Tool Lab
- , scriptc HIDDEN GEM
(Vercel Labs; surfaced by Guillermo Rauch; updated Sun July 26). Compiles ordinary TypeScript (type-checked by the real tsc) to a native binary with no Node and no V8 engine inside. Rauch's numbers compiling the Vercel CLI: 1.28MB binary, 1.5ms startup, compiled in 2.94s. Why: every throwaway TS script or internal CLI becomes a single file a teammate runs with nothing installed. macOS is primary; Linux/Windows by cross-compile. First step: go to scriptc.dev, compile one script you already have, hand the binary to someone with no Node.
Featured Jul 27, 2026 · that episode - , the recall safety net TECHNIQUE
(from CodeRabbit's Opus 5 numbers). Opus 5's reviewer recall dropped ~6 points, it's precise but misses more, so don't let it review alone. Run Opus 5 for the sharp pass, then a second recall-oriented model (Fable 5 or Sonnet 5) whose only job is to catch what Opus skipped. First step: wire it as a pre-commit step, Opus writes, the second model sweeps, then you merge.
Featured Jul 27, 2026 · that episode - , the research-folder / AGENTS.md pattern TECHNIQUE
(Guillermo Rauch, Vercel). Do all research with an agent CLI + the filesystem: a `research/` folder, an `AGENTS.md` inside describing your preferred format and best practices, then launch the agent and ask. It finds and correlates knowledge from previous sessions; scales infinitely because it's just files; sync via iCloud/git across machines. No app, no knowledge graph, no UI. First step: make the folder, write one AGENTS.md with your output format, point your CLI at it. "The factory is the product."
Featured Jul 26, 2026 · that episode - , T3 Code HIDDEN GEM
(Theo / Ping). Free, open-source GUI for Claude Code and Codex that runs on the subscription you ALREADY pay for, no extra bill, no key to buy. 76 PRs merged in a week: Opus 5 support, Auto mode for approvals, Claude Code skills in the composer, worktree isolation, parallel multi-agent threads. First step: `npx t3@nightly`, drive two agents at once on a repo you know.
Featured Jul 26, 2026 · that episode - , Opus 5 effort dialing TECHNIQUE
set hard-reasoning routes to xhigh, NOT max. Anthropic's launch charts (flagged by Alex Albert) show xhigh beats every other model on the board while using ~25% fewer output tokens than max. One-line config change that pays every request. Credit: Anthropic / Alex Albert.
Featured Jul 25, 2026 · that episode - , Dari's open-sourced router model HIDDEN GEM
(daridotdev): cache-aware, per-step routing, 79.8% Terminal-Bench 2.1 for $76. Clone it, point at your model set, route a high-volume job through it. Credit: Dari / kimmonismus.
Featured Jul 25, 2026 · that episode - , the dispatcher pattern TECHNIQUE
(Offloop's D1; surfaced by TestingCatalog + kimmonismus; built by the team behind Manus). Most multi-agent setups fail because every agent acts at once. Add one tiny cheap router that, at each step, decides which agent moves next, when to stop, and when to escalate to a human (in their overnight run, D1 escalated two stuck tasks to a person at 2am rather than burning money). Copy it in Claude Code: one "dispatcher" subagent that never does the work, only picks the next agent or flags you; workers run on a cheap model and act only when called. Receipts: a 220-task benchmark run cost $363 vs $3,100 on Claude Code; 84.9% on GDPval, ~1/5 the cost of rival multi-agent systems.
Featured Jul 24, 2026 · that episode - , Atomic Agent HIDDEN GEM
(AtomicBot-ai, MIT). Local-first open-source agent: llama.cpp inference, MCP for tools, SQLite persistent memory, and time-travel file history (every touched file snapshotted before/after, one-click diff or restore). No token bill. First step: clone, point it at a local model via llama-server, hand it a small repo task, use snapshots as your undo button.
Featured Jul 24, 2026 · that episode - , Gemini 3.5 Flash Cyber TRENDING
(Google, July 22). A small, cheap model built specifically to find and patch critical software vulnerabilities; live on AI Studio + Vertex. Why: the Claude Security plugin is your deep in-terminal audit; Flash Cyber is cheap enough to run on EVERY change. First step: schedule an agent that runs Flash Cyber over each PR and posts findings. Credit: Google.
Featured Jul 23, 2026 · that episode - , the two-model security pair TECHNIQUE
(framing: Nathan Lambert). Don't review code with the model that wrote it: closed frontier models refuse a lot of security work and miss things. Have your builder model write, then send the diff to a DIFFERENT model (Flash Cyber, or an open model like Kimi) whose only job is to attack it. First step: add one Claude Code pre-commit hook that fires the second, adversarial model automatically on every commit. Two brains (one adversarial) beat one brain grading its own homework.
Featured Jul 23, 2026 · that episode - , Strip the scaffolding TECHNIQUE
(Simon Willison, w/ Cat Wu + Thariq Shihipar, Claude Code team). Stop overloading prompts with examples and "don't do X" lists; Anthropic cut Claude Code's own system prompt by ~80%. Examples constrain the model to "make it look like this," and negative-rule walls narrow it onto the very failure you feared. First step: open your fattest CLAUDE.md/system prompt, delete the example blocks and negative rules, replace with one clear goal + definition of done, then re-run your hardest task. Usually better on Fable 5 / Opus 4.8.
Featured Jul 22, 2026 · that episode - , claude-in-mobile HIDDEN GEM
(Alex Gladkov). Open-source MCP server: drives Android via ADB, iOS Simulator via simctl, and desktop apps, the cross-platform version of what Anthropic shipped iOS-only today. First step: clone, wire it as an MCP server, point Claude at a running Android emulator, have it walk one screen flow (install/tap/screenshot). Android builders don't have to wait for the official feature.
Featured Jul 22, 2026 · that episode - , planner/worker swarm split TECHNIQUE
(Cursor team). What: define one Claude Code subagent as PLANNER (reads spec, builds a task tree, delegates, never codes); worker subagents on a cheaper model (Sonnet 5 / Kimi K3) each complete one leaf and push; grade every worker output against a held-out test the workers never saw. Why: this structure turned a $20K job into $1.3K in Cursor's run. First step: clone the reference harness off the Cursor blog, or wire the split with existing Claude Code subagents. Credit: Cursor team.
Featured Jul 21, 2026 · that episode - , Seedream 5.0 Pro on BytePlus Lumina HIDDEN GEM
(ByteDance). Not one-shot image gen, editable design: layer separation (splits a poster into text/subject/background/decorations as assets), region edits via point/lasso/box selection, 14-language text. Surfaced this week by kimmonismus + Min Choi. First step: drop one marketing asset in, select just the headline, edit it while everything else stays pixel-locked. Replaces the designer round-trip for small changes.
Featured Jul 21, 2026 · that episode - , Devin for Startups TRENDING
(Cognition, surfaced by swyx). $65k in credits to run Devin across Cloud, Desktop, and CLI. Why: a real budget to offload background engineering. First step: early-stage founders apply this week, then hand Devin a backlog of small, well-scoped fixes to grind in the Cloud while you build. Credit: Cognition.
Featured Jul 20, 2026 · that episode - , Multi-model Claude Code setup HIDDEN GEM
(Theo / T3). A config pattern, not a product: run GPT-5.6 Sol and Kimi K3 as the underlying model INSIDE Claude Code, keep the harness (skills, browser, file editing), swap the brain per task. Why: Claude Code is the best agent harness, but you don't have to pay frontier rates for every keystroke. First step: copy Theo's config, route routine coding to K3/Sol, reserve Fable for hard reasoning. Credit: Theo (T3).
Featured Jul 20, 2026 · that episode - , Anthropic Workbench, rebuilt Build section + Fable 5 fallback TRENDING
(Anthropic). Send a Messages API request, see Claude's response, copy shippable code in-place. New today: configure a fallback model for Fable 5, a blocked (safety-classifier) or failed Fable request auto-reruns on Opus 4.8 instead of dying. First step: open the Console → Build, set a fallback under your primary model; or on the API, pass the `fallbacks` parameter. ~90 seconds; a blocked request never strands you mid-session.
Featured Jul 19, 2026 · that episode - , Cerebras-hosted GPT-5.6 Sol at 750 tok/s HIDDEN GEM
(Cerebras). Wafer-scale inference serving frontier Sol at up to 750 tokens/sec vs ~40-120 tok/s on GPU clusters (~10x). Rolling to customers now; hype carried by kimmonismus ("projects finished 10x faster is worth premium pricing"). First step: route only your latency-bound work (live agent chat, interactive iteration, autocomplete-style loops) to Cerebras Sol; keep hard reasoning on Claude. You're buying the feel of instant, not more IQ. Credit: Cerebras.
Featured Jul 19, 2026 · that episode - , Gemini Notebook code execution TRENDING
(Google, July 16). NotebookLM renamed to Gemini Notebook, and every notebook now gets a secure cloud computer: upload data, it writes and runs Python in an isolated container and returns real charts/tables grounded in your sources (not prose). Live for AI Ultra + Workspace AI Ultra/Expanded; rolling to Pro on web soon. First step: drop one ugly CSV in, ask it to clean the data and find the pattern. CAVEAT: output looks authoritative even when the method is wrong, make it show the code before you trust the chart. Credit: Google.
Featured Jul 18, 2026 · that episode - , Kimi Code CLI HIDDEN GEM
(Moonshot). Everyone tested the K3 model; almost nobody wired the harness it runs in, and that harness is what scored joint #5 on the Artificial Analysis Coding Agent Index. Shares an API key with Fable 5 on the AI/ML API, 1M-token context, a fraction of frontier output price. First step: install Kimi Code, point it at K3, run one boilerplate-heavy job vs your Claude baseline, log cost + quality; you get a cheap high-volume route wired before the July 27 open weights. Credit: Moonshot / Artificial Analysis.
Featured Jul 18, 2026 · that episode - , AgentMail via Vercel TRENDING
(AgentMail team, surfaced by Guillermo Rauch). One command gives your coding agent its own real email inbox: no signup, automatic setup, unified billing. Why: your agent can now do email-gated work (catch a confirmation, receive a webhook, email you a finished build). First step: `vercel install agentmail`, then tell your agent to email you a summary when its next long run ends.
Featured Jul 17, 2026 · that episode - , Anthropic fallback model / fallbacks parameter HIDDEN GEM
(Anthropic; Fable-5 fallback per TestingCatalog). In the upgraded Console Build workbench, set a fallback model under your primary; or pass the `fallbacks` parameter on the API. Any refused/failed request silently re-runs on the model you name, at that model's rate. With Fable metering and dropping out, this is the resilience move of the week (~90 seconds). Point it at Sonnet 5.
Featured Jul 17, 2026 · that episode - , grok-build source HIDDEN GEM
(xAI; teardown by Simon Willison). 844,530 lines of readable Rust: the definitive reference for how an agent harness loads/invokes skills, plugins, hooks, MCP servers, subagents. Runs fully local-first (compile, point at your own inference, drive from config.toml). First step: clone, read `xai-grok-tools` before anything else to see shell-command dispatch. Credit: xAI / Simon Willison.
Featured Jul 16, 2026 · that episode - , Inkling TRENDING
(Thinking Machines / John Schulman team). First major open-weights model under Apache 2.0; 975B total / 41B active, 1M context, multimodal (text/image/audio); 77.6% SWE-bench Verified. Killer feature: controllable thinking-effort dial, matches Nemotron 3 Ultra on Terminal-Bench 2.1 at ~1/3 the tokens. On Databricks via Unity AI Gateway; connects to Cursor and OpenCode. First step: point Cursor/OpenCode at Inkling, set effort low, route boilerplate there. Credit: Thinking Machines.
Featured Jul 16, 2026 · that episode - context-quarantine pruner HIDDEN GEM
(via GetBindu/awesome-claude-code-and-skills). Reads your real session transcripts to find skills, MCP servers, and agents loaded into context but never fired, then safely quarantines them. Cross-agent (Claude Code, Codex, Hermes, OpenCode, Cursor, OpenClaw), zero telemetry, single static Go binary; scan/report/prune keeps removals reversible; Homebrew + npm, MIT. First step: install, run scan on your heaviest repo, prune skills that never fire. Why now: leaner context = cheaper every message = stretches your Fable allowance to Sunday. Pairs with `/checkup` but goes cross-agent.
Featured Jul 15, 2026 · that episode - last30days-skill TRENDING
(mvanhorn). AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web, then synthesizes a grounded, recency-weighted summary. NEW: free arXiv + Techmeme sources (zero keys; first-run setup installs their CLIs). Install: `npx skills add mvanhorn/last30days-skill` in Claude Code, or the one-click .mcpb bundle in Claude Desktop.
Featured Jul 15, 2026 · that episode - education-agent-skills HIDDEN GEM
(Gareth Manning). Open-source library of 165 evidence-based pedagogical skills across 20 domains; works in Claude Code, Claude.ai (via MCP), OpenAI Codex, Hermes. Why it's here for ALL builders: authoring discipline — every skill grounded in NAMED research, empirically-weak frameworks deliberately excluded and documented in EXCLUSIONS.md. First step: clone it, open one SKILL.md, study the tight scoping so it only fires when relevant; copy the structure for your own domain. Credit: Gareth Manning.
Featured Jul 14, 2026 · that episode - Atomic Chat DFlash TRENDING
(AtomicBot-ai). Local inference app exposing an OpenAI-compatible server at (drop any agent/CLI/IDE on top). NEW: DFlash block-diffusion speculative decoding for Qwen 3.6, Gemma 4, Kimi K2.5 — small drafter proposes token blocks, full model verifies, output identical. Independent bench: ~80–83 tok/s vs ~38–40 baseline on Qwen 3.6 27B (one reviewer). CATCH: Apple Silicon only for now; can't combine with MTP; use the Q8 drafter (Q4 collapses acceptance). First step: on an M-series Mac, install Atomic Chat, enable DFlash w/ Q8 drafter, point Claude Code at localhost:1337 — ~2x local throughput, byte-for-byte same output. Credit: AtomicBot-ai.
Featured Jul 14, 2026 · that episode - /checkup TRENDING
(Anthropic / Boris Cherny, July 8). What: one-command setup audit + confirmable fix for context bloat, stale version, hooks, permissions. First step: `claude --version` to confirm you're current (or let /checkup update you), then type `/checkup` in your biggest repo and read the proposed plan before approving. Changes your day: a leaner, faster, cheaper session on every message after.
Featured Jul 13, 2026 · that episode - Cursor v3.11 side chats HIDDEN GEM
(Cursor team, July 10). What: spin up a parallel agent conversation that inherits main-chat context, runs read/search/answer by default, is durable, and can be @-mentioned back into the main thread — each can even use a different (cheaper) model. First step: update Cursor, type `/side` (or `/btw`, or the + at the top of the chat panel), ask your tangent, then @-mention it back. Why: stop killing a long, expensive main run just to ask a clarifying question. (Note: afterAgentResponse/stop cloud hooks may not fire correctly yet in v3.11.)
Featured Jul 13, 2026 · that episode - llm-router HIDDEN GEM
(Yali Pollak, via awesome-claude-code). A local-first router that sits under Claude Code (and Codex/Gemini CLI) and sends each prompt to the cheapest capable model, with three-layer token compression and automatic provider fallback — protects premium quota, zero config on a Claude subscription. Credibly engineered: 1,900+ tests, independent RouterArena placement, local-only (no hosted proxy). Why now: stretches your 50% Fable allowance to July 19 by auto-dropping trivial work to cheap models. Credit: Yali Pollak.
Featured Jul 12, 2026 · that episode - claude-statusbar TRENDING
(leeguooooo, via awesome-claude-code). The most complete Claude Code status line: 5-hour and 7-day rate-limit usage with reset countdowns and learned end-of-window projections, context window, prompt-cache-expiry countdown, per-session cost, live todo/tool/git activity; 3 styles, 9 themes; PyPI, 320+ tests, sub-1%-CPU daemon. Why now: Fable draws from a shared pool and burns it faster — a live reset countdown lets you pace instead of hitting a wall mid-session. Credit: leeguooooo.
Featured Jul 12, 2026 · that episode - Cursor side chats + Cmd+F transcript search TRENDING
(Cursor, July 10). Why: run parallel investigations without interrupting your main agent; find past chats by content, not just names/PR numbers. First step: update Cursor, type /side (or hit + at the top of the chat panel), ask your question, then @-mention the side chat to bring its answer into the main thread. Changes your day: no more killing a long run to ask a clarifying question. Credit: Cursor team.
Featured Jul 11, 2026 · that episode - Anthropic dual-use "GRAM" research HIDDEN GEM
(AE Studio × Anthropic, July 8). Not a product — a lens for builders shipping on filtered/safety-sensitive workflows. GRAM (Gradient-Routed Auxiliary Modules) adds removable per-category compartments so dangerous knowledge can be isolated without hurting general performance; preliminary, not in production. Builder takeaway (not a keyboard step): expect model access to increasingly be gated by capability category, not just per-request classifiers — design fallbacks accordingly. Credit: AE Studio, Anthropic (Cem Anil, Alex Cloud).
Featured Jul 11, 2026 · that episode - Tencent Hy3 TRENDING
(Tencent Hunyuan Hy Team). 295B MoE, 21B active/token, 256K context, Apache 2.0; free on OpenRouter (tencent/hy3:free) through July 21. Serving reliability is the real story: SWE-bench variance within a few points across Claude Code-style harnesses, Cline, and KiloCode. Ceiling: GLM-5.2 beats it across the agentic coding suite (per Tencent's own appendix) — keep hard reasoning on Claude. FIRST STEP: point one high-volume coding job (boilerplate/scaffolds/refactors) at the free OpenRouter route, reasoning_effort=high on hard steps; compare quality + cost vs your Claude baseline before July 21.
Featured Jul 10, 2026 · that episode - GPT-5.6 multi-agent beta HIDDEN GEM
(OpenAI Responses API). One instance coordinates subagents on independent workstreams and synthesizes findings in a single request — less orchestration glue to maintain.
Featured Jul 10, 2026 · that episode - Grok 4.5 in Cursor TRENDING
(SpaceXAI / Cursor). Cheapest near-frontier coding agent right now: Grok Build task $2.49 vs Fable 5 in Claude Code $11.80 and GPT-5.5 in Codex $5.07. Free for a limited window in Grok Build + Cursor. FIRST STEP: open Cursor, switch model to grok-4.5, re-run one real multi-file task vs your Claude baseline, log gap + cost. (Not in EU at launch.)
Featured Jul 9, 2026 · that episode - llm-coding-agent 0.1a0 HIDDEN GEM
(Simon Willison). A minimal, fully-readable coding agent on top of the `llm` CLI. Install/run: `uvx --prerelease=allow --with llm-coding-agent llm code`. Implements just edit_file (exact-string replace, returns a diff) + execute_command. Why: every rented agent (Claude Code, Codex, Cursor) is this same loop with more scaffolding — read it end-to-end and the category stops being magic. Credit: Simon Willison.
Featured Jul 9, 2026 · that episode - sqlite-utils 4.0 HIDDEN GEM
(Simon Willison, July 7): first major bump since 2020 (124th release). The lesson isn't migrations; it's METHOD. Willison used his remaining Fable window to get to a comfortable stable release: one prompt in Claude Code for web on iPhone ("final review before shipping a stable 4.0 — spot breaking changes") surfaced 5 "release blockers," incl. a silent-write-discard bug. 37 prompts / 34 commits / ~$149.25. MOVE this week: before any breaking change, have Fable run a pre-release audit that hunts for anything forcing a bad major version later — while it's still included. Credit: Simon Willison.
Featured Jul 8, 2026 · that episode - skills.sh directory TRENDING
(via Vercel eve; Guillermo Rauch): community + Vercel skill directory. The CLI is agent-aware: run it and it auto-detects your project, offering to install into an eve agent OR your local AI coding agent (Cursor/Codex/Claude Code).
Featured Jul 8, 2026 · that episode - Dev Browser TRENDING
(Sawyer Hood, via awesome-claude-code): browser-automation plugin/skill letting Claude Code drive a browser to test and verify its own work. Full Playwright API + pixel/DOM-level computer-use tools; connects to your running Chrome or fresh Chromium; scripts run in a QuickJS WASM sandbox (no host filesystem/network). Benchmarked beating Playwright MCP on time, cost, and turns.
Featured Jul 7, 2026 · that episode - jacobian-lens HIDDEN GEM
(Anthropic; Wes Gurnee, Nicholas Sofroniew, Jack Lindsey + team): open-source J-lens implementation (Apache-2.0, Python) + Neuronpedia demo on open-weights models. Detects evaluation-gaming, silent misaligned reasoning, and prompt-injection ("injection"/"fake" appear in J-space) BEFORE output.
Featured Jul 7, 2026 · that episode - Claude Code fallbackModel + org defaults FRESH
Configure up to 3 fallback models tried in order when the primary is overloaded/unavailable; `--fallback-model` now applies to interactive sessions. Org admins can set the default model across chat, Cowork, and Claude Code so routine work doesn't default to the priciest option. `claude --version` to confirm you're current, then set your chain.
Featured Jul 6, 2026 · that episode - Simon Willison's lower-power-subagent instruction HIDDEN GEM
(July 3). Add to CLAUDE.md: "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent." Implementation spawns to a cheap model; trivial edits go lower; design/audit/judgment stays in the main loop. Willison: getting a ton done, Fable allowance shrinking slower. Credit: Simon Willison.
Featured Jul 6, 2026 · that episode - Condense TRENDING
(condense.chat team): drop-in compression proxy, no key swap, byte-exact skeleton + code preserved. Install `curl -fsSL bash`, run one long Claude Code session through it, read savings off the dashboard, keep only if byte-exact quality holds. 100M free tokens to benchmark.
Featured Jul 5, 2026 · that episode - OpenWiki HIDDEN GEM
(LangChain / Brace Sproul): CLI that generates+maintains an agent-consumable repo wiki (in `openwiki/`) and appends a reference to your AGENTS.md/CLAUDE.md so the agent retrieves context on demand instead of carrying a bloated instruction file. Add the included GitHub Action (`openwiki-update.yml`) to open a docs-update PR once a day. Supports OpenRouter, Fireworks, Baseten, OpenAI, Anthropic. (Named Sat; today's add is the setup.)
Featured Jul 5, 2026 · that episode - pxpipe HIDDEN GEM
(teamchong): run as a local proxy in front of Claude Code, keep default models (Fable 5 / GPT-5.6), open the dashboard. Per-request savings logged vs a free count_tokens counterfactual in ~/.pxpipe/events.jsonl — re-derive the number yourself. Route byte-exact tasks to a text subagent (imaging off).
Featured Jul 4, 2026 · that episode - Claude Science PICK
(trending): AI workbench for scientists, beta on macOS/Linux (Pro/Max/Team/Enterprise). 60+ scientific databases; every figure bundles exact code + environment + description + full history. Grant program: 50 projects, up to $30K credits, apply through July 15.
Featured Jul 3, 2026 · that episode - K-Dense scientific-agent-skills TRENDING
#1 science skills library (~140 skills + 100+ databases, 160k+ users), works with Claude Code / Cursor / Codex. Install: `/plugin marketplace add K-Dense-AI/claude-scientific-skills`, or reach all skills via their MCP server. Only install the Python packages you need (each SKILL.md lists its own). Credit: K-Dense AI.
Featured Jul 2, 2026 · that episode - K-Dense claude-skills-mcp HIDDEN GEM
MCP server that semantic-searches + progressively loads Agent Skills so your agent finds the right skill on demand instead of you hand-wiring each. Lightweight 15MB frontend starts instantly (no Cursor timeout); 250MB backend loads in background; ~90 skills from Anthropic's official repo + K-Dense's science collection out of the box; no API keys. Install: `uvx claude-skills-mcp` (or drop the mcpServers block into config). Credit: K-Dense AI.
Featured Jul 2, 2026 · that episode - shot-scraper video FRESH
(Simon Willison, shot-scraper 1.10, Jun 30): a `storyboard.yml` file + Playwright records a video of a routine against your web app. Your coding agent can write the storyboard and film its own demo. Install: `pip install shot-scraper`. The `--help` output doubles as a SKILL.md so agents drive it directly. Use: auto-generate a client-ready demo video at the end of every build.
Featured Jul 1, 2026 · that episode - claude-api skill HIDDEN GEM
(ClaudeDevs, shipped with Sonnet 5): drop-in Claude Code skill that tunes prompts for Sonnet 5, recommends effort levels, and configures advisor mode. Because effort level + token count decide your Sonnet 5 bill, letting the skill pick the effort setting is the difference between a cost cut and a cost surprise. Drop the skill folder in your project.
Featured Jul 1, 2026 · that episode - ClinePass TRENDING
(Cline team): route high-volume/low-stakes work (boilerplate, scaffolds, refactors) to GLM 5.2 / Kimi K2.7 inside Cline; keep Opus for hard reasoning. Install CLI, set ClinePass as provider, run a real multi-file refactor, measure vs your Claude baseline. Cline: GLM 5.2 with medium reasoning moves meaningfully higher on Terminal-Bench.
Featured Jun 30, 2026 · that episode - GitNexus HIDDEN GEM
(Abhigyan Patwari): zero-server code-intelligence engine. Client-side knowledge graph + Graph RAG agent; drop in a repo/ZIP, get an interactive graph. CLI runs fully local (no network calls; index in gitignored `.gitnexus/`); web runs in-browser (no code uploaded). Runs as an MCP server (`gitnexus-mcp`) so Claude Code/Cursor/Codex query structure instead of grepping. Gives even small models full architectural clarity. Install: `git clone`, `npm install`, `npm run dev`, drag a ZIP into localhost.
Featured Jun 30, 2026 · that episode - BrowserBC TRENDING
(ViDA team): record one human web flow, distill to a reusable skill, run on a cheaper model. Clone, record a real flow (vendor login + invoice export), let cheap models replay. Pay frontier prices once at record time, pennies per execution after.
Featured Jun 29, 2026 · that episode - browse.sh HIDDEN GEM
(Browserbase): open catalog of 100+ curated browser skills, install with one CLI command. Install: `npm i -g browse`. In Claude Code: `bb skills` to install reusable workflows; also works in Cursor/Codex. BrowserBC makes skills; browse.sh is the library you pull from. Use both this week.
Featured Jun 29, 2026 · that episode - DESIGN.md TRENDING
(Google Labs, ~19k stars). A format spec describing a visual identity to coding agents: YAML token front matter (exact values) + markdown prose (the why). Drop a DESIGN.md in your repo root with real colors/type/spacing; tell Claude to build to it. Validate with the CLI: `npx @google/design.md` — checks token refs, WCAG contrast, emits JSON agents can act on. Credit: google-labs-code.
Featured Jun 28, 2026 · that episode - awesome-design-md HIDDEN GEM
(VoltAgent). Ready-made DESIGN.md files reverse-engineered from real brands (Stripe, Airbnb, Coinbase, Linear-style tools, dozens more). Copy one in, tell the agent "build a page that looks like this," get consistent UI from analyzed tokens + rules, not surface guesses. When a client says "make it feel like Stripe," hand the agent the file instead of prose. Credit: VoltAgent.
Featured Jun 28, 2026 · that episode - davila7/claude_subagents PICK
(Daniel Avila, updated ~June 22): drop a markdown agent file in `.claude/agents/`. Each subagent runs in its own context window and returns only the result, so grep/read noise never pollutes your main thread. Frontmatter: name, description (Claude matches tasks to it), tools, model, system prompt. Key trick: a subagent can spawn children only if `Agent` is in its `tools` list.
Featured Jun 27, 2026 · that episode - VILA-Lab/Dive-into-Claude-Code HIDDEN GEM
a systematic teardown of Claude Code internals (not on trending). Surfaces: orchestration script fans out to up to 1,000 subagents with state held outside the context window; agent teams in plan mode cost ~7x tokens; subagent summary-only returns prevent context blow-up; auto-approve rates climb 20%→40%+ with experience (a rubber-stamp warning).
Featured Jun 27, 2026 · that episode - LLM PICK
(Simon Willison): CLI + Python lib for running prompts against any model (OpenAI, Anthropic, Gemini, local). Now does tool calling. Install: `pip install llm`; `llm keys set openai`; `cat myfile.py llm -s "Explain this code"`. The Unix pipe for LLMs; pairs with local-model plugins for a frontier-outage fallback.
Featured Jun 26, 2026 · that episode - CodexBar HIDDEN GEM
(Peter Steinberger / steipete): macOS menu-bar app showing usage + reset windows for Codex, Claude, Cursor, Gemini, Copilot, Grok, OpenRouter + more. Per-provider session/weekly/monthly countdowns so you stop guessing whether to start a long job. One status item per provider; spend + credit scans.
Featured Jun 26, 2026 · that episode - Context/output compressor TRENDING
compresses tool outputs, logs, files, and RAG chunks before they hit the LLM; claims 60-95% fewer tokens for the same answers. Ships as library, proxy, or MCP server. How-to: run it as a drop-in proxy endpoint, point one token-heavy workflow at it, measure tokens before/after, keep if quality holds.
Featured Jun 25, 2026 · that episode - Code-intelligence MCP server HIDDEN GEM
indexes a codebase into a persistent knowledge graph (avg repo in milliseconds), 158 languages, sub-ms queries, ~99% fewer tokens, single static binary, zero deps, 100% local; auto-syncs on code changes; works with Claude Code, Codex, Gemini, Cursor, OpenCode and more. How-to: drop in the binary, point it at your repo root, connect as an MCP server in your editor so your agent reads a map instead of grepping file-by-file. (Same GitHub Trending source as above)
Featured Jun 25, 2026 · that episode - Bumblebee HIDDEN GEM
(Perplexity AI): read-only supply-chain scanner for npm, PyPI, Go modules, RubyGems, Composer, MCP servers, editor + browser extensions. Requires Go 1.25+, zero non-standard deps. Install Go, `go install` the binary, run against your project root before installs land. Caveat: pre-1.0, treat as fast first-pass audit. (via Firecrawl's 2026 repo roundup)
Featured Jun 24, 2026 · that episode - agent guardrails CLAUDE.md PICK
(Forrest Chang): single-file Claude Code behavioral skill encoding Karpathy's four rules — stops silent wrong assumptions, over-engineering (50→500 lines), and orthogonal edits. 156k stars. Install = drop the file in your repo root.
Featured Jun 24, 2026 · that episode - anthropics/skills PICK
(trending, +519 stars/day): Anthropic's public repo of agent "skills" for Claude. A skill = a folder with a SKILL.md describing a trigger + steps (+ optional scripts). Clone, drop a skill folder in, and turn tribal knowledge into version-controlled, reusable capability.
Featured Jun 23, 2026 · that episode - simonw/files-to-prompt HIDDEN GEM
(by Simon Willison): concatenates a whole directory into one clean prompt. Install: `pip install files-to-prompt`. Use: `files-to-prompt ./your-project -c` then pipe into your model (pairs perfectly with `llm`). Collapses copy-paste gymnastics into one command for cross-file review/refactors.
Featured Jun 23, 2026 · that episode
Every story, number, and link in your inbox.
The written brief from each episode, free, every morning.