~/.agents

devs only want agents who have great SKILLs

TL;DR: agent skills, commands, and configs.

This repo contains agent skills, subagents, slash commands, hooks and harness-specific setup specs for Claude Code, Codex, Cursor, Copilot, etc. ~/.agents/skills/ is a cross-harness convention, so most harnesses can read the skills as soon as they're cloned. Skills follow the open Agent Skills standard.

Install

Clone the repo to ~/.agents:

git clone https://github.com/thefilesareinthecomputer/dotagents ~/.agents

Most harnesses pick the skills up with no further step. Claude Code reads only ~/.claude, so link the skills, subagents and commands into it:

bash ~/.agents/sync-skills.sh --dry-run   # preview
bash ~/.agents/sync-skills.sh

The script adds one symlink in ~/.claude/skills, agents and commands for each entry in this repo. It leaves any entry already in ~/.claude alone and removes links whose source was deleted. It is safe to re-run; run it after any pull that adds or removes an entry. Skills hot-reload; a new subagent or command registers after a session restart.

Harnesses

Harness	Setup	Note	Spec
Claude Code	sync-skills.sh	Reads skills, subagents and commands; the fullest station spec	SPEC-CLAUDE-CODE
Antigravity CLI	one skills.json pointer	Reads .agents/ per project, not ~/.agents/ globally	SPEC-ANTIGRAVITY
Codex CLI	none	Its privacy switches are a five-key config.toml fragment	SPEC-CODEX
GitHub Copilot	none	Strictest skill parser of the set	SPEC-COPILOT
Cursor	none	Subagents load per repo: sync-skills.sh --cursor <repo>	SPEC-CURSOR
deepagents	in code	Skill paths are passed to create_deep_agent	SPEC-DEEPAGENTS
Gemini CLI	none	Retired for consumer tiers; see Antigravity	SPEC-GEMINI
goose	none	If it misses ~/.agents, use .goose/skills/ in the project	SPEC-GOOSE
Kimi Code CLI	none		SPEC-KIMI
opencode	none	opencode debug skill shows what it found	SPEC-OPENCODE
Pi	none	Finds only <name>/SKILL.md folders	SPEC-PI

Every harness in the table can load skills/. The subagents in agents/ and the slash commands in commands/ are Claude Code formats. Claude Code reads both through the sync, Cursor reads the subagents in a repo after sync-skills.sh --cursor <repo>, and the other harnesses ignore both folders. To make a subagent or command work everywhere, rewrite it as a skill. Two optional skill fields keep the original behavior in Claude Code and Copilot: disable-model-invocation: true runs the skill only when you type its slash command, and context: fork runs it in its own subagent. Other harnesses ignore those fields and load the skill as usual.

specs/ has one station spec per harness. Each lists the config, plugins, CLI tools, hooks and permission rules that harness needs, including its privacy and remote-control switches. seed.py plan --harness all in the agent-cc-configs-sync skill checks a machine against them. specs/HARNESS-PROFILES.md compares the harnesses with dated evidence.

Because every harness reads the same tree, sessions in different harnesses running side by side share one skill set and the same working rules in AGENTS.md. agent-mail lets those sessions message each other through files, with no daemon and no network.

Models. An Ollama :cloud model is served remotely and has the same exposure as any hosted API, whatever the local-looking command suggests. Ollama stays on loopback: OLLAMA_HOST is never set to 0.0.0.0 or any routable address, because that publishes an unauthenticated inference endpoint to the network.

What's inside

Each entry's own file is authoritative: a skill's frontmatter description is its trigger contract and its body is the workflow.

Skills

Session lifecycle

Skill	What it does
hi	Session-start orientation: reads the living docs, memory and git state, then names the next action. Read-only.
focus	/focus for a session that drifted: the supervisor re-derives the user's requests from the record.
cmon	Restates the last verbose reply in as few words as possible, then holds that register.
notes	End-of-session docs sweep that also corrects any repo text the session made false.
reflect	End-of-session memory reconciliation, applied only after the user approves it.
wrap-up	Closes a session: file review, docs and handoff, commit and push, then memory.
repo-device-sync	Multi-device git ritual: fetch and diff before work, re-fetch before every push.

Orchestration and delegation

Skill	What it does
meta-loop	Plan, fan out to parallel workers, verify each result, synthesize, with an advisor consulted off the hot path.
deep-research	Multi-angle web research with parallel researchers, cross-validated and cited.
task	/task <harness> hands one bounded task to Codex, Cursor or Antigravity in an isolated worktree; /task tryout benchmarks a candidate harness.
agent-review	One outside model reviews a user-approved file set in a single headless call with no tools.
agent-conference	Panel review for high-stakes questions: several harnesses answer, rank each other blind, and a chair synthesizes.
local-model-task	Hands one bounded text or search task to a local Ollama model on loopback.
support-local	A Haiku worker drives the local models over a named corpus and returns one JSON answer whose quotes the script checks.

Knowledge, vaults and planning

Skill	What it does
markdown-kg	Offline section-level knowledge graph over any markdown corpus.
code-kg	Offline symbol-level knowledge graph over a codebase: imports, liveness, subsystems, coupling hotspots.
obsidian	Obsidian markdown standard plus a per-vault authoring workflow.
project-knowledge-vault	Seeds, finds and lints a flat project vault and its deliverables folder: the contract, a scaffolder, a linter, and the BRD, solution design, ADR and knowledge-base entry templates.
gather-learnings	Sweeps a vault's dated notes and the running session into its learnings file, and keeps its question register current.
compact-notes	Rewrites one meeting-note section into a fixed summary format, with a backup and a verify.
get-ready	Claims audit before a meeting or a piece of work, written into the vault's run sheet.
pr-review	Drafts PR comments and replies in the user's voice, in both directions.
sprint-board	Plans, writes and audits agile backlogs as markdown, with a scaffolder and a linter.
teach-me	Teaches a topic and certifies understanding with a scored retrieval check.
ai-engineering	AI and agent stack knowledge from a dated catalog, with read, write and build modes.

Writing, UI and media

Skill	What it does
ai-slop-magic-eraser	Rewrites prose to the user's voice with the meaning fixed, and flags what looks invented.
frontend-aesthetics	Raises UI past the defaults that read as AI-generated, and verifies the rendered result in a browser.
compress-media	/compress-media shrinks or converts video, audio and images with ffmpeg into a _small copy beside each original.

Security and station

Skill	What it does
sanitize	Deterministic scrubbing engine, and the commit and push gate behind the global git hooks.
my-security-review-checklist	Pre-merge security gate for agent tooling, including checks for hidden Unicode.
scan	/scan <path> updates ClamAV's signatures if needed, scans a file or folder, and checks apps, installers and programs with Gatekeeper, the signer, the quarantine tag and a SHA-256; --sweep lists what in a folder was made on the Mac rather than downloaded. Report-only.
skill-authoring	House profile for authoring and auditing skills, subagents, commands and hooks.
agent-cc-configs-sync	Seeds or reconciles a machine's harness configs and hooks against specs/.
agent-mail	File-based messages between agents working separate repos on the same machine.

Subagents (agents/*.md)

Subagent	What it does
advisor	Consulted by meta-loop on strategy, decomposition, risk and proportion.
ai-engineer	Fresh-context builder for delegated AI and agent work.
curator	Owns how reader-facing work reads and looks; edits in place with ai-slop-magic-eraser.
my-security-reviewer	Applies the security checklist to a diff before it ships.
reader	Read-only fan-out worker: answers one bounded question with findings.
researcher	Source-cited researcher for one angle; the deep-research worker.
supervisor	Spawned only by focus; flags drift and landmines with a root-cause section.
support-local	The support-local worker, pinned to a Haiku model; cannot spawn agents.
worker	Fan-out worker that changes the tree and returns evidence of what it changed.

Commands (commands/*.md)

Command	What it does
/advisor	Asks the advisor subagent one question about a plan.
/agent-mail	One agent-mail action: send, read, list or reply.
/cmon	Restates the last verbose reply in as few words as possible.
/my-security-review	The agent-tooling security review, with my-security-reviewer for depth.
/reflect	Memory reconciliation: propose, wait for approval, apply.
/wrap-up	Closes the session: review, inbox and notes, commit and push, then reflect.
/spec /plan /build /test /review /ship /code-simplify	The house procedure for each stage, self-contained.

The seven stage commands defer to their counterparts in Addy Osmani's agent-skills plugin when it is installed, and work without it. Upstream skills come from installed plugins, never from this tree; SPEC-CLAUDE-CODE.md section 3 lists them.

In practice

meta-loop runs long sessions as an orchestration loop. Explore subagents search in parallel and return only their findings. The orchestrator curates them into a synthesis, sends it to the advisor for review, then delegates the writing and editing to a second wave of subagents. Each result is checked against acceptance criteria and evidence rather than the worker's own summary.

No single agent fills its context window, so several short threads cost less in total than one long thread that maxes out and compacts. Each call names its model tier rather than inheriting one.

The status line shows the model, folder, branch, reasoning effort, context, session cost and rate-limit use on every prompt. Both scripts and what each field means are in specs/claude-code/, section 9.

Design principles

Every skill here fixes something the model gets wrong on its own. frontend-aesthetics exists because default UI taste is bad. There is no generalized backend skill, because there is no generalized backend mistake to correct, and a skill that only repeats what the model already knows never fires anyway. Retired skills wait in skills-archive/ until their stack is in play again.

Prose does not stop an agent from doing anything. Tell a subagent it is read-only and hand it Bash, and it will edit your files. Advisory agents get a read-only tools: allowlist instead, because that is the part the harness enforces.

Nothing grades its own homework. A checklist run by the model that wrote the code is theater, so the mechanically checkable parts ship as scripts that exit non-zero: slop_check.py for machine-writing tells, unicode_smuggle_check.py for instructions hidden in invisible characters. Judgment stays in the prose.

Layout

~/.agents/
  skills/          one folder per skill, each with a SKILL.md and its own scripts and tests
  skills-archive/  retired skills; not synced, restored with git mv
  agents/          one .md per subagent
  commands/        one .md per slash command
  hooks/           cross-harness hook scripts and their manifest
  hooks/git/       global git hooks: the sanitize gate at commit and push
  specs/           one folder per harness: station spec and seed files
  tests/           station hook suites and repo-wide checks
  app/             the agents chat app, its own uv project
  sync-skills.sh   links skills, agents and commands into ~/.claude
  AGENTS.md        the rules every agent works under

Third-party plugins load from their own cache under ~/.claude/plugins/ and never read or write here, so updating one cannot change a skill in this tree.

More

• AGENTS.md - the rules every agent works under, in every harness. Claude Code loads it through the one-line pointer in specs/claude-code/CLAUDE.md.example.
• SPEC.md - what this repo is, its scope and invariants.
• CONTRIBUTING.md - setup, branches, skill conventions, verification and style.
• SECURITY.md - reporting a vulnerability in a repo whose payload is instructions an agent executes.

License

MIT.