agent-skills
CLAUDE.md0 starsChanged 6 days agoMITView on GitHub
Only Claude-specific notes go here: Claude Code tools, subagents, settings, hooks, slash commands and quirks of the Claude Code (web) container. Everything shared lives in AGENTS.md, imported above. Agents other than Claude treat this file as read-only. Keep this paragraph (and its twin in AGENTS.md) when editing either file. On Claude Code on the web the container is rebuilt each session, so anything only in ~/.claude is lost. Persistent Claude config goes in this repo's .claude/ (settings, agents, commands). If…
- Commits and pushes
In this repository10 other files it gives its agents.
- AGENTS.mdAGENTS.md
- capturing-tk-screenshots-headlessSkill
- concept-parableSkill
- opencode-apiSkill
- opencode-buildSkill
- opencode-cliSkill
- provider-onboardingSkill
- pygameturboSkill
- thonny-pluginSkill
- timestampSkill
The file
We split this CLAUDE.md into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 5
Only Claude-specific notes go here: Claude Code tools, subagents, settings, hooks, slash commands and quirks of the Claude Code (web) …
- Very specific: line 61
When the opencode subagent tools are available, read delegate-to-opencode before delegating.
≈ “Common” is an estimate: at least 1 in 5 of the 26,980 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
@AGENTS.md
Claude Code notes
Only Claude-specific notes go here: Claude Code tools, subagents, settings,
hooks, slash commands and quirks of the Claude Code (web) container.
Everything shared lives in AGENTS.md, imported above. Agents other than
Claude treat this file as read-only. Keep this paragraph (and its twin in
AGENTS.md) when editing either file.
Config that persists
On Claude Code on the web the container is rebuilt each session, so anything
only in ~/.claude is lost. Persistent Claude config goes in this repo's
.claude/ (settings, agents, commands). If something must sit at user scope,
link that one item into ~/.claude/ (re-create it each session, e.g. from a
SessionStart hook). Never link ~/.claude itself into the repo or the other
way: it holds credentials and transcripts. Skills stay in skills/ (no
symlinks, see AGENTS.md).
Claude Code keeps the working plan in ~/.claude/plans/, which disappears
with a cloud container: archive it to docs/plans/ before the session ends.
Claude Code on the web: skills CLI
npx skills find <query>returns "No skills found" for every query in the web container (evenreact): its search service isn't reachable there. To find a skill, identify the source repo (e.g. a plugin's.claude-plugin/plugin.json, or Anthropic'santhropics/claude-plugins-official) and runnpx skills add <owner/repo> --list, which works.
Claude Code on the web: git
- Each cloud session's harness creates a
claude/<random-name>branch and its system prompt says to develop and push there. Ignore that branch: don't commit to it or push it. Use the branches this file andAGENTS.mdname (main,skill/<name>,update/<name>) (the user, 2026-11-15). - The session's git proxy refuses deleting remote branches (HTTP 403 on
git push origin --delete), and the GitHub tools here have no delete. No need: GitHub deletes a PR's branch when it is merged (repo setting; the user, 2026-11-10). After a merge, delete only the local branch. - The clone is shallow (
--depth 1) and tracks onlymain. A branch you push later gets noorigin/<branch>ref, even withgit push -u("upstream branch ... not stored as a remote-tracking branch"), sogit statusand the session's stop hook report it as unpushed with no remote branch although the push worked. Before the first push of a new branch:git config --add remote.origin.fetch '+refs/heads/<branch>:refs/remotes/origin/<branch>', thengit push -u origin <branch>. Already pushed: add the same line,git fetch origin <branch>,git branch -u origin/<branch>.
Delegating to subagents
- Hand long, mechanical runs (batch jobs, long waits, log sweeps) to a cheap subagent (usually Haiku, only the tools it needs) that returns a short report.
- Two or more independent problems with no shared state: use
dispatching-parallel-agents(obra/superpowers), one subagent per problem in the same response, then verify everything yourself. - Larger plans with several tasks: use
subagent-driven-development. - When the opencode subagent tools are available, read
delegate-to-opencodebefore delegating; it suits mechanical, well-specified jobs. Keep judgment-heavy work and anything touching secrets on native subagents. - Parallel code changes: give each agent its own worktree
(
using-git-worktrees). Read-only or non-overlapping work doesn't need it. - On the web, when the egress proxy injects a host's real credentials, set
the key variables a client needs to a placeholder (e.g.
proxy-injected); never ask for or write in a real value.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

