Every week the OpenClaw community surfaces something worth stealing. Here are this week's three best.


1. The 50+ Tip Master Cheatsheet — All of Steinberger's Best Advice, Finally in One Place

Peter Steinberger has been scattering OpenClaw tips across replies, comments, and threads on X for months. @aiedge_ noticed, and did the work nobody else had bothered to: hunting down all 50+ of them and consolidating them into a single master cheatsheet at moltfounders.com/openclaw-mega-cheatsheet. The collection runs the full gamut — from foundational mindset ("treat it like a competent, tireless personal assistant you text, not a coding tool") to advanced moves like using the workspace folder as a shared context layer, switching models mid-conversation, defining custom skills in YAML, security-hardening your deployment, and chaining multi-step tasks with approval checkpoints. If you've ever felt like you're only using 30% of what OpenClaw can do, this is the fastest way to close that gap. It's the unofficial manual Steinberger never wrote but clearly meant to — and the community is treating it like a bible.

https://x.com/aiedge_/status/2039825229472747704


2. PinchBench — The Scientific Way to Pick Your OpenClaw Model

With Anthropic's Claude subscription restriction still reshaping the landscape, the question on everyone's mind is: which model should I actually run my Claw on? Now there's a benchmark built specifically to answer it. The team at kilocode (makers of Kilo Code) built PinchBench — 32 models put through 192 standardized runs on real-world OpenClaw tasks, scored across three axes simultaneously: success rate, speed, and cost. The leaderboard at pinchbench.com is interactive, so you can filter by whatever matters most to your setup — maximize task completion, minimize spend, or find the sweet spot in between. Charly Wargnier (@DataChaz) highlighted it as the missing tool the community desperately needed, and the reception has been enormous — Jensen Huang even cited it publicly. No more vibes-based decisions between GLM 5.1, Qwen3.5-35B, and Kimi K2.5: run the data, pick the model, don't look back.

https://x.com/DataChaz/status/2029244775870906414


3. v2026.4.23-beta — OpenClaw Can Now Edit Your Images Too

The latest OpenClaw beta quietly ships something that solo creators and content-automation builders have been asking for since the video and music generation landed in v2026.4.5: native reference-image editing. The v2026.4.23-beta wires in support for OpenAI's gpt-image-2 (via Codex OAuth — no separate API key required for most flows), OpenRouter-compatible image generation models, and xAI's grok-imagine for visual generation and editing. In practice, this means your agent can now take an existing image, mask a section, and regenerate it with natural-language instructions — all embedded in an automated workflow alongside your existing text, code, and research tasks. The beta also ships memory/CLI embedding fixes (resolving the long-standing local embedding inconsistency at #70836), improved plugin install linking, and ongoing provider security hardening. For anyone already running content pipelines with Dreaming memory or scheduled skill chains, this turns OpenClaw from a capable text-and-code agent into something closer to a full production suite.

https://github.com/openclaw/openclaw/releases


Spotted something great on X about OpenClaw? Share it — the community is the best part.