https://epoch-research.github.io/ai-productivity-digest/ High-signal AI productivity tweets selected by Epoch's daily digest bot. en-us Sat, 25 Jul 2026 13:31:26 +0000 1440 https://x.com/lucasmeijer/status/2080761379007471980 https://x.com/lucasmeijer/status/2080761379007471980 Fri, 24 Jul 2026 21:05:46 +0000 Search across your past Claude/agent sessions in a project to recall prior context — a high-leverage recurring prompt pattern Source: @lucasmeijer https://x.com/thdxr/status/2080723541490573601 https://x.com/thdxr/status/2080723541490573601 Fri, 24 Jul 2026 18:35:24 +0000 Use Cloudflare Access to gate /admin/* routes of small AI-built tools behind Google login while keeping the rest public Source: @thdxr https://x.com/intellectronica/status/2080693046157693357 https://x.com/intellectronica/status/2080693046157693357 Fri, 24 Jul 2026 16:34:14 +0000 If you draft in markdown but collaborate in Google Docs/Word/Notion, have an agent maintain a linked doc and keep both in sync Source: @intellectronica https://x.com/steipete/status/2080690608105255057 https://x.com/steipete/status/2080690608105255057 Fri, 24 Jul 2026 16:24:32 +0000 Use ChatGPT Voice on desktop as a planning partner: start with a fuzzy idea and let it ask sharp questions to sharpen it Source: @steipete https://x.com/Dimillian/status/2080662531102716371 https://x.com/Dimillian/status/2080662531102716371 Fri, 24 Jul 2026 14:32:58 +0000 Run ChatGPT Voice alongside Codex to ramble ideas, check task/PR status, spawn parallel worktrees, refine prompts, and generate concepts hands-free Source: @Dimillian https://x.com/intellectronica/status/2080616187399184539 https://x.com/intellectronica/status/2080616187399184539 Fri, 24 Jul 2026 11:28:49 +0000 Co-write documents by leaving ChatGPT Voice on alongside your editor — dictate, critique, and ask for edits in real time Source: @intellectronica https://x.com/ivanfioravanti/status/2080527462216147241 https://x.com/ivanfioravanti/status/2080527462216147241 Fri, 24 Jul 2026 05:36:15 +0000 When a coding agent gets stuck, ask it to write a handoff .md summarizing failed attempts, then reference that file in a fresh chat to reset context Source: @ivanfioravanti https://x.com/geoffreylitt/status/2080433142456864773 https://x.com/geoffreylitt/status/2080433142456864773 Thu, 23 Jul 2026 23:21:28 +0000 Ask your coding agent to attach annotated screenshots to PRs so reviewers can see UI changes at a glance Source: @geoffreylitt https://x.com/petergyang/status/2080340987516539256 https://x.com/petergyang/status/2080340987516539256 Thu, 23 Jul 2026 17:15:16 +0000 For app UI work: generate a target design image in ChatGPT, then use /goal with the image attached to make the code match Source: @petergyang https://x.com/gabrielchua/status/2080300469298041267 https://x.com/gabrielchua/status/2080300469298041267 Thu, 23 Jul 2026 14:34:16 +0000 Pair Codex Computer Use with WhatsApp to auto-push daily summaries, to-dos, or meeting notes into your chats and channels Source: @gabrielchua https://x.com/Saboo_Shubham_/status/2080170745934786650 https://x.com/Saboo_Shubham_/status/2080170745934786650 Thu, 23 Jul 2026 05:58:48 +0000 Ramble your thoughts to an agent via voice, then have it summarize what it heard back to you before it acts — catches misinterpretations early Source: @Saboo_Shubham_ https://x.com/clairevo/status/2080128077229862996 https://x.com/clairevo/status/2080128077229862996 Thu, 23 Jul 2026 03:09:15 +0000 Concrete Codex computer-use tasks worth trying: finding front-end bugs, synthetic browser persona research, LinkedIn inbox cleanup, and travel shopping Source: @clairevo https://x.com/kr0der/status/2080126510149853432 https://x.com/kr0der/status/2080126510149853432 Thu, 23 Jul 2026 03:03:01 +0000 When Claude hits usage limits mid-task, screenshot the chat into Codex to resume where it left off without losing context Source: @kr0der https://x.com/petergyang/status/2079989966298616039 https://x.com/petergyang/status/2079989966298616039 Wed, 22 Jul 2026 18:00:26 +0000 Use a 25/50/25 writing workflow: draft manually, then use AI (with an anti-slop skill) to edit the middle, then manually polish the end Source: @petergyang https://x.com/petergyang/status/2079943830024188105 https://x.com/petergyang/status/2079943830024188105 Wed, 22 Jul 2026 14:57:07 +0000 Grab the open-source /no-ai-slop Claude skill to strip 20+ recurring AI cliché patterns (binary contrasts, throat-clearing openers, fake-profound endings) from edited writing Source: @petergyang https://x.com/Sourcegraph/status/2079709820589056079 https://x.com/Sourcegraph/status/2079709820589056079 Tue, 21 Jul 2026 23:27:15 +0000 Comparison of five ways AI coding tools handle codebase context (prompts, indexing, repos, MCP, long-context) with tradeoffs for picking one Source: @Sourcegraph https://x.com/dexhorthy/status/2079658750441586838 https://x.com/dexhorthy/status/2079658750441586838 Tue, 21 Jul 2026 20:04:18 +0000 Treat every agent-written codebase as brownfield from day one—invest in structure early because yolo-prompting breaks down within weeks Source: @dexhorthy https://x.com/Dimillian/status/2079650471250116631 https://x.com/Dimillian/status/2079650471250116631 Tue, 21 Jul 2026 19:31:25 +0000 In Codex Code Review, add repeated reviewer nitpicks as rules in AGENTS.md so the agent enforces them automatically Source: @Dimillian https://x.com/karpathy/status/2079610838143623371 https://x.com/karpathy/status/2079610838143623371 Tue, 21 Jul 2026 16:53:55 +0000 Use voice mode to ramble stream-of-consciousness at an LLM for 10 minutes; it reconstructs your intent better than typing Source: @karpathy https://x.com/Dimillian/status/2079588844450480416 https://x.com/Dimillian/status/2079588844450480416 Tue, 21 Jul 2026 15:26:32 +0000 Kick off ChatGPT cloud tasks from desktop/mobile with configured plugins to run long-running work in the background Source: @Dimillian https://x.com/itsolelehmann/status/2079391997639720989 https://x.com/itsolelehmann/status/2079391997639720989 Tue, 21 Jul 2026 02:24:20 +0000 Use Claude's `find-skills` skill to have Claude search for and install expert-built skills before doing a task, improving output quality Source: @itsolelehmann https://x.com/geoffreylitt/status/2079366818221740529 https://x.com/geoffreylitt/status/2079366818221740529 Tue, 21 Jul 2026 00:44:16 +0000 Prompt template: ask a smart LLM to teach a topic 3blue1brown-style, building intuition in layers and first asking about your background to personalize Source: @geoffreylitt https://x.com/petergyang/status/2079257646939742542 https://x.com/petergyang/status/2079257646939742542 Mon, 20 Jul 2026 17:30:28 +0000 For subjective outputs, use a separate verification agent with a rubric to review the work — avoids self-preferential bias when a model grades its own output Source: @petergyang https://x.com/clairevo/status/2079248971625844819 https://x.com/clairevo/status/2079248971625844819 Mon, 20 Jul 2026 16:56:00 +0000 AI social-post workflow: scan company data for ideas, interview yourself via AI panel, train voice, run an editorial council loop, convert feedback into reusable skills Source: @clairevo https://x.com/petergyang/status/2079209983984677210 https://x.com/petergyang/status/2079209983984677210 Mon, 20 Jul 2026 14:21:04 +0000 Before building, ask Claude to inspect the codebase/plan and surface blind spots and unknown unknowns that could change your approach Source: @petergyang https://x.com/dexhorthy/status/2079016387876692098 https://x.com/dexhorthy/status/2079016387876692098 Mon, 20 Jul 2026 01:31:47 +0000 Default your coding agent to create new git worktrees off the latest remote main so parallel agent runs stay isolated and current Source: @dexhorthy https://x.com/_catwu/status/2079011428380602526 https://x.com/_catwu/status/2079011428380602526 Mon, 20 Jul 2026 01:12:05 +0000 Use Claude Cowork to manage your calendar with a prompt that caps meeting hours, dedupes conflicts, learns your decline patterns, and asks before editing invites Source: @_catwu https://x.com/steipete/status/2078914010871992443 https://x.com/steipete/status/2078914010871992443 Sun, 19 Jul 2026 18:44:59 +0000 Keep agent memory as a plain folder of markdown files — it improves automatically with each new model release, no migration needed Source: @steipete https://x.com/steipete/status/2078907212731400488 https://x.com/steipete/status/2078907212731400488 Sun, 19 Jul 2026 18:17:58 +0000 Have your browser-controlling agent record network traffic to a HAR file, then read it back to debug or reverse-engineer API calls Source: @steipete https://x.com/petergyang/status/2078846124828545179 https://x.com/petergyang/status/2078846124828545179 Sun, 19 Jul 2026 14:15:13 +0000 Don't glaze over an agent's plan — read it carefully; use HTML artifacts to plan and learn, and set up /loop and /goal to run agents for long periods Source: @petergyang https://x.com/thdxr/status/2078727284865827140 https://x.com/thdxr/status/2078727284865827140 Sun, 19 Jul 2026 06:23:00 +0000 Instead of having an agent drive a browser repeatedly, have it record network requests to a HAR file and derive a reusable API client Source: @thdxr https://x.com/kr0der/status/2078703381766631692 https://x.com/kr0der/status/2078703381766631692 Sun, 19 Jul 2026 04:48:01 +0000 In Claude Code, block the deep-research skill via settings.json deny list to prevent it from burning your usage when you mention 'research' Source: @kr0der https://x.com/dexhorthy/status/2078592010852982977 https://x.com/dexhorthy/status/2078592010852982977 Sat, 18 Jul 2026 21:25:28 +0000 Before building a feature with AI, write the customer-facing launch blog post first to surface edge cases and clarify priorities Source: @dexhorthy https://x.com/HamelHusain/status/2078567781621219377 https://x.com/HamelHusain/status/2078567781621219377 Sat, 18 Jul 2026 19:49:11 +0000 In Codex, tell an agent to fan out subtasks into separate threads and tag other threads by name to orchestrate work Source: @HamelHusain https://x.com/petergyang/status/2078502566095474807 https://x.com/petergyang/status/2078502566095474807 Sat, 18 Jul 2026 15:30:03 +0000 To actually learn from AI output, ask Claude to generate an HTML report of its changes and quiz you on it to expose knowledge gaps Source: @petergyang https://x.com/dexhorthy/status/2078223341652189423 https://x.com/dexhorthy/status/2078223341652189423 Fri, 17 Jul 2026 21:00:30 +0000 Have your AI agent ask multiple clarifying questions in one pass so you can voice-ramble answers down the list instead of one-at-a-time Source: @dexhorthy https://x.com/NickADobos/status/2078159728975429646 https://x.com/NickADobos/status/2078159728975429646 Fri, 17 Jul 2026 16:47:44 +0000 Treat agent memory as a curated file library: grow and prune a garden of reusable prompts rather than one giant rulebook Source: @NickADobos https://x.com/cathrynlavery/status/2078153739740012809 https://x.com/cathrynlavery/status/2078153739740012809 Fri, 17 Jul 2026 16:23:56 +0000 Point an agent at your own setup: mine Claude Code/Codex sessions daily for recurring friction and stage fixes to tools, skills, memory, context Source: @cathrynlavery https://x.com/Dimillian/status/2078134928282182021 https://x.com/Dimillian/status/2078134928282182021 Fri, 17 Jul 2026 15:09:11 +0000 Ask Codex to review your past sessions, archive one-off queries, and reorganize projects to keep your workspace tidy Source: @Dimillian https://x.com/petergyang/status/2078125090659811585 https://x.com/petergyang/status/2078125090659811585 Fri, 17 Jul 2026 14:30:05 +0000 Set a daily Codex scheduled task to triage your inbox: list action items, unsubscribe candidates, and prep meeting briefs Source: @petergyang https://x.com/davis7/status/2077905227874013263 https://x.com/davis7/status/2077905227874013263 Thu, 16 Jul 2026 23:56:26 +0000 Route tasks across coding agents by strength: Codex for computer use and grunt work, Fable for planning/copy, Pi for quick edits Source: @davis7 https://x.com/davis7/status/2077892926911475943 https://x.com/davis7/status/2077892926911475943 Thu, 16 Jul 2026 23:07:33 +0000 Skim just ~10 lines of an agent's reasoning trace or diff mid-run to catch wrong directions early and steer it back on course Source: @davis7 https://x.com/davis7/status/2077892444121935974 https://x.com/davis7/status/2077892444121935974 Thu, 16 Jul 2026 23:05:38 +0000 Treat agent code like pottery: let it draft, then reshape by pasting a bad function back, explaining why it sucks, and citing docs to consult Source: @davis7 https://x.com/thdxr/status/2077853471483498800 https://x.com/thdxr/status/2077853471483498800 Thu, 16 Jul 2026 20:30:46 +0000 After a large refactor, ask your coding agent to exhaustively exercise every config/keybinding to surface regressions Source: @thdxr https://x.com/pontusab/status/2077776933761294379 https://x.com/pontusab/status/2077776933761294379 Thu, 16 Jul 2026 15:26:38 +0000 Audit your LLM API usage for prompt caching opportunities — teams have found ~$10k/month in savings from a same-day fix Source: @pontusab https://x.com/cathrynlavery/status/2077524086444658996 https://x.com/cathrynlavery/status/2077524086444658996 Wed, 15 Jul 2026 22:41:55 +0000 Export your ChatGPT and Claude data via Settings, then prompt an LLM to organize/analyze your accumulated AI history Source: @cathrynlavery https://x.com/Dimillian/status/2077463568031854793 https://x.com/Dimillian/status/2077463568031854793 Wed, 15 Jul 2026 18:41:26 +0000 Iterative game-building workflow: generate concept art, refine via chat, get a pitch/vertical slice, then loop prompt-play-prompt on visuals and feel Source: @Dimillian https://x.com/bcherny/status/2077460395279692197 https://x.com/bcherny/status/2077460395279692197 Wed, 15 Jul 2026 18:28:50 +0000 Encode domain knowledge into CLAUDE.md, REVIEW.md, and skills files so agents (and new contributors) can work in your codebase with zero extra context Source: @bcherny https://x.com/NickADobos/status/2077456792788090956 https://x.com/NickADobos/status/2077456792788090956 Wed, 15 Jul 2026 18:14:31 +0000 Brainstorm on a walk in ChatGPT voice mode, then @-attach that voice thread into Codex/desktop as a plan for your coding agent Source: @NickADobos https://x.com/emollick/status/2077431952253112621 https://x.com/emollick/status/2077431952253112621 Wed, 15 Jul 2026 16:35:48 +0000 When you need a small utility, have an AI agent spec, build, and install it for you rather than hunting for existing software Source: @emollick https://x.com/cathrynlavery/status/2077235288389079200 https://x.com/cathrynlavery/status/2077235288389079200 Wed, 15 Jul 2026 03:34:20 +0000 Use the open-source Diagram Design Claude Code skill to generate 27 diagram types as branded HTML, skipping Figma Source: @cathrynlavery https://x.com/davis7/status/2077157941371949285 https://x.com/davis7/status/2077157941371949285 Tue, 14 Jul 2026 22:26:59 +0000 Chain coding agents as subagents: use one model to research/plan, another to implement, then a third to review for simplicity and correctness Source: @davis7 https://x.com/kr0der/status/2077136198162891122 https://x.com/kr0der/status/2077136198162891122 Tue, 14 Jul 2026 21:00:35 +0000 Get an AI agent to 'watch' a video by extracting screenshots at intervals and feeding them in for analysis Source: @kr0der https://x.com/Dimillian/status/2077110043451814066 https://x.com/Dimillian/status/2077110043451814066 Tue, 14 Jul 2026 19:16:39 +0000 In Codex, have it research first and then call set_goal (instead of using /goal upfront) for better task framing Source: @Dimillian https://x.com/clairevo/status/2077094651010449544 https://x.com/clairevo/status/2077094651010449544 Tue, 14 Jul 2026 18:15:30 +0000 Concrete list of AI projects execs can build to learn: morning briefings, meeting-note processors, calendar optimizer, Slack triage, me-as-a-skill, etc Source: @clairevo https://x.com/steipete/status/2076897109060849970 https://x.com/steipete/status/2076897109060849970 Tue, 14 Jul 2026 05:10:32 +0000 In Codex, ask the agent to research first and then call set_goal (instead of /goal) to get better-scoped, more accurate goals Source: @steipete https://x.com/steipete/status/2076886451455992249 https://x.com/steipete/status/2076886451455992249 Tue, 14 Jul 2026 04:28:11 +0000 Use 'stress test' as a prompt to have coding agents probe edge cases and weaknesses in your code or plan Source: @steipete https://x.com/clairevo/status/2076802367626924500 https://x.com/clairevo/status/2076802367626924500 Mon, 13 Jul 2026 22:54:04 +0000 Use Codex/ChatGPT for life admin: dedupe and tag photo libraries, build approval webapps, convert notes to playlists, format docs Source: @clairevo https://x.com/thdxr/status/2076734201760493709 https://x.com/thdxr/status/2076734201760493709 Mon, 13 Jul 2026 18:23:12 +0000 Create a dedicated 'issues' channel where GitHub issues (and even social posts) go, then have an agent reproduce and fix them Source: @thdxr https://x.com/petergyang/status/2076677615583379921 https://x.com/petergyang/status/2076677615583379921 Mon, 13 Jul 2026 14:38:21 +0000 Use a main agent to split work, launch focused sub-agents in parallel, and verify their outputs (screenshots, tests) before you review Source: @petergyang https://x.com/intellectronica/status/2076584134735241413 https://x.com/intellectronica/status/2076584134735241413 Mon, 13 Jul 2026 08:26:53 +0000 Delegate tedious support-chat negotiations to Codex + Chrome — it can patiently handle in-site chat with service providers on your behalf Source: @intellectronica https://x.com/dexhorthy/status/2076381087203475772 https://x.com/dexhorthy/status/2076381087203475772 Sun, 12 Jul 2026 19:00:03 +0000 When models upgrade, audit and remove old skills/prompts that force subagent use or over-advertise—new models pick tools well on their own Source: @dexhorthy https://x.com/Saboo_Shubham_/status/2076380344631398606 https://x.com/Saboo_Shubham_/status/2076380344631398606 Sun, 12 Jul 2026 18:57:06 +0000 The open-source /last30days skill lets an agent pull recent content from X, Reddit, YouTube, arXiv, HN, etc. in one command Source: @Saboo_Shubham_ https://x.com/steipete/status/2076359405000646770 https://x.com/steipete/status/2076359405000646770 Sun, 12 Jul 2026 17:33:53 +0000 Import Wispr Flow's learned dictionary from its local SQLite DB into Codex so your custom vocabulary carries over Source: @steipete https://x.com/gabrielchua/status/2076300026612269105 https://x.com/gabrielchua/status/2076300026612269105 Sun, 12 Jul 2026 13:37:56 +0000 Ask Codex (or coding agents) to ping you when blocked so you can walk away and only return when input is needed Source: @gabrielchua https://x.com/Dimillian/status/2076254387131797640 https://x.com/Dimillian/status/2076254387131797640 Sun, 12 Jul 2026 10:36:35 +0000 ChatGPT Work runs on an always-on cloud computer with your connectors, so tasks continue even when your laptop is offline Source: @Dimillian https://x.com/Saboo_Shubham_/status/2076131184145494349 https://x.com/Saboo_Shubham_/status/2076131184145494349 Sun, 12 Jul 2026 02:27:01 +0000 You can install a Codex plugin inside Claude Code to delegate code reviews or subtasks to Codex without leaving your Claude workflow Source: @Saboo_Shubham_ https://x.com/lucasmeijer/status/2076066050148299234 https://x.com/lucasmeijer/status/2076066050148299234 Sat, 11 Jul 2026 22:08:12 +0000 After any AI coding task, run a follow-up prompt in the same context asking it to review its work, remove dead code, simplify, and consolidate duplicates Source: @lucasmeijer https://x.com/itsolelehmann/status/2075952515762257969 https://x.com/itsolelehmann/status/2075952515762257969 Sat, 11 Jul 2026 14:37:03 +0000 GPT-5.6 tips: state each instruction once, use concrete tone descriptions, set verbosity globally, and dial reasoning effort down one level from prior defaults Source: @itsolelehmann https://x.com/petergyang/status/2075943229430833234 https://x.com/petergyang/status/2075943229430833234 Sat, 11 Jul 2026 14:00:09 +0000 Parallel agent pattern: have one master agent spin up multiple sub-agents on the same task with separate contexts, then judge which solution is best Source: @petergyang https://x.com/emollick/status/2075741472712728924 https://x.com/emollick/status/2075741472712728924 Sat, 11 Jul 2026 00:38:27 +0000 In ChatGPT, type @study to invoke study mode, which acts more like a tutor and improves learning outcomes vs the default assistant Source: @emollick https://x.com/RayFernando1337/status/2075726264082788863 https://x.com/RayFernando1337/status/2075726264082788863 Fri, 10 Jul 2026 23:38:01 +0000 Build a custom Claude skill (e.g. /apple-design) by distilling knowledge from a favorite video/doc series so the model applies it on demand Source: @RayFernando1337 https://x.com/emollick/status/2075695604575903767 https://x.com/emollick/status/2075695604575903767 Fri, 10 Jul 2026 21:36:11 +0000 For research over many files, try NotebookLM instead of ChatGPT Work — it surfaces process and sources rather than just an answer Source: @emollick https://x.com/cathrynlavery/status/2075680629791023504 https://x.com/cathrynlavery/status/2075680629791023504 Fri, 10 Jul 2026 20:36:41 +0000 Set up a dedicated AI 'house/life manager' agent that researches options and pauses for approval before booking to offload personal admin Source: @cathrynlavery https://x.com/github/status/2075629694792409356 https://x.com/github/status/2075629694792409356 Fri, 10 Jul 2026 17:14:17 +0000 GitHub found rewriting Copilot's review instructions—not adding tools—was what improved code review quality; iterate on prompts first Source: @github https://x.com/0xSero/status/2075544637930893622 https://x.com/0xSero/status/2075544637930893622 Fri, 10 Jul 2026 11:36:18 +0000 For shopping research, open a project dir in Codex, give it a Chrome skill, and loop it over listings until it outputs a wiki of options Source: @0xSero https://x.com/xeophon/status/2075468440874143767 https://x.com/xeophon/status/2075468440874143767 Fri, 10 Jul 2026 06:33:31 +0000 Multi-model pipeline: use a planner model to draft the plan, an explorer to deepen it, and a stronger model to implement Source: @xeophon https://x.com/davis7/status/2075301275479224589 https://x.com/davis7/status/2075301275479224589 Thu, 09 Jul 2026 19:29:15 +0000 Route high-volume, well-defined prompts (classification, title-gen, repeatable markdown workflows) to small models; reserve big models for loosely-defined tasks Source: @davis7 https://x.com/emollick/status/2075294040971464959 https://x.com/emollick/status/2075294040971464959 Thu, 09 Jul 2026 19:00:31 +0000 Feed a full book/long PDF to Codex for a 30-min automated proofreading pass to surface dozens of editing notes Source: @emollick https://x.com/amirmxt/status/2075253183618732138 https://x.com/amirmxt/status/2075253183618732138 Thu, 09 Jul 2026 16:18:09 +0000 Build an AI-native workflow: numbered context folders per project (01-COMPANY-OVERVIEW…), persona skills encoding named experts, and a shared Obsidian vault where every agent writes results back so the next agent can read them Source: @amirmxt https://x.com/bcherny/status/2074997570317779038 https://x.com/bcherny/status/2074997570317779038 Wed, 08 Jul 2026 23:22:27 +0000 In Claude Code, run /checkup to prune unused skills/MCPs, dedupe CLAUDE.md files, disable slow hooks, and pre-approve safe commands Source: @bcherny https://x.com/clairevo/status/2074935469381722154 https://x.com/clairevo/status/2074935469381722154 Wed, 08 Jul 2026 19:15:40 +0000 Build a custom agent harness (e.g. with Claude SDK) for a repeat task like bug triage—stops the micromanaging-chat treadmill Source: @clairevo https://x.com/emollick/status/2074930119173460007 https://x.com/emollick/status/2074930119173460007 Wed, 08 Jul 2026 18:54:25 +0000 Use Pinokio to run local open-source AI models and expose them to Claude Code/Codex as a cheap alternative to paid APIs Source: @emollick https://x.com/steipete/status/2074923306596389234 https://x.com/steipete/status/2074923306596389234 Wed, 08 Jul 2026 18:27:21 +0000 Draft essays by brain-dumping into ChatGPT Voice on a walk, then have it transcribe and structure into a first draft Source: @steipete https://x.com/Saboo_Shubham_/status/2074899257652597051 https://x.com/Saboo_Shubham_/status/2074899257652597051 Wed, 08 Jul 2026 16:51:47 +0000 Improve agents by capturing user corrections as labeled examples and feeding them back into context/memory — more valuable than synthetic tests Source: @Saboo_Shubham_ https://x.com/0xSero/status/2074803793804292301 https://x.com/0xSero/status/2074803793804292301 Wed, 08 Jul 2026 10:32:27 +0000 Give a coding agent persistent Chrome/Gmail/Drive/browser tools plus a long-running /goal to grind through multi-country tax prep with human review Source: @0xSero https://x.com/NickADobos/status/2074561088495260091 https://x.com/NickADobos/status/2074561088495260091 Tue, 07 Jul 2026 18:28:01 +0000 Have an agent pre-research all restaurants near you (hours, menus, macros) so you can later query it with natural constraints like time, distance, and diet Source: @NickADobos https://x.com/petergyang/status/2074527669883162829 https://x.com/petergyang/status/2074527669883162829 Tue, 07 Jul 2026 16:15:14 +0000 Use an AI agent as a chief of staff: at week start, have it review your calendar to consolidate meetings into focus blocks and flag ones that could be a DM Source: @petergyang https://x.com/gabrielchua/status/2074505580207862185 https://x.com/gabrielchua/status/2074505580207862185 Tue, 07 Jul 2026 14:47:27 +0000 For long-running Codex tasks, take a screenshot of the current state and ask Codex to incorporate it to keep the agent oriented and steerable Source: @gabrielchua https://x.com/geoffreylitt/status/2074501584596226505 https://x.com/geoffreylitt/status/2074501584596226505 Tue, 07 Jul 2026 14:31:34 +0000 Walk-driven development: dictate a long audio note on a walk, then have an agent turn it into docs/tasks and kick off coding agents Source: @geoffreylitt https://x.com/cathrynlavery/status/2074289056452878656 https://x.com/cathrynlavery/status/2074289056452878656 Tue, 07 Jul 2026 00:27:04 +0000 Prompt template to audit agent session logs across machines, extract recurring failures via SSH/jq, and surface the single highest-leverage fix Source: @cathrynlavery https://x.com/HamelHusain/status/2074242605634949306 https://x.com/HamelHusain/status/2074242605634949306 Mon, 06 Jul 2026 21:22:29 +0000 For AI evals: don't ask the AI to 'find issues'—build a review UI, annotate traces into a failure taxonomy, and re-review data periodically Source: @HamelHusain https://x.com/petergyang/status/2074206798631071796 https://x.com/petergyang/status/2074206798631071796 Mon, 06 Jul 2026 19:00:12 +0000 Five copy-paste prompts for a top-tier model: surface high-leverage tasks, get a business assessment, pre-ship bug hunt, plan a big project, and refactor your AI skills library Source: @petergyang https://x.com/gabrielchua/status/2074158639791948273 https://x.com/gabrielchua/status/2074158639791948273 Mon, 06 Jul 2026 15:48:50 +0000 In Codex desktop, double-tap ⌘⌘ to send an 'Appshot' capturing the frontmost window plus its full text and off-screen content as richer context Source: @gabrielchua https://x.com/emollick/status/2074137233607590328 https://x.com/emollick/status/2074137233607590328 Mon, 06 Jul 2026 14:23:46 +0000 Push AI models with maximally ambitious asks first to find their ceiling, then dial requests back — don't underassign Source: @emollick https://x.com/petergyang/status/2074135084324774312 https://x.com/petergyang/status/2074135084324774312 Mon, 06 Jul 2026 14:15:14 +0000 Prototype UI faster than code: screenshot the current UI, ask Image Gen for 4-5 variants, then turn a chosen mock into a working prototype Source: @petergyang https://x.com/altryne/status/2073950674543600019 https://x.com/altryne/status/2073950674543600019 Mon, 06 Jul 2026 02:02:27 +0000 AI code review routing: READ critical code (auth/money), decompose big PRs, verify with tests/evals, and never let the writing agent also review Source: @altryne https://x.com/itsolelehmann/status/2073838079899537774 https://x.com/itsolelehmann/status/2073838079899537774 Sun, 05 Jul 2026 18:35:02 +0000 Claude/Fable prompting: delete old verbose instructions, give it your hardest problem, ask for proof before 'done', use a second fresh agent to review Source: @itsolelehmann https://x.com/HamelHusain/status/2073807874036543728 https://x.com/HamelHusain/status/2073807874036543728 Sun, 05 Jul 2026 16:35:01 +0000 Enable always-on remote control for Claude Code by setting remoteControlAtStartup:true in ~/.claude/settings.json Source: @HamelHusain https://x.com/_catwu/status/2073806626965049686 https://x.com/_catwu/status/2073806626965049686 Sun, 05 Jul 2026 16:30:03 +0000 Use Claude Code to source candidates: describe the role, have it find 100 with links/pitches, output an artifact emailed to you to review later Source: @_catwu