How to use the Agent Startup and Context Ingestion agent skill

What the skill makes the agent do

The skill teaches an agent how startup context should be designed and audited. It covers AGENTS.md overlays, role files, skills, rig specs, workflow specs, startup checklists, refocus messages and the rig context surface. It also names four failure modes that make startup context fail: an old rig spec that misses the current operating mode, current agents never told about new guidance, a startup file that turns into a dumping ground, and an orchestrator that passes on implementation without the product intent behind it.

The skill tells the agent to look up the current task first and read only the context that task needs, rather than loading everything. Startup files are treated as maps that point to canonical sources, not as copies of those sources.

Install the skill from the ZIP

Download the ZIP for this skill from this page and unzip it. The skill is delivered as a folder containing SKILL.md. You can place it in one of two locations: ~/.claude/skills/ for your personal skills, or .claude/skills/ inside a project if you want it available only to that project.

If you use the Claude apps instead, you can upload the skill under Settings → Capabilities → Skills.

• Personal: unzip into ~/.claude/skills/
• Per project: unzip into the project's .claude/skills/ folder
• Claude apps: upload under Settings → Capabilities → Skills

Trigger the skill with natural requests

You do not need a special command. Describe the job in plain words and mention the situation. The skill's description is written around phrases like designing or auditing how an agent becomes useful after launch, so requests that use those ideas are the most likely to match.

Good requests name the artefact or problem: a new role file, an agent that seems to be running old instructions, or a startup checklist that has grown too long.

• Help me design the startup files for a new engineer agent in our OpenRig rig.
• Audit whether the running agents in this rig picked up the new operating guidance.
• Our startup/context.md has become huge. Show me how to split it so it only points to canonical sources.
• Review this refocus message and check that it preserves the product intent, not just the implementation steps.

Example: auditing a rig's startup

Suppose you changed the operating mode for your rig and want to know whether the agents will act on it. Ask the agent to audit the startup path. Using the skill, it will check the four failure modes, in particular whether the rig spec reflects the current operating mode and whether running seats were told about the change.

The skill notes that editing a file does not update a session that is already running. Adoption needs an authorised refresh or a fresh launch, so the audit should look for evidence of that rather than assume it.

What you need before it works

The skill assumes an OpenRig installation that provides the rig command-line tool, with the rig whoami and rig context commands. It also refers to other OpenRig docs and sibling skills, such as mission-slice-sop and openrig-operating-model. Without these, several steps cannot run as written, so the skill is of limited use outside OpenRig.

The skill does not run scripts and does not use the network, according to its metadata.

• OpenRig installation providing the rig CLI (rig whoami, rig context)
• Referenced OpenRig docs and sibling skills for full functionality
• Scripts: none. Network: none.

Safety notes

The skill itself runs no scripts and makes no network calls. It does read and reason about startup files and rig state, so review what it reads in sensitive projects. The skill also says that permission to read an input is not permission to rewrite its source. Writes follow the active project or rig policy and assignment.

The proof standard in the skill asks for tests in a disposable fresh session. It advises against clearing or re-priming a live seat just to test wording.

• Read-only in intent: it audits and designs, and writes follow your project policy.
• Test changes in a disposable fresh session, not on a live seat.
• Check the actual reads and next action the agent takes, not only the edited prose.

Tips for better results

Keep startup files short and point them at canonical sources. The skill treats a startup file that tries to hold everything as a failure mode. Keep reusable methods in skills and identity content in startup files, as the skill's comparison table explains.

When you ask for an audit, give the selected task or slice so the agent can resolve the current path instead of reading old packets.

• Put role, culture and context identity in startup files; put reusable SOPs in skills.
• Do not expect per-file acknowledgements or quizzes; the skill advises against them.
• An old packet does not choose the current work. Name the current task explicitly.