefully. Send to /codex for an adversarial review. Only incorporate *real* improvements into the plan. Keep iterating until codex has no substantive suggestions."

@gauravmc @claudeai @OpenAI @conductor_build @Railway Codex planning sucks for me. Might be a skill issue.

Very similar to mine. The new Codex plugin for CC makes #1 easy! I skip Cowork for #2 (plain English plan). For complex work with lots of decisions/complexity: - Ask CC: “describe the key decisions and trade-offs in plain English - use tables and mermaid charts - and push to leafmill.net” (really nice ephemeral markdown rendering) Or: - Use CC + visual explainer skill: github.com/nicobailon/vis…

i'm sure we all have our little coding 'hacks,' here are my top 5, please share yours (or help me improve mine! 🙏): 1. Plan -> Deepen Plan -> Then let Codex review the plan, then hand it back to Claude Code 2. Let Co-Work read the plan and build you a PDF of the plan in plain english along with flowcharts (vs just "go to work!"), this is a great for overall logic agreement 3. If I'm unsure of a stack or an algorithm choice (e.g. best algo for clustering objects with vector embeddings), give it to the beast models and let them deep research it for 20 mins 4. If have something big to tackle, always quit and restart Claude Code 5. On big PRs, I always let CC, Codex, and @greptile view it (at the same time), never fails to find some P1s

@JeyasankarKavin Our approach makes this easy actually: everything, including stopping, is a tool call. For example: any agent run that doesn’t have a “stop” tool call after x turns / minutes gets a little prompt injected into the next turn to check if it’s still on track etc

@daviddbwilson the overnight burn is one of those things you only learn the painful way. how did you handle the cases where the agent hit a guardrail but the task was legitimately complex, did you give it a way to signal that vs just silently stopping?

@kevinrose looking forward to listening on the muni if you ask claude to upload the audio to airloom.fm, people can listen/subscribe in their podcast app

I used ai to reverse-engineered anthropic’s claude code prompt architecture and turned it into a standalone playbook for building best in class LLM system prompts: layering, variable injection, trust boundaries, tool policy, verifier roles, memory, caching, and real templates.

@steveruizok - cache invalidation - naming things - rich text editing

Other hard problems in web development seeking extreme engineers: - efficient drop shadows for svg paths - distinguishing between a scroll-wheel and a trackpad in mousewheel events - unfurling urls to access social metadata - rich text editing - bottom navbars

@oliviakoshy @_hex_tech omg

the @_hex_tech brand & design team have once again truly outdone themselves i'm OBSESSED with our new careers page, there's so many fun details and easter eggs

@YucefBsf @ron_joshi We’re seeing mainly personalized daily briefings (your calendar, news), articles and docs read to you, shortened versions of podcasts. A cool one is researching your latest X bookmarks and creating short podcasts with more info.

@daviddbwilson @ron_joshi What's the use case you're thinking of for airloom ? like publishing Claude conversations as a podcast feed, or more for sharing individual audio clips? Trying to figure out the best way to integrate it

Introducing Kitten TTS V0.8: open-source TTS that fits in 25MB. Three variants: 80M | 40M | 14M (<25MB) Highly expressive. Runs on CPU. Built for edge. No GPU? No problem. Ship voice anywhere. Check it out:

@YucefBsf @ron_joshi Have you guys tried airloom.fm with your plugin/model yet?

@ron_joshi This is awesome ! just added Kitten TTS as a model in our Claude Code TTS plugin. Runs fully local which is nice. Can't wait to test the new voices github.com/MatiousCorp/cl…

@packyM ...and it's so short! Incredible book.

John McPhee's Levels of the Game is even better than advertised. I just don't understand how someone writes that well. I expected it to paint the match in vivid detail (it did), but the way he weaves in all of the stories, the callbacks, the character tics, it's just so good.

@nikunj You can also use skills to spawn subagents! One of my frequently used skills assembles an advisory board of distinct people/perspectives and uses subagents to get their advice in parallel and without cross contamination.

TIL - you can spawn subagents for skills in Claude Code. What.. I feel so stupid now. This would have saved me SO much time. Every day, you learn something new.

Super ambitious. Chat/code + canvas + kanban + slides... This is a glimpse of the future, where agents will adapt their UI to the task.

The Final Agent UI will feel like: - one living UI - keeping you in flow - managing many agents & projects - serving you the ideal context for each decision - to saturate your cognition

@fletchrichman thanks for sharing! Do the agents also check the dashboards btw, or are they for you?

Each of the 5 agents i outlined has a weekly report/dashboard that it updates. My favorites: Data analyst pulls from posthog + read only postgres database and gives me a customer health dash Marketing manager pulls from google analytics, x, and reports on top of funnel + opportunities

I’m increasingly finding planning to be the bottleneck! Sufficiently in-depth planning, reviewed in a loop by opus and codex (with warnings not to overengineer), seems to reduce testing/fixing iterations. Really enjoying the ability in conductor to have parallel convos in a single worktree with Codex and Opus.

This week I ran parallel coding agent sessions for the first time, using @conductor_build. Thoughts – • It is good. • I max out at 3 concurrent projects, mentally. Switching costs are too high past that. • Tasks need to be broken apart until they are small enough for an AI to one-shot. • My projects are usually in different phases. One in planning, another in development, and the last in testing. • Testing is a bottleneck. • Anything that requires human input is a bottleneck. • Models constantly hack the tool calls you permit. It can't `rm` but it can run `python -c "rm /"` • You really, really need to put Codex in a verifiable loop. It will move mountains if it can reliably test itself. • Models can use Chrome to read the Dev Tools console automatically • When a model can't fix a bug, it never volunteers to add debugging code. I think this can be solved. • I like Conductor a lot. Linear -> Open In Conductor. • Claude writes, Codex reviews. • I still hand write something in 70%+ of PRs. • Models can't write PR descriptions. I'm not sure this can be solved. • Claude still seems bad at Frontend. • I like to walk, dictate plans to Claude, have it write a spec. I don't like evaluating code from my phone. • CI/CD needs to be fast. Running tests needs to be fast. Dev builds need to be fast. CPU cores are cheap, human time is not. • Humans should still be in code review. Use models to help you grok a PR. Don't outsource thinking here yet. We are just hitting the surface of what these models can do and how they'll be deployed. Extremely bullish.

@nikunj @conductor_build @WisprFlow for yapping at Claude

current pipeline for building internal tools + projects: > claude code w/ @conductor_build for building > opus thinker > codex planner > opus doer > codex / devin reviewer > nanobanana for design assets and gemini 3 flash for large context LLM calls what's wrong or missing?

Just shared a bunch of tactics for how to make agents work reliably in the background: x.com/daviddbwilson/… incl: - making everything a tool call (including stopping) - treating tool metadata as prompts and rewriting them all - using deterministic binary checks with cheap LLMs plus optional review from expensive LLMs for quality control - using multiple simple memory approaches instead of one complex system.

if you’re watching all of this and your first instinct is to start building your own agent from scratch, i want to be your friend drop one of your favorite unique agent building tactics here, and if i like it, i’ll invite you to a small DM group for sharing ideas and questions around building better autonomous agents (i’m rebuilding now and have lots of fun ideas and very specific questions but don’t want to spam public feed)