agentsclimarketplace

Onboarding

Skill latent-variable/agentscope/skills/onboarding

One user-scope brain every agent CLI reads: Claude Code, Codex, Antigravity/Gemini, Pi. Clone to ~/.agents, let your agent onboard you.

Install
npx -y skills add latent-variable/agentscope --skill onboarding

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

First-run setup for agentscope. Interview the user, fill the AGENTS.md template + memory with their identity and preferences, detect which agent CLIs are installed, wire them all to ~/.agents, and tune optional features (ELI5 explanations, the Speak-to-me TTS block, the review-cycle workflow). Run this the first time someone clones agentscope, or when they say "onboard me" / "set up agentscope".

SKILL.md

5.4 KB, as published. Nobody here has run it

Onboard a new user to agentscope

Goal: turn the shipped template into this person's user-scope brain, and wire every agent CLI on their machine to it. After this, every tool (Claude, Codex, Gemini/Antigravity, Pi) reads the same identity, preferences, skills, and memory. Be warm and quick, this is someone's first impression of the system.

0. Detect state first (don't re-onboard a configured user)

ls -d ~/.agents 2>/dev/null || echo "NOT INSTALLED at ~/.agents"
grep -l '{{' ~/.agents/AGENTS.md 2>/dev/null && echo "TEMPLATE still has placeholders → onboard"
  • If ~/.agents doesn't exist: the repo was cloned elsewhere. Tell them to git clone ... ~/.agents (or move it), then re-run.
  • If AGENTS.md has no {{...}} markers: already onboarded. Confirm and offer to re-wire (bin/sync.sh) or edit a specific preference instead.

Detect installed CLIs (only these get wired):

for d in .claude .codex .gemini .pi; do [ -d "$HOME/$d" ] && echo "found: ~/$d"; done

1. Interview (keep it short, 6 quick prompts, conversational)

Ask these, adapting to what they volunteer. Don't interrogate; one message with the list is fine.

  1. Who are you? Name + a sentence or two: role, background, what you build. (→ bio + user_profile.md)
  2. Where does your work live? Top project directories (e.g. ~/Documents/..., ~/Code/...) + GitHub handle. (→ "Where my work lives")
  3. Explanation style, do you want ELI5? Plain-language analogies layered on top of full technical substance, or straight technical only? (→ keeps/strips the ELI5 line)
  4. Spoken summaries? Do you use (or want) text-to-speech reading a short "here's what I did" at the end of each reply? (→ keeps/strips the 🔊 Speak to me block). If yes and they're on macOS, mention Yap (github.com/latent-variable/Yap), local-first on-device TTS+STT built for this, and offer to help install it (§4).
  5. Workflow, structured review cycle? Do you want agents to branch / open PRs / run automated review before merging code, or keep it lightweight? (→ keeps/strips the review-cycle bullet)
  6. Attribution name for commits (defaults to "Claude" / the agent's own name + model id). Most people keep the default.

2. Fill the template

Edit ~/.agents/AGENTS.md, replace every {{...}} block:

  • {{YOUR_NAME}} in the title and "Who I am" → their name + bio paragraph.
  • {{BIO ...}} → the one-paragraph bio.
  • {{PROJECT_DIRS ...}} → their actual dirs + GitHub: <handle>.
  • {{ELI5_TOGGLE ...}} → if yes, replace with the literal sentence: Use ELI5 framing when explaining: concrete analogies and everyday objects, full technical substance kept. If no, delete the toggle text (leave the surrounding sentence clean).
  • {{REVIEW_TOGGLE ...}} → if they want it, delete just the brace note (keep the bullet). If not, delete the whole review-cycle bullet.

Speak-to-me block (between <!-- BEGIN speak-to-me --> and <!-- END speak-to-me -->):

  • TTS yes → keep the block; delete the two HTML comment markers so it reads clean. Keep the Yap pointer only on macOS.
  • TTS no → delete the entire block including both markers.

Verify nothing is left: grep -n '{{' ~/.agents/AGENTS.md must return nothing.

3. Seed memory

  • Write ~/.agents/memory/user_profile.md from the bio (use the remember skill's format: frontmatter + body). Fill the template that ships there.
  • Add/confirm the one-line pointer in ~/.agents/memory/MEMORY.md.
  • If they gave project dirs worth pinning, optionally seed memory/reference_project_dirs.md too.

4. (macOS + TTS only) Offer Yap

Only if they said yes to spoken summaries and are on macOS:

# Public, MIT-licensed. Easiest install is Homebrew; or DMG from releases; or build from source.
echo "Yap: local-first on-device TTS + STT for macOS, github.com/latent-variable/Yap"
echo "  brew install --cask latent-variable/tap/yap   # or grab a DMG from /releases"

Offer to walk them through install. Don't block onboarding on it, note it and move on if they defer.

5. Wire every installed CLI

~/.agents/bin/sync.sh
~/.agents/bin/sync.sh --check

sync.sh only touches tools it finds; it backs up anything real it replaces into ~/.agents/backups/.

6. Back it up (recommended)

Encourage them to make ~/.agents their own private repo so canon is versioned + backed up, and so the self-correcting loop has somewhere to push:

cd ~/.agents && git add -A && git commit -m "onboarding: my user scope" && git push   # after they add a remote

7. Verify + hand off

~/.agents/bin/verify.sh

Then a short confirmation: who you recorded them as, which CLIs you wired, which optional features are on (ELI5, Speak-to-me, review-cycle), and the two things they control going forward, edit AGENTS.md for global behavior, drop facts via the remember skill. Point them at the other skills by name so they know what's available.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.