project-agent — Claude Code skill: setup and capabilities
This Claude Code skill keeps an ARIS project agent working across sessions with Herdr or Orca. It restores project context and guides bounded delegation between roles.
MIT ★ 173 updated 2026-11-16
npx skills add wanshuiyin/SuperARIS --skill project-agent -a claude-codeWhat project-agent does
The skill starts or resumes an ARIS project agent while preserving project memory. Before assigning new work, it directs the agent to read CLAUDE.md, applicable AGENTS.md files, the current status, and the recovery entrypoint, then identify the owner of active work.
Separate instructions cover startup, role coordination, project memory, and receipt handling. The skill distinguishes Herdr from Orca and checks which session owns the work. Without a verified session association, it prepares startup instructions or handles ordinary file tasks. Product commands run through installed superaris or python -m superaris.
What it helps with
- restore an active project task from its status and recovery entrypoint
- delegate work to roles with bounded responsibilities
- process receipts and continue the next authorized step
- verify which Herdr or Orca session owns a process
Example prompts for the agent
- Resume the active project task using its memory and recovery entrypoint.
- Prepare a Herdr agent startup for this project and verify its session association.
- Split this task across roles and process receipts as work progresses.
How to install project-agent
1Skills CLI (Vercel)
npx skills add wanshuiyin/SuperARIS --skill project-agent -a claude-code2Manually: put the files in place
~/.claude/skills/project-agent/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/project-agent/
1Skills CLI (Vercel)
npx skills add wanshuiyin/SuperARIS --skill project-agent -a codex2Manually: put the files in place
.agents/skills/project-agent/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add wanshuiyin/SuperARIS --skill project-agent -a cursor2Manually: put the files in place
.cursor/skills/project-agent/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install wanshuiyin/SuperARIS src/superaris/resources/claude/project-agent2Manually: put the files in place
.github/skills/project-agent/SKILL.md1Skills CLI (Vercel)
npx skills add wanshuiyin/SuperARIS --skill project-agent -a gemini-cli2Platform CLI
gemini skills install https://github.com/wanshuiyin/SuperARIS --path src/superaris/resources/claude/project-agentManually: put the files in place
.devin/skills/project-agent/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about project-agent
How do I install project-agent in Claude Code?
Run npx skills add wanshuiyin/SuperARIS --skill project-agent -a claude-code. The skill file is at ~/.claude/skills/project-agent/SKILL.md.
What does project-agent do?
It starts or resumes an ARIS project agent with project memory, role delegation, and receipt handling. Before new assignments, it restores the active task and its owner.
What does project-agent need to run?
The project needs CLAUDE.md and applicable AGENTS.md files. Product commands require installed superaris or the ability to run python -m superaris. Herdr or Orca use depends on the environment; Herdr work requires a verified session association.
Source text
SKILL.md · MIT
--- name: project-agent description: Start or continue a Herdr- or Orca-backed ARIS project agent with existing project memory, bounded role delegation and evidence-based follow-up. Use for a persistent project owner or a supported workbench; ordinary one-off skills remain directly usable. --- # Project Agent Use these bundled instructions to maintain the requested project's continuity with ARIS pipelines, workflows or individual skills. Their original contracts still govern. Product commands use installed `superaris` or `python -m superaris`; this does not require globally installing the skill or an ARIS product wrapper. Read the project's `CLAUDE.md` and applicable `AGENTS.md`, current status and recovery entrypoint first. Restore active work and its real owner before making new assignments; reading this skill does not designate you as coordinator or resume paused work. For startup, role presets and helper resolution, read [references/launch.md](references/launch.md). Read the packet/navigation workbench host and original launch.json binding; missing host on an older binding means Herdr. For Orca, read [references/orca-host.md](references/orca-host.md) and follow its terminal/runtime identity checks; Herdr environment markers and commands do not identify an Orca session. Before Herdr scheduling, verify the actual environment: `HERDR_ENV=1` is one signal, but backend shell calls may lack it. In that case use an explicit user-confirmed session/pane association and official API verification of the caller's native session/process and cwd. Fully read `herdr --skill`, then check local `herdr --help` and the relevant command group's help. An explicitly user-invoked launcher can bootstrap the named project/session. Otherwise, without verified association, prepare startup instructions or do ordinary file work. Never synthesize the marker or take over the user's focused pane. For delegation, writing roles, project memory, receipt handling or recovery, read [references/coordination.md](references/coordination.md). Use native callbacks when available, otherwise the project-limited runtime bridge; inspect real outputs and continue the next authorized step. Protect human input and reconcile lost events before repeating operations. For the executable start → watch → report → receiver → acknowledge chain and explicit memory events, read [references/receipts.md](references/receipts.md). Load the selected host-correct ARIS skills and their required references in each executing role. Files being present or discoverable does not prove the role actually read them or has the required capabilities. Retain existing MCP routes, reviewer freshness, writer ownership, acceptance and budget boundaries. Do not substitute persistent panes for every temporary worker or independent reviewer.
© wanshuiyin · MIT · wanshuiyin/SuperARIS