Horizon AIOS — Skills Reference

AIOS skills are slash commands exposed in the Claude Code harness UI. Each skill is a SKILL.md file that tells Claude how to drive a task: what steps to take, what scripts to call, and what to surface to the user. They are distinct from Claude Code marketplace plugins (code-review, update-config, verify, etc.), which are installed separately from Anthropic's marketplace and are not part of this repo.

Skills live in one of two OS-level directories:

$HORIZON_SYSTEM/skills_bin/ — available to all users, including brain accounts
$HORIZON_SYSTEM/skills_sbin/ — owner only; brain accounts have an explicit Deny ACL on this directory

At bootstrap (or when /resync-user-skills is run), the owner's ~/.claude/skills/ is pointed at skills_sbin/, and symlinks are created inside skills_sbin/ for each skills_bin/ skill so the owner sees the full set flat. Brain users' ~/.claude/skills/ points directly at skills_bin/.

Quick-scan table

Slash command	Location	What it does
`/context-cost`	`skills_bin/`	Report KB, word count, and token estimates for all files the harness auto-loads above a given path
`/create-brain`	`skills_sbin/`	Provision a new brain OS user, groups, workspace, shell profile, and keystore credential (Admin/root)
`/doctor`	`skills_bin/`	Run the AIOS health check — env vars, skills symlink, hooks, registry, and privileged-dir Deny ACLs
`/handoff`	`skills_bin/`	Write a structured session handoff document for the next session or a human reviewer
`/harden`	`skills_sbin/`	Apply the authoritative brains-group ACL model to the AIOS layer (Admin/root)
`/agent-teams`	`skills_sbin/`	List, create, or edit agent-team definitions in `local.agent_teams.md` at any scope; manage custom role flags
`/horizon_aios_dev_consistency_check`	`skills_sbin/`	Run an iterative docs/implementation consistency validation pass against the AIOS check standard
`/horizon_aios_documentation_index_update`	`skills_sbin/`	Rebuild `documentation/index.md` so every doc is registered with a stable path-based ID
`/horizon_aios_wiki_upkeep`	`skills_sbin/`	Run a consistency pass between the operational wiki and its source documentation; fix unambiguous drift, surface judgment calls
`/terseness-check`	`skills_sbin/`	Evaluate every file in the Terseness Contract Index for context overhead; report FAIL/ADVISORY findings with file:line evidence and cut suggestions
`/test-agent-teams`	`skills_sbin/`	End-to-end self-test of the Agent Teams system — walk every team, spawn each role to echo a nonce + role + model, report PASS/FAIL per team
`/model-catalog-refresh`	`skills_bin/`	Fetch live model+pricing data from Anthropic, OpenAI, Gemini, and Ollama; diff against the current model-preference config
`/userguides`	`skills_bin/`	Browse the operational wiki: no argument summarizes all sections via parallel Haiku agents; `/userguides N` displays section N in full
`/model-prefs`	`skills_sbin/`	Author or inspect model groups, per-session slots, and task-class routing rules in the gitignored extend file
`/model-prefs-assign`	`skills_sbin/`	Audit skills for model-preference group callouts and assign or refresh them; keep indexes in sync
`/model-prefs-test`	`skills_bin/`	Test how each model group resolves in the current runtime (dry-run or live spawn)
`/monitor`	`skills_bin/`	Start the AIOS filesystem integrity monitor (watches system dirs, logs events as JSON lines; requires elevation)
`/objective`	`skills_sbin/`	Create, list, show, or update durable multi-session objectives that handoffs chain back to
`/pre-flight-tooling-validation`	`skills_sbin/`	Validate the repo ships full-lifecycle tooling per platform, then emit a ready-to-run test prompt per platform
`/remove-brain`	`skills_sbin/`	Deprovision a brain — remove its OS user, per-brain group, workspace, profile, and credential (Admin/root)
`/resync-user-skills`	`skills_sbin/`	Report skill inventory and rebuild symlinks so the owner view matches the filesystem
`/skill-creation`	`skills_sbin/`	Scaffold a new AIOS skill with correct structure, frontmatter, and index registration

Per-skill entries

/context-cost

Location: skills_bin/ (available to all users including brains) Underlying tool: $HORIZON_SYSTEM/bin/context_cost.py

Reports the Claude Code harness context overhead for a given path — how many KB, words, and estimated tokens will be loaded into every session that starts under that directory. It walks the ancestor chain collecting every CLAUDE.md, CLAUDE.local.md, agents.md, and @-import file the harness would pull in, and flags totals above 1,000 tokens as worth reviewing.

Onboarding: Gives immediate visibility into how much context overhead any directory imposes, enabling the token economy discipline described in dev_values.md. Offboarding: You lose the integrated harness-aware report format. The underlying context_cost.py script remains at $HORIZON_SYSTEM/bin/ and can be run directly: python context_cost.py [path] [--json].

/handoff

Location: skills_bin/ (available to all) Underlying tool: None — Claude drives the session from conversation memory

Captures the current session state and writes a structured handoff document (summary, decisions, deferred items, next-session entry point) to the handoffs directory. The directory defaults to $HORIZON_ROOT/handoffs/ and can be overridden per project via aios_overrides.md. Supports objective linkage so handoffs chain a durable goal forward across sessions.

Onboarding: Provides a consistent handoff format and path convention across all sessions, making session continuity reliable rather than ad hoc. Offboarding: You lose the structured format and automatic directory routing. Handoff documents are plain Markdown and remain readable; future sessions can still read them directly. The skill is the entry point — no underlying script.

/horizon_aios_dev_consistency_check

Location: skills_sbin/ (owner only) Underlying tool: $HORIZON_DOCS/development_tools/consistency_checks.md (the check definitions); delegates to a general-purpose subagent for file-heavy work

Runs the AIOS consistency validation protocol (check IDs CC-*) over the repo or a defined scope. Fixes unambiguous drift in place (stale paths, missing index entries, one-sided platform implementations), surfaces judgment calls for user input, and re-runs until 100% clean or blocked. The check definitions file — not this skill — is the source of truth for what is checked.

Onboarding: Provides a single, repeatable entry point for the consistency standard, ensuring docs and implementation stay synchronized after every change. Offboarding: You lose the automated iteration loop. The check definitions remain in $HORIZON_DOCS/development_tools/consistency_checks.md and can be followed manually; the underlying standard is not a skill.

/horizon_aios_documentation_index_update

Location: skills_sbin/ (owner only) Underlying tool: None — Claude reads and writes documentation/index.md directly

Rebuilds $HORIZON_DOCS/index.md by globbing all in-scope docs, deriving each entry's title and one-line purpose, and writing a grouped, path-keyed catalog. Add/remove/rename operations are reflected deterministically. Satisfies the CC-G4 consistency check requirement that every doc has a stable index entry. Delegates to a subagent to avoid bloating the main session's context.

Onboarding: Keeps the documentation index accurate without manual upkeep; a new doc added without running this skill will be flagged by the CC-G4 consistency check. Offboarding: You lose the automated rebuild. The index file remains at documentation/index.md and can be maintained by hand; the format is documented in the skill and in the index file's own header.

/model-catalog-refresh

Location: skills_bin/ (available to all users including brains) Underlying tool: Web/bash access — the agent fetches live data at runtime

Fetches current model ids and pricing from Anthropic, OpenAI, Google Gemini, and Ollama. If the current ## Model Groups block is already in context, automatically diffs it against live data and flags stale member ids, newer alternatives, and pricing changes. Companion to /model-prefs, which writes the extend file.

Onboarding: Provides an up-to-date model+pricing reference so horizon_aios_model_prefs.local.md is populated with valid, current ids rather than training-cutoff guesses. Offboarding: You lose the guided, multi-provider fetch. Provider docs remain publicly accessible and can be read directly; the catalog format and fetch strategy are documented in the skill's SKILL.md.

/model-prefs

Location: skills_sbin/ (owner only) Underlying tool: None — Claude reads and writes the extend file directly

Authors or inspects the gitignored model-preference extend file ($HORIZON_ETC/horizon_aios_model_prefs.local.md): model groups, per-session slot preferences, and task-class routing rules. The mechanism is in-context — the acting model reads the file each session and honors it by direct instruction; no scripts or env vars are wired. Writes nothing to the OS-tracked base file.

Onboarding: Provides a guided entry point for configuring the model-preference layer without needing to know the file grammar or fallback-order rules. Offboarding: You lose the guided edit and resolution report. The extend file is plain Markdown and can be edited directly; the grammar and fallback order are documented in $HORIZON_ETC/horizon_aios_model_prefs.md and $HORIZON_DOCS/system/model_preferences.md.

/objective

Location: skills_sbin/ (owner only) Underlying tool: None — Claude reads and writes objective files directly

Creates, lists, shows, and updates durable objective files (plain Markdown) in the objectives directory (default: $HORIZON_ROOT/objectives/). Objectives hold the long-term goal behind a stream of work — not a task tracker, but a durable notepad that handoffs reference by number so the goal survives across sessions. Pruning of stale objectives is handled externally by horizon_aios_maintain_logs.py.

Onboarding: Provides a lightweight, persistent goal-tracking layer that handoffs can chain back to, preventing goal drift across many sessions. Offboarding: You lose the structured create/list/show/update interface. Objective files are plain Markdown in the objectives directory and remain fully readable; they can be managed by hand. The pruning logic in horizon_aios_maintain_logs.py remains independent of the skill.

/resync-user-skills

Location: skills_sbin/ (owner only) Underlying tool: $HORIZON_SYSTEM/sbin/horizon_aios_register_user_skills.py

Reports whether the owner's aggregated skill view (skills_bin/ + usr_skills/ linked into skills_sbin/) matches the filesystem, then heals any drift by rebuilding symlinks. Also compares the on-disk state to what is loaded in the current session and advises on restarts. An upstream sync that refreshes skills_sbin/ can drop these links; running this skill or horizon_aios_register_user_skills.py --check detects that.

Onboarding: Provides a single command to verify and repair the skill wiring after any sync or new skill is added, without needing to understand the symlink mechanics. Offboarding: You lose the check/heal UX. The underlying horizon_aios_register_user_skills.py script remains at $HORIZON_SYSTEM/sbin/ and can be run directly with --check or without flags to rebuild.

/skill-creation

Location: skills_sbin/ (owner only) Underlying tool: None — Claude creates files and updates indexes directly

Scaffolds a new AIOS skill at the correct tier (skills_bin/, skills_sbin/, or usr_skills/), writes a valid SKILL.md with required YAML frontmatter, updates the tier's index.md, and for skills_sbin/ skills also updates .gitignore to whitelist the new directory. Enforces the invariant that name, directory, and frontmatter name: field must all match. OS skills require a restart; the symlink is live immediately.

Onboarding: Ensures every new skill is correctly structured, registered, and git-tracked from the start — misstructured skills are silently ignored by Claude Code. Offboarding: You lose the guided scaffolding. Skills can be created by hand by following the structure in any existing SKILL.md; the requirements are also documented in skill-creation/SKILL.md and both tier index files.

/agent-teams

Location: skills_sbin/ (owner only) Underlying tool: $HORIZON_SYSTEM/bin/resolve_agent_teams.py (for discovery); writes local.agent_teams.md and local.agent_team_flags.md directly

Bare /agent-teams lists all resolved agent teams in effect at the current path by running resolve_agent_teams.py. With arguments, creates or edits team definitions and custom role flags in the scope-appropriate local.agent_teams.md (gitignored, never clobbered by sync). Never modifies the shipped $HORIZON_ROOT/agent_teams.md.

Onboarding: Provides a guided entry point for creating and managing multi-agent workflow definitions without hand-editing markdown; discovers teams at the correct scope automatically. Offboarding: Edit local.agent_teams.md at the desired scope directly. See $HORIZON_ROOT/agent_teams.md for the team definition format and $HORIZON_DOCS/system/agent_teams.md for the full invocation and cascade reference.

/horizon_aios_wiki_upkeep

Location: skills_sbin/ (owner only) Underlying tool: None — delegates section checks to #investigate subagents, then edits the wiki directly

Runs a consistency pass between $HORIZON_DOCS/user_guides/using_your_aios.md and its source documents. Spawns one subagent per section to check for factual drift, stale paths, outdated examples, and coverage gaps. Applies unambiguous fixes; surfaces judgment calls for user decision. Re-runs until all sections are CLEAN or remaining issues are explicitly deferred.

Onboarding: Keeps the operational wiki accurate after source docs change, without requiring a full manual audit. Offboarding: Audit sections manually by reading the wiki and source docs side-by-side. The source-doc mapping is in $HORIZON_SYSTEM/skills_sbin/horizon_aios_wiki_upkeep/SKILL.md.

/terseness-check

Location: skills_sbin/ (owner only) Underlying tool: None — Claude reads the Terseness Contract Index and evaluates each tracked file directly

Evaluates every file listed in $HORIZON_DOCS/terseness_contract_index.md against the seven terseness criteria defined there. Reports FAIL findings with file:line evidence and concrete cut suggestions. Gitignored / user-controlled files are evaluated but findings are ADVISORY only. Auto-fixes pure-removal FAIL findings (no new content written); surfaces everything else for confirmation. Also delegated by CC-T2 in the consistency check.

Onboarding: Ensures context-loaded files stay token-efficient; a single command confirms the Terseness Contract is met before pushing changes. Offboarding: Apply the seven criteria from $HORIZON_DOCS/terseness_contract_index.md manually to each tracked file.

/test-agent-teams

Location: skills_sbin/ (owner only) Underlying tool: $HORIZON_SYSTEM/bin/resolve_agent_teams.py (team discovery); spawns real subagents per role

End-to-end integration test of the Agent Teams system. Runs resolve_agent_teams.py --json to enumerate all resolved teams, then spawns one real subagent per role across all teams. Each role subagent echoes a nonce, its role name, and the model it actually ran as. Verifies that every role spawned, the #model-group routed correctly, and the chain executed. This is a deliberate, potentially costly integration test — run to verify, not routinely.

Onboarding: Proves the full agent-team loop works end-to-end (resolution, model routing, nonce proof of execution) before relying on agent teams for real work. Offboarding: Test teams manually by invoking a team and observing which roles spawn and on which models.

/userguides

Location: skills_bin/ (available to all users including brains) Underlying tool: None — Claude reads the wiki file and optionally spawns Haiku summarizer agents

Bare /userguides reads $HORIZON_DOCS/user_guides/using_your_aios.md, discovers all numbered sections dynamically, spawns one #lowcost (Haiku) agent per section in parallel to write a 3–5 sentence summary, and presents summaries in section order. /userguides N displays the full text of section N. The section map in the SKILL.md is informational; the live file read is authoritative.

Onboarding: Provides a one-command overview of the full operational wiki without reading thousands of words; section N access gives targeted reference lookup. Offboarding: Read $HORIZON_DOCS/user_guides/using_your_aios.md directly.

/create-brain

Location: skills_sbin/ (owner only) Underlying tool: $HORIZON_SYSTEM/sbin/horizon_aios_create_brain.py

Provisions a new brain: creates the OS user account, the shared brains group and a per-brain group, the workspace folder at $HORIZON_ROOT/brains/<name>/, a login shell profile, and stores an auto-generated 64-char password in the OS native keystore. Requires Administrator/root; delegates to horizon_aios_create_brain.py.

Onboarding: Provides a single guided entry point for brain provisioning that handles all OS-level steps and surfaces the post-provisioning checklist. Offboarding: Run horizon_aios_create_brain.py directly; see the "Adding a Brain" section of getting_started/ReadMeToSetupYourSystem.md.

/remove-brain

Location: skills_sbin/ (owner only) Underlying tool: $HORIZON_SYSTEM/sbin/horizon_aios_remove_brain.py

Deprovisions a brain by removing its OS user account, per-brain group, workspace folder, profile config (including the skills symlink), and stored credential. The shared brains group is left intact. Requires Administrator/root.

Onboarding: Provides a safe, guided teardown path that mirrors provisioning. Offboarding: Run horizon_aios_remove_brain.py directly (see utilities.md).

/harden

Location: skills_sbin/ (owner only) Underlying tool: $HORIZON_SYSTEM/sbin/horizon_aios_harden.py

Applies the authoritative brains-group ACL model to the AIOS layer: grants the brains group Read+Execute on bin/ and skills_bin/, applies a no-write Deny across $HORIZON_SYSTEM, and adds an explicit full Deny on sbin/, skills_sbin/, and logs/. Run after horizon_aios_doctor.py reports a missing Deny ACE, or after structural changes to $HORIZON_SYSTEM. Requires Administrator/root.

Onboarding: Provides a single command to re-establish the security invariants without needing to recall the icacls/chmod commands manually. Offboarding: Run horizon_aios_harden.py directly (see utilities.md).

/doctor

Location: skills_bin/ (available to all users including brains) Underlying tool: $HORIZON_SYSTEM/sbin/horizon_aios_doctor.py

Runs the AIOS health check: verifies env vars, the ~/.claude/skills/ symlink, git hooks, local config (aios_local.conf), the AIOS registry, and that sbin/, skills_sbin/, and logs/ have an explicit Deny ACE for the brains group. Reports PASS / WARN / FAIL for each check. Optional --post-setup flag adds sound, statusline, and GPG signing checks.

Onboarding: Provides immediate post-install verification and ongoing health checks without needing to know which files to inspect. Offboarding: Run horizon_aios_doctor.py directly (see utilities.md).

/monitor

Location: skills_bin/ (available to all users including brains) Underlying tool: $HORIZON_SYSTEM/sbin/horizon_aios_monitor.py

Starts (or explains) the AIOS filesystem integrity monitor, which watches the AIOS system directories for create, modify, delete, and move events and appends each as a JSON-line audit record to $HORIZON_SYSTEM/logs/horizon_aios_monitor/. Requires elevation because writing to logs/ is an admin-only action.

Onboarding: Provides a guided entry point for enabling filesystem audit logging, including the watch set, log format, and service-registration options. Offboarding: Run horizon_aios_monitor.py directly; see security/audit_logging.md for full setup and SIEM integration guidance.

/model-prefs-assign

Location: skills_sbin/ (owner only) Underlying tool: None — Claude reads and edits skill files directly

Audits all skills in skills_sbin/ and skills_bin/ for a Model preference: callout in the body. Adds the callout to skills that lack one (defaulting to #midcost), updates it where a different group is more appropriate, and keeps the Model group columns in both tier index files in sync with the actual skill bodies.

Onboarding: Ensures every skill declares its model group so the model-preference layer routes work correctly without per-invocation overrides. Offboarding: Skills can be audited and updated by hand; the format is documented in system/skill_model_groups.md.

/model-prefs-test

Location: skills_bin/ (available to all users including brains) Underlying tool: None — Claude resolves and optionally spawns agents directly

Tests how each model group from the current model-preference config resolves in this runtime: dry-run mode reports which member would be selected without spawning anything; --live mode spawns a small agent per group to confirm the harness actually honors the selection and has each agent self-report its model.

Onboarding: Provides confidence that the model-preference config is wired correctly before relying on it for real work. Offboarding: Test group resolution manually by reading the extend file and checking which members are available in the current runtime.

/pre-flight-tooling-validation

Location: skills_sbin/ (owner only) Underlying tool: None — Claude inspects repo contents directly

Validates that the AIOS repo ships every tool needed to run the full lifecycle (install → create brain → create a second AIOS → switch → backup → delete) on each supported platform (Windows, Linux, macOS), then emits a ready-to-run admin/sudo test prompt per platform. Intended as a pre-test gate before handing off a fresh clone to someone running a lifecycle test on a clean machine.

Onboarding: Catches missing or misnamed tooling before a test run so failures are a tooling gap, not a test environment issue. Offboarding: Check the tool list manually against file_structure_invariants.md and utilities.md.

Onboarding and offboarding summary

When AIOS skills are registered, an owner gains 22 slash commands across two tiers. skills_bin/ (available to brains): /context-cost, /doctor, /handoff, /model-catalog-refresh, /model-prefs-test, /monitor, /userguides. skills_sbin/ (owner only): /agent-teams, /create-brain, /harden, /horizon_aios_dev_consistency_check, /horizon_aios_documentation_index_update, /horizon_aios_wiki_upkeep, /model-prefs, /model-prefs-assign, /objective, /pre-flight-tooling-validation, /remove-brain, /resync-user-skills, /skill-creation, /terseness-check, /test-agent-teams.

Skills that wrap underlying scripts: /context-cost, /doctor, /create-brain, /harden, /monitor, /remove-brain, /resync-user-skills (all delegate to a sbin/ or bin/ Python script). /model-catalog-refresh and /model-prefs-test perform live fetches or agent spawns at runtime. The remaining skills have no separate script — Claude is the engine and the SKILL.md is the procedure.

Unregistering the skills (by removing the ~/.claude/skills/ symlink or clearing the skills directories) does not delete any underlying data or scripts. Handoff documents, objective files, the documentation index, and the model-preference extend file remain on disk exactly as written. The Python scripts in $HORIZON_SYSTEM/bin/ and $HORIZON_SYSTEM/sbin/ remain fully functional from any shell. What is lost is convenience: the single slash command, the guided step-by-step procedure, and the integration with the harness context. A user offboarding from AIOS retains all artifacts produced by the skills; they lose only the automated workflow that produced them.

Brain accounts see only skills_bin/ (/context-cost, /doctor, /model-catalog-refresh, /model-prefs-test, /monitor) and gain no access to the owner-only skills. This is by design and enforced at the ACL level.