Developer Trial Feedback - `local-ai-use` onboarding, cleanup, and naming

[danielholanda]

Dogfooding feedback: local-ai-use onboarding, cleanup, and naming

Onboarding / walkthrough fixes

1. Node.js isn't listed as a prerequisite. walkthroughs/local-ai-use.md tells the user to run npx skills add ... but the walkthroughs README only promises claude code is installed. A beginner won't infer the Node dependency. One line alongside the existing "assumes claude code is setup" is enough.

2. The skills CLI link reads as "go install this repo." The walkthrough links vercel-labs/skills next to an npx command with no explanation, and the reviewer lost ~10 minutes trying to figure out how to install that repo before realizing npx fetches it. NVIDIA's README hits the same shape but spells out the resolution right after; ours says something similar but truncated enough that it reads as a comment on npx generally rather than a pointer to the CLI being invoked. Clarify in both the main README and the walkthrough.

3. Offer the manual install path in the walkthrough. Add a one-liner (git clone + copy the skill folder) or link the main README's manual-install section, so users who prefer that route aren't forced through npx.

4. Warn that the first run is slow, and how to survive it. The first model pull may take long enough to look hung. Note this under "Default modality models" or troubleshooting, including raising the timeout on whatever command the agent issues, rather than leaving the agent to volunteer the warning.

5. Typo: "Participatns" in walkthroughs/README.md.

Missing cleanup / uninstall story

reference.md covers only a fraction of teardown. The reviewer had to write a private Claude skill just to reset the machine between iterations. A supported cleanup path should cover: uninstalling the Lemonade server and backends, reverting user-group changes, disabling server/file autostart, and deleting temp files and cached models.

Remote / SSH execution is under-discoverable

The reviewer ran the agent on one machine and SSH'd into the test system. This works once you tell the agent to execute only on the remote, but the agent hit initial hiccups because it never read the remote-host guidance that already exists in reference.md — it needs remote_ip:port instead of localhost for the Lemonade endpoint, and AGENTS.md on the host. Adding a few words about "remote/host system connection and execution" to the Reference section of SKILL.md would raise the odds the agent actually opens reference.md. This also covers the common case of a headless home server.

Naming and scope boundary

local-ai-use reads like it sets up everything local, including text generation — the most popular LLM use case, and the one thing the skill deliberately excludes (a separate skill is planned, per reference.md). Users judge on the name alone, so this is a discoverability problem. Suggested renames: local-multimodal-use, local-media-ai, local-vision-speech, or local-lemonade-use (the skill is Lemonade-specific by design).

Alternatively, keep local-ai-use as a dispatcher that decides what to set up and delegates to modality- or vendor-specific skills such as local-lemonade-use. That's the more modular long-term shape if more local backends land later.

Scope itself was rated appropriate: staying Lemonade-only avoids the complexity and wrong-choice risk of covering ComfyUI, Stability Matrix, and friends.

local-ai-app-integration: two near-miss triggers

The skill fired on "local llm + img gen setup" and on "saving costs + local text llm gen only", then correctly recused itself using the "If the user instead wants a system-wide Lemonade Server" note. Self-correction worked both times, so this is not a bug — but tightening the description to exclude system-wide setup and text-only requests would save the recovery turn.

Measurements supporting the skill's value

Same task (local image generation), with and without the skill:

Setup	Output tokens	Total tokens	Notes
Opus + skill	26k	2.2M	Fastest, essentially plug-and-play
Opus, no skill	~63k	~7.5M	Correct, but asked which backend to install
Sonnet 4.6 + skill	10k	1.8M	Slower than Opus + skill, still clean
Sonnet 4.6, no skill	~21k	~4.3M	Kept installing ROCm system-wide, refused venv until forced, pulled wrong ROCm/torch versions, defaulted to ComfyUI
Sonnet 5, no skill	~16k	~3.4M	Lemonade only when told explicitly; tried Vulkan and ROCm

Totals include cache reads. The skill saved tokens and time in every pairing, and the gap widened on smaller models — which is where the value lands for users running local models on consumer GPUs.

Feedback from Hassan Nasir