Ultracode Skill gives Codex a dynamic workflow layer for serious coding tasks.
Use it when a task needs more than one pass: discovery, implementation, review, tests, and a final integration step. The skill tells the agent how to decide between a direct edit, a packetized workflow, or native agent delegation.
Codex is the main target. Claude Code already has first-class workflow concepts; Codex has native multi-agent tools, but users still need a clean operating pattern around them. Ultracode fills that gap with one small SKILL.md folder.
Ultracode asks the agent to:
- Define the goal and success criteria.
- Decide whether the task needs direct, workflow, or delegated mode.
- Create workflow artifacts for non-trivial work.
- Use native agents when they are useful and allowed by the host.
- Keep integration in the parent session.
- Verify the result before answering.
For Codex, that means the skill can use native spawn_agent for independent packets, then integrate the results in the parent session.
The installable skill folder is ultracode/.
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
cp -R ultracode "${CODEX_HOME:-$HOME/.codex}/skills/"Restart Codex after installing.
mkdir -p "$HOME/.claude/skills"
cp -R ultracode "$HOME/.claude/skills/"Workspace install:
mkdir -p .agents/skills
cp -R /path/to/ultracode .agents/skills/User install:
mkdir -p "$HOME/.gemini/antigravity/skills"
cp -R ultracode "$HOME/.gemini/antigravity/skills/"Start with the short form:
Use $ultracode to build this feature end to end.
You do not have to manually design the workflow or ask for subagents. Ultracode decides whether the task needs a direct edit, a workflow, or native agents.
If you want to force a delegated run, you can still say it directly:
Use $ultracode. Split this across agents where it is safe, keep integration in the parent session, and verify the final patch.
| Mode | Use it for | Behavior |
|---|---|---|
| Direct | Small, clear edits | No workflow files unless needed. Run the narrowest useful check. |
| Workflow | Multi-step work without useful delegation | Write plan, packet, result, integration, and final report files. |
| Delegated | Independent packets where native agents are available | Spawn bounded agents, keep the critical path local, integrate results, verify. |
For non-trivial tasks, Ultracode writes plain Markdown and JSON:
.workflow/ultracode/<run-slug>/
plan.md
orchestration.md
state.json
packets/
results/
integration.md
final-report.md
These files make the run inspectable. You can see what was delegated, what changed, what passed, and what risk remains.
When Codex exposes native agent tools, Ultracode prefers real delegation over fake packet simulation.
The parent session should:
- keep the critical path local
- spawn
exploreragents for read-only discovery - spawn
workeragents only with clear file ownership - avoid combining
agent_typewith a full-history fork - wait only when a result blocks the next parent step
- close agents after collecting their results
If native agents are unavailable or blocked by policy, Ultracode falls back to workflow mode and says so.
This skill was tested in Codex with a fresh demo repository.
Result:
$ultracodeloaded the skill.workflow/ultracode/...artifacts were created- Codex used native
spawn_agent - tests passed
- the parent session integrated the agent results
Ultracode Skill does not ship a background service, hidden runner, or required scripts. It is a small skill folder that teaches the agent how to run high-effort coding workflows with the tools already available in the current host.
The repo is ultracode-skill.
The skill name is ultracode, so users can invoke it with:
$ultracode
MIT.