Daily How-To
Short, practical how-to guides you can apply to your AI workflow today.
Claude's prompt cache fails silently below 1,024 tokens
Claude's prompt caching needs a minimum prefix (512–4,096 tokens by model); below it, no cache forms and no error is
Claude Code resends full context every turn — Headroom cuts it
Headroom compresses Claude Code's tool outputs, JSON, and logs before LLM send — 15–20% fewer tokens per session.
OpenMontage has 45k stars and no proprietary orchestrator
OpenMontage routes video production through your coding agent: 12 pipelines, 700+ skill files, no API keys required.
Penpot's MCP server runs on five tools — Figma's uses dozens
Penpot's official MCP server exposes design tokens, component trees, and styles to AI agents via just five endpoints —
Agent-Reach absorbed Bilibili's 412s — your agent kept working
Agent-Reach v1.5.0 routes Claude Code to 13 channels — Jina, yt-dlp, gh CLI, feedparser — MIT-licensed, Python 3.10+.
Variant Claude Code design skill close design-quality gap vibe-coded websites repo setup a
AI coding agents will hand you a working React and Tailwind page in minutes — and it will look exactly like everyone else's working React and Tailwind page. That sameness is the problem worth fixing before you ship. The Median-Aesthetic Problem: Why Vibe-Coded Pages Look Alike Vibe-
Strix solved 100 of 104 real-world exploits — at $3.37 each
Strix (usestrix/strix) orchestrates recon, exploitation, and post-exploitation agents to deliver working PoCs — here's
Claude Code frontend plugin bans three looks it kept producing
Claude Code's Frontend Design plugin forces aesthetic lock-in before CSS and bans three common AI-generated patterns.
101k stars and can't do long-form: MoneyPrinterTurbo's limit
MoneyPrinterTurbo, ShortGPT, and OpenMontage compared for LLM-scripted, TTS-narrated long-form YouTube video automation.
Sub-agents can hold MCP servers the parent session doesn't
Claude Code sub-agents can hold MCP servers the parent lacks. Covers tools, disallowedTools, and mcpServers frontmatter.
Claude can't grade its own loops — /goal uses Haiku instead
Claude Code's /goal routes stop-condition checks through a separate Haiku model — self-preference bias drove the design.
Claude Code skips bash guards in agentic mode
Claude Code's agentic bash guard is off by default. Covers Telegram control planes and broker API safety for trading.
Meetily runs Whisper and Ollama locally — no cloud, no bot
Meetily is an open-source Tauri desktop app with 27.9k GitHub stars that transcribes and summarizes meetings on-device
Claude Code can drive a quant backtester via Vibe-Trading MCP
Vibe-Trading ships an MCP server so Claude Code or Codex CLI can drive its 88 finance tools and 9 backtest engines.
React detection → Playwright MCP; payments → security subagent
How claude-code-setup maps React to Playwright MCP and payments code to a security-reviewer subagent, explained.
claude -p is all that stands between you and a hanging CI job
Run Claude non-interactively with -p: subprocess wiring, allowed tools, output formats, GitHub Actions integration.
Strix claims 96% on XBEN — that number is vendor-reported
Strix: agentic pentest agent, 96% XBEN (vendor-reported), OWASP Top 10 coverage, SARIF 2.1.0 output, vs Garak and PyRIT.
Two lines swap OpenAI for 160 free NVIDIA-hosted models
NVIDIA NIM free tier: 160+ models via OpenAI-compatible API, 40 req/min, no card. API key at build.nvidia.com.
OpenMinis on-device Alpine Linux agents iPhone GitHub — iSH vs Termux vs mobile Claude wor
OpenMinis is the rare mobile app whose headline feature is architectural, not a chat window: it drops a full Alpine Linux shell onto your iPhone and lets a model drive it. Here is what actually ships inside. What OpenMinis Packs: ARM64 iSH Fork, PRoot, and iPhone API Bridging OpenMinis is
Two BitChat releases in 48 hours moved the offline mesh stack
BitChat v1.6–v1.7 (July 7–8, 2026) adds store-and-forward, courier drops, and live voice DMs to its BLE mesh.
One missing checkpoint can break every approval gate
LangGraph approvals, LangSmith observability, checkpoints, CrewAI, and AutoGen compared for production agents.
Checkpointing is where while-loop agents break
LangGraph checkpointing, human review, retries, and durable execution for loop-style agents.
Handoffs can turn one task into a 15x token bill
LangGraph multi-agent workflow guide to token usage tracing, state pruning, handoffs, and official docs.
The prompt is the smallest part of a coding agent
Claude Code agent architecture, harness control, execution loops, and minimal system prompt design explained.
Disabling thinking can 400 your Opus 5 calls
Claude Opus 5 API setup, Claude Code routing, benchmarks, cost controls, and Opus 4.1 fallback risks.
Agent Reach installs the tools, then gets out of the way
Agent Reach CLI setup, doctor checks, GitHub, OpenCLI, Firecrawl, Jina Reader, and Browserbase compared.
Ego says 2.5x faster; the catch is your Mac
Ego Lite ego-browser skill setup, macOS limits, Space isolation, and logged-in agent workflows.
OpenAI's Codex plugin hands your code review to a second LLM
OpenAI's Codex plugin runs inside Claude Code to delegate code review, adversarial critique, and tasks to Codex locally.
World Monitor hit 67k stars — here's what the MCP endpoint
Open-source OSINT dashboard by @koala73: 60+ Vercel Edge Functions, Ollama/Groq AI layer, MCP at /mcp, 56 map layers,
Codex CLI dropped chat-wire — OpenCodex picks up the routing
OpenCodex translates Codex CLI's Responses API to Anthropic, Gemini, or Ollama — 40+ providers, zero patches to Codex
OpenAI ships Codex into Claude Code — two commands, or four?
OpenAI's codex-plugin-cc adds Codex review to Claude Code. The real install path is four commands, not two.
LongCat-Video-Avatar 1.5 cuts inference to 8 steps — here's
Meituan's MIT-licensed LongCat-Video-Avatar 1.5 replaces Wav2Vec2 with Whisper-Large-v3, adds DMD2 distillation at 8
MengTo/Skills has no tagged releases — pin the commit
78 MIT agent skills across web design, GSAP, Three.js, Codex. No releases in MengTo/Skills — pin a commit.
LingBot-Map runs 10,000 frames on monocular video — no
Ant Group's Robbyant open-sources LingBot-Map, a feed-forward monocular 3D scene reconstruction model running at ~20
36 languages parsed locally — Graphify needs no model call
Graphify indexes your codebase with tree-sitter across 36 languages — local-only, no embeddings, no API calls.
OmniRoute's "unlimited free" claim — 90 providers, not 200
OmniRoute gives Claude Code 90+ free-tier providers via quota fallback — not the 200+ free that viral coverage claims.
Voicebox clones your voice in 3 seconds — but read the fine
jamiepine/voicebox bundles 7 TTS engines, 23-language cloning, system-wide dictation, MCP integration, and a local
One Claude session, 20 front-ends — that's Hermes
Hermes: one Claude session across Telegram, Slack, and 20+ apps. Install, provider config, and trigger setup covered.
The env var that silently routes Fable agents to Haiku
How CLAUDE_CODE_SUBAGENT_MODEL precedence works in Claude Code, and why it silently overrides Fable 5 pinning.
One SKILL.md replaces CLAUDE.md, AGENTS.md, and .cursorrules
Agent Skills open standard: SKILL.md replaces CLAUDE.md, AGENTS.md, and .cursorrules for Claude Code, Codex, and Cursor.
Claude Code subagents concurrent parallel architecture context isolation limits 2026
A single operator running a "department" of AI employees — one researching, one writing, one analyzing — is the pitch behind a viral demo, but Claude Code's real parallelization story is more precise, and more limited, than the headline suggests. How Claude Code's Parallelization Architecture Evolved
For Dhan and Shoonya, the live order code path doesn't exist
Vibe-Trading v0.1.11 adds NSE/BSE backtesting; Dhan and Shoonya are hard-coded paper-only with no live order path.
20+ agent backends, one Artifacts canvas — self-hosted
Apache-2.0, local-first canvas supporting 20+ coding agents as a self-hosted Claude Artifacts alternative.
launchd KeepAlive + tmux detach = respawn loop — the fix
Why launchd KeepAlive + detached tmux respawn-loops Claude Code and the safe RunAtLoad bootstrap pattern to avoid it.
Seedance 2.5 targets 30s 4K — nothing independent confirms it
Seedance 2.5: native 4K, 30s clips, GA early July 2026. Practical access now: Seedance 2.0 on BytePlus ModelArk.
96% of cuBLAS, no `unsafe`: what cuTile Rust proves
cuTile Rust applies Rust ownership to GPU kernels. Covers sm_80+ prereqs, the partition-dispatch pattern, and Grout.
Fugu hides the seams: multiple AIs, billed as a whole
Fugu wraps a swappable LLM pool in an OpenAI-compatible endpoint. Setup, tier comparison, billing, and EU restrictions.
Three packages claim 'SkillsGuard'. One shipped malware.
The SkillsGuard that actually ships: AgentGuard v1.1.28. Four commands to first scan, 24 detection rules, known gaps.
An offline AI you power with your arm — 48 tok/s on a Pi 5 CPU
Squeez Labs' hand-cranked offline voice AI on Pi 5: the edge_voice_agent stack, parts, and commands to recreate it.
GitHub Copilot App can now merge PRs while you're offline
How the GitHub Copilot App works: agent sessions, PR delegation, automations, and AI Credits pricing as of June 2026.
Creative Agent spans the CC suite now. Here's the paid gate.
Firefly AI Assistant (public beta, April 2026) is cross-app. This covers eligibility, invocation steps, and supervision.
Skip RAG entirely — SubQ loads your whole codebase in one pass
SubQ (Subquadratic) is an OpenAI-compatible LLM in private beta with a subquadratic sparse-attention architecture, base URL api.subq.ai/v1, bearer auth, subq-preview model ID, SubQ Code CLI, and a 1M-validated (not 12M) context. Here's how to wire it in.
Firefly's Creative Orchestrator Is Live. After Effects Isn't.
Firefly AI Assistant in Photoshop, Premiere & Illustrator June 18. Eligibility, steps, and what's still waitlisted.
GLM-5.2's FrontierSWE 74.4 is vendor-only. Does it hold up?
Z.ai's GLM-5.2: 1M-token span, Anthropic-compatible endpoint, MIT weights, and a benchmark reality check.
K2.7 Code is 30% lighter — but chain-of-thought is locked on
Kimi K2.7 Code (June 2026): mandatory chain-of-thought, 256K context, HighSpeed at 180 t/s. CLI and API guide for two distinct entry points.
Kog hits 3K t/s on MI300X, no kernel switches — test it now
Kog AI's monokernel collapses AMD MI300X LLM decode into one GPU-resident kernel — 3,000+ t/s at batch 1, no CPU hand-off. How to try it and what replication actually costs.
Fable 5 refusals are 200 OK — your error handler misses them
Fable 5 and Mythos 5 are offline under a U.S. export-control order. Wire up refusal handling and Opus 4.8 fallback now.
Hand off any session to the GUI — Codex CLI 0.138.0
OpenAI Codex CLI 0.138.0 (June 8, 2026): GUI hand-off via /app, v2 PAT auth, faster resume. Full tutorial here.
macOS 27 skips Intel — Siri AI is queued, Liquid Glass is not
macOS 27 Golden Gate drops Intel, Siri AI is gated. Device eligibility, install steps, and what to test in WWDC 2026.
Ideogram 4 was trained on JSON — plain prompts are second-class
Ideogram 4.0: open weights, JSON-first prompting, bounding-box layout, native 2K. API starts at $0.03/image.
A 10-second Grok Imagine 1.5 clip at 720p runs $1.41
Image-to-video only; audio prompting added; $0.14/sec at 720p. Frame prep, REST and SDK steps, rate limits, known constraints.
Gemma 4 12B skips the audio encoder. Is 16 GB enough?
Encoder-free 12B open-weight model (June 2026): Ollama, Transformers, llama.cpp GGUF, and audio caveats.
Meta Business AI went global — gated rollout, paid plans TBD
Meta Business AI: activation flow, four controls, WhatsApp messaging fees, and the fine print on a gated market rollout.
MiniMax M3 benchmarks at $0.30/M: verified vs. vendor-only
MiniMax M3 at $0.30/M: what the 1M-sequence benchmarks mean, credential selection, and a quickstart.
NVIDIA's 550B finally lands: free to use, expensive to host
Nemotron 3 Ultra, 550B MoE (June 4 2026): hardware minimums, hosted API quickstart, NIM steps, benchmark check.
Qwen3 in the browser, zero keys — WebLLM 0.2.83 hands-on
WebLLM 0.2.83: run Qwen3 in Chrome via WebGPU, no server. Setup steps, streaming code, VRAM requirements, and gotchas.
NeMo out, GGUF in: how parakeet.cpp ports NVIDIA ASR to C++
parakeet.cpp v0.1.0: NVIDIA Parakeet in GGUF — no NeMo needed. CMake steps, quant tradeoffs, and whisper.cpp status.
Is Omni's conversational video editor as good as the demos?
Gemini Omni in Google Flow: credit costs, regional limits, and iterative editing — no callable API yet.
Windsurf is Devin Desktop now. Cascade has 27 days left.
Windsurf is now Devin Desktop: Agent Command Center, Spaces, and Devin Local replacing Cascade by July 1.
Nemotron 3 Ultra went live June 4. Here's the call that works.
NVIDIA Nemotron 3 Ultra GA June 4: how to call via NIM/OpenRouter, hardware floor, and the base-checkpoint caveat.
Composer 2.5 hits near-frontier at 60× lower spend
Composer 2.5: third on the Artificial Analysis Coding Index at $0.07/task vs $4.10 for its nearest rival. Billing choice, effective prompting, and what the independent scores actually show.
Opus 4.8 kills budget_tokens — here's what else moved
Opus 4.8: fast mode, mid-session system prompts, 1K cache floor. Old budget_tokens syntax returns 400.
llama-bench skipped FA on capable GPUs — b9437 corrects it
llama.cpp b9437 (May 30): -fa goes auto, -ngl to -1 in llama-bench. Your pre-b9437 comparisons need a flag audit.
Qwen3.6-35B NVFP4 runs on one H100 — A100 owners are out
FP4-quantized Qwen3.6-35B fits in ~23 GB on Hopper. vLLM serve commands, env vars, DGX Spark config, and gotchas.
Step 3.7 Flash is a drop-in — except for one endpoint detail
StepFun Step 3.7 Flash: 198B MoE with native vision, Advisor Mode, and an OpenAI-compatible API you can call today. Includes endpoint gotchas and reasoning_effort examples.
You don't pick the RL algorithm — SIA's Feedback loop does
SIA co-evolves scaffold and LoRA weights in one loop. Install, run LawBench, and add custom evals — Hexo Labs, May 2026.
'Gemini Omni 3.5' doesn't exist. Here's the real split.
SDK setup, video generation calls, and conversational editing for Gemini Omni — Google's new world model from I/O 2026.
What openai-codex Beta Gets Wrong on First Install
Official openai-codex first beta: how to pin v0.1.0b1, start a thread, and avoid the beta quirks. Released May 28 2026.
What langchain-fireworks 1.4.x Changed for Your Code
What the 1.4.x patch sequence changed — and a runnable ChatFireworks setup from scratch.
Opus 4.8 Thinking Blocks Were Silently Corrupting on Retry
Thinking blocks on Opus 4.8 were corrupting on retry. v2.1.156 is the hotfix — update, verify, and see what else landed.
Your Claude Code Skills Now Hot-Reload Without Restart
Claude Code v2.1.157 adds .claude/skills/ live-loading, worktree unlocking, and OTEL telemetry. Annotated guide.
openai-codex b2 Has a Renamed Config Class Worth Knowing
v0.1.0b2 ships named Sandbox presets and a renamed config class. A runnable walkthrough from pip install to first thread.
Claude Code Now Gates Execution on Bedrock and Vertex
v2.1.158 enables classifier-gated execution on managed inference platforms. Here's the env var, what it does, and what to verify before upgrading.
How to Add SuperGrok to Kilo Code in Any Environment
Set up grok-build-0.1 in Kilo Code using your SuperGrok or X Premium+ subscription — VS Code, JetBrains, CLI, and SSH.
Headless Auth and Streaming With openai-codex in CI/CD
Practical patterns for async, streaming, and headless auth using openai-codex 0.1.0b2 in CI/CD pipelines.
SuperGrok 티어별로 Kilo Code 설정이 달라진다
Set up grok-build-0.1 in Kilo Code using your SuperGrok or X Premium+ subscription — VS Code, JetBrains, CLI, and SSH.
AI developer tools and ecosystem news for developers and technical founders