Personal agent plugin: global coding instructions, a secrets guard, code review agents, review workflow skills and a UI design canvas, shared across devices and accounts. Built to sit next to superpowers, not to replace it.
global-rules.md: the instructions that would otherwise live in ~/.claude/CLAUDE.md. They cover reporting style, git branches and commits, migrations, dependencies, comments and code review. Loaded at the start of every session, and again after /clear and compaction.
hooks/protect-secrets.py runs before every Read, Grep and Bash call, subagents included, and denies:
- Reading secrets:
.envand.env.*(.env.localand.env.example,.sample,.dist,.templatestay readable), private keys (id_*,*.pem,*.key,*.p12,*.pfx), Composerauth.json, and credential stores in the home directory:~/.ssh(exceptconfig,known_hosts,*.pub),~/.gnupg,~/.aws,~/.azure,~/.config/gcloud,~/.kube, Docker, Git and npm credentials,~/.netrc,~/.pypirc,ghandglabconfigs, and the Claude and Codex auth files. - Remote shells and transfers:
ssh,scp,sftp,rsync,sshpass,mosh,autossh, also behindsudo,timeoutorenvand insidebash -c,evalor$(...). - Uploads:
curlandwgetsending data (-d,--data*,--json,-F,-T,--post-*,--body-*, or a method other than GET or HEAD). Downloads stay allowed.
Claude gets the reason and asks you instead. It is a guardrail, not a sandbox: indirect reads such as printenv or scripts that load .env themselves are not caught.
| Agent | Runs | Does | Edits code |
|---|---|---|---|
larapulse:code-simplifier |
Before code review or a merge request, or when asked | Simplifies the branch's changes without changing behavior; runs the tests before and after | Yes |
larapulse:silent-failure-hunter |
During code review, or when asked | Reports swallowed errors, unjustified fallbacks and unhelpful error messages | No |
Claude dispatches them on its own at those points, or ask by name ("run code-simplifier").
| Skill | Used when | Does |
|---|---|---|
larapulse:mr |
Claude writes a merge request description | Keeps it to one or two sentences plus only what the diff doesn't make obvious: env variable names, breaking changes, related merge requests in other repositories. Follows the project's template when there is one. |
Skills load on their own when the task fits; each costs a one-line description in context.
| Command | Does |
|---|---|
/larapulse:handoff [purpose] |
Writes a brief of the current work for another agent, by default an independent reviewer. After review feedback, writes a follow-up covering only the new changes. |
/larapulse:review-handoff <brief or path> |
Reviews the changes a brief describes in every repository it lists and gives a merge verdict. Fixes minor non-logic issues itself, lists logic findings for the developer. |
/larapulse:wait-what |
Re-pitches the last message with the missing context, in full sentences of plain English, when a terse report didn't land. |
/larapulse:ui-review [files] |
Checks UI code against about 100 web interface rules (accessibility, focus, forms, motion, layout, i18n, copy) and lists findings as file:line. Without arguments, reviews the branch's changed UI files. |
/larapulse:design-canvas [what to design] |
Designs or redesigns UI with you on a live browser canvas: mockups you choose between, comment on, edit, draw on and tweak, round by round. Ends with a spec, a standalone mockup and an implementation handoff in docs/design/. |
A review round: run /larapulse:handoff where the code was written, paste the brief into the reviewer (another agent, such as Codex, on any machine) with /larapulse:review-handoff, and forward its report back. After fixing, /larapulse:handoff again writes the follow-up for the next round.
The short forms /handoff, /review-handoff, /wait-what, /ui-review and /design-canvas work too while no other command uses those names. Commands only run when typed, so their descriptions cost no context.
- Context: the global instructions add about 1,200 tokens to every session, again after
/clearand compaction. - Duplicates: instructions still present in
~/.claude/CLAUDE.md(or~/.codex/AGENTS.md) load twice. Remove them there. - Branches: on
master/main, Claude creates a feature branch before changing code, unless the project's CLAUDE.md or AGENTS.md or an explicit instruction says otherwise. - Agents: each run is a subagent with its own context and token cost.
code-simplifieredits files and runs the project's tests;silent-failure-hunteronly reads. - Handoff files: briefs and reports longer than about 40 lines go to
.handoff/in the repository root, which is added to.git/info/excludeso it is never committed. Shorter ones are printed in chat. - Review fixes:
/larapulse:review-handofffixes minor non-logic issues (typos, stale comments) directly in the reviewer's checkout. - Design canvas: needs Node.
/larapulse:design-canvasruns a local server on localhost, protected by a random key, and waits up to 30 minutes per round for you to send. Scratch screens go to.design-canvas/, which it adds to.git/info/exclude; the deliverable goes todocs/design/. - Secrets guard: about 30 ms per Read, Grep and Bash call, and no tokens unless it blocks. It needs
python3: without it, every Read, Grep and Bash call is blocked until it is installed.cp .env.example .envis blocked as well; run it yourself. - Updates: with auto-update on, every push to
masterreaches every device. There are no version bumps.
Add to ~/.claude/settings.json:
{
"extraKnownMarketplaces": {
"larapulse-skills": {
"source": { "source": "github", "repo": "larapulse/skills" },
"autoUpdate": true
}
},
"enabledPlugins": {
"larapulse@larapulse-skills": true
}
}Or from a session: /plugin marketplace add larapulse/skills, then /plugin install larapulse@larapulse-skills, then /plugin > Marketplaces > larapulse-skills > Enable auto-update. Third-party marketplaces don't auto-update by default.
- Update by hand:
claude plugin update larapulse@larapulse-skills - Remove:
claude plugin uninstall larapulse@larapulse-skills
Cloud sessions ignore plugin settings from devices and repositories. An organization Owner adds the same JSON in claude.ai > Organization settings > Claude Code > Managed settings. Not yet tested.
The skills CLI installs this repo's skills and commands into Claude Code, Codex, OpenCode, Cursor and many other agents:
npx skills add larapulse/skills -gIt asks which skills and agents to use; --skill <name> and -a <agent> (for example -a codex) pick them up front, and --list shows what's available. Update with npx skills update.
It installs skills only. The global instructions, the secrets guard and the agents need the plugin, so on Claude Code use the plugin instead; installing both would load every skill twice.
Codex reads Claude Code plugin manifests, so the global instructions load through the same hook. Not yet tested.
codex plugin marketplace add larapulse/skillscodex plugin add larapulse@larapulse-skillsThen trust the hooks in /hooks. Codex asks again after every update. Whether Codex runs the secrets guard is not verified. Codex updates plugins at startup.
The commands come with the plugin as Codex skills. Agents are not available: Codex agents use a TOML format, and plugins cannot ship them.
OpenCode doesn't read Claude Code plugins. Point its instructions at the file in ~/.config/opencode/opencode.jsonc; it always reads the latest version from master. Not yet tested.
{
"instructions": ["https://raw.githubusercontent.com/larapulse/skills/master/global-rules.md"]
}For the commands, use the skills CLI:
npx skills add larapulse/skills -g -a opencodeAgents are not available: OpenCode uses a different agent format.
Gemini CLI, Cursor and GitHub Copilot CLI are not set up. Each has its own manifest and hook format.
Plugins and skills that pair well with this one, what each costs, and what to avoid installing next to it: RECOMMENDED.md.
Conventions for changing this repo are in CLAUDE.md. Adapted third-party files and their licenses: THIRD_PARTY_NOTICES.md.