Shipping ClawHub skills that agents (and humans) actually install

The bar: installable beats impressive

A prompt pasted into a chat window is not a product. An OpenClaw skill that agents can install from ClawHub — with a version, a failure mode, and a non-agent escape hatch — is closer to shipping.

I use a blunt gate before I call something done: Can a stranger install it from ClawHub in one flow? Can a human without an agent runtime still run it via Docker / CLI? Are downloads visible so I cannot lie to myself with GitHub stars?

AI Test Case Generator and Trading Assistant Core both cleared that gate. The rest of this note is the concrete packaging and product decisions behind the download curve.

Download proof (refreshed on the homepage)

ClawHub downloads are the metric I trust more than repository stars. Stars are optional social proof. Installs mean an agent runtime or a curious human pulled the package.

Trading Assistant Core currently leads on installs; Test Case Generator follows close behind. Together they clear 1.6k+ downloads on the profile — numbers this site refreshes live via /api/metrics.

Case A — AI Test Case Generator

QA teams drown in PDFs, screenshots, and half-written notes. Raw LLM dumps create more review debt than coverage. The product bet was multimodal intake plus a Test / Dev / Product review loop, then packaging that loop for agents and humans.

Distribution is intentionally redundant: ClawHub for agents, Docker Hub for laptop / CI runs, npm for JS-native workflows. One pipeline, three install surfaces — not three demos.

• Inputs: PDF / Word / images / video — not text-only prompts
• Review personas reduce noisy drafts before export
• Exports teams already use: Excel / Markdown / XMind
• Tradeoff: slower first answers in exchange for reviewable artifacts

Case B — Trading Assistant Core

Trading workflows were stuck in notebooks and chat prompts. Agents could not install a durable skill with clear risk guardrails. Shipping a “signal fireworks” skill would have been irresponsible.

The skill packages support/resistance, multi-indicator signals, position sizing, and risk alerts — with risk-first defaults. CLI and Docker remain for people who are not inside an agent runtime.

• Risk sizing and alerts outrank flashy entry calls
• ClawHub install for agents; CLI / Docker for humans
• Same logic across surfaces — avoid “agent-only truth”
• Downloads ~900+ — stronger signal than a polished landing page alone

Packaging matrix I refuse to shrink

Every serious skill gets at least two paths: agent install and human install. Shrinking to ClawHub-only feels modern until a teammate without the runtime needs to reproduce a bug.

                      Surface          Test Case Generator     Trading Assistant
---------------  ----------------------  --------------------
ClawHub skill    required                required
Docker image     required                required
CLI / pip/npm    npm path                pip / CLI path
Docs + failures  required                required
Live metric      ClawHub downloads       ClawHub downloads

                    

Review loops beat hotter models

For testcase generation, the quality jump did not come from chasing the newest model name. It came from forcing Test / Dev / Product perspectives before export. Drafts got quieter; QA leads accepted more of them.

For trading, the quality jump was saying no to signal-only packaging. An installable skill that only shouts entries is a liability. Defaults that emphasize sizing and risk are slower to market — and safer to put on ClawHub.

Release gate (copy this)

Before each publish I answer four questions out loud. Any “no” blocks the release — even if the demo video looks great.

• ClawHub: can a stranger install in one documented flow?
• Human path: Docker or CLI works without the agent runtime?
• Failure mode: is the first error message actionable?
• Metric: will downloads show up somewhere I check weekly?

What I optimize next

More skills are useless if the install bar drops. The next investments are clearer failure docs, tighter versioning notes on ClawHub, and keeping Docker / CLI paths in lockstep with skill releases so agents and humans never diverge.

If you are packaging your own OpenClaw skill: publish the install path first, then the clever prompt. The download curve will tell you whether you shipped a product or a performance.