Skip to content
CarryCtx

Agent Skills

CarryCtx ships a single agent skill — a SKILL.md-based instruction pack that any supporting agent (Claude Code, Codex, OpenCode, Cursor, Antigravity, Gemini CLI, …) can load. It teaches the agent how to use CarryCtx properly: not just which commands exist, but when to run them, in what order, and with what discipline.

The skill lives in the carryctx-skills repository and installs with the Vercel Labs Skills CLI:

Terminal window
npx skills add https://github.com/Xuepoo/carryctx-skills --skill use-carryctx -y

use-carryctx: the commander entry point

Section titled “use-carryctx: the commander entry point”

Load it once per main session, at the start of any multi-step engineering effort. The skill casts your main-session agent as the commander: you plan and verify; subagents implement. The loop:

  1. Plan — decompose the goal into tasks with explicit dependencies and scopes. Small one-line fixes stay inline; tightly-coupled work batches to one subagent instead of fanning out.
  2. Assign — encode the plan durably so nothing depends on the conversation surviving: task create --depends-on/--team/--required-role plus task scope add for edit-conflict detection.
  3. Dispatch — hand each subagent exactly its slice: a task ID plus team context --agent-for <sub> output, ideally isolated in its own Git worktree via worktree create <TASK_REF> (<path>/.worktrees/<task-id> on branch carryctx/<task-id>).
  4. Accept & verify — never trust a subagent’s self-report. Read state back with team status, team context, task show, and search; verify the diff and run the tests yourself before task complete.

Supporting discipline is baked into the same doctrine: record durably because subagents die silently (progress todo/note/block/risk, checkpoint, decision add); keep your own context lean (--compact output, outlines before reads, re-reading team status between rounds instead of holding state in conversation); and close what you open — end sessions cleanly and checkpoint before merging branches away.

Why a single skill?

Section titled “Why a single skill?”

Earlier releases shipped five separate skills covering daily usage, rules, workflows, personas, and handoffs. They are now consolidated into use-carryctx:

  • One install, one invocation per session. The commander loop needs all of these capabilities anyway, so there is nothing left to pick between — installing one skill covers the full surface.
  • Progressive disclosure via bundled references. The SKILL.md keeps the commander doctrine and a verified command quick-reference inline; deep detail lives in references/ guides that the agent reads on demand when operating in that area:
Reference guide Read it when…
references/task-lifecycle.md Working with task states, transitions, dependency gating, file scopes
references/team-coordination.md Running teams, read-only projections, dispatch patterns, recording discipline
references/sessions-and-checkpoints.md Managing session lifecycle, stale detection, checkpoint policy
references/handoffs.md Writing and taking over handoffs, routing documents, accept/reject/close
references/presets-rules-personas.md Installing SOPs, domain rules, and personas into .carryctx/
references/troubleshooting.md Interpreting error codes, recovery steps, diagnostics

Handoffs deserve a special note: when a session ends with unfinished work, the same skill walks you through capturing state into a routed handoff document (handoff create --target …) and taking one over on the receiving side (handoff list → handoff accept --claim-task) — the record answers what happened, the document answers what to do next.