How to use the Harbor Intake Triage agent skill

What does the Harbor skill make the agent do?

Harbor sweeps incoming issues and pull requests, checks each claim against the repository and the tracker, and reuses decisions the maintainer has already made. The agent acts on its own only for reading, searching, reproducing under safe conditions, gathering evidence and carrying out actions a maintainer has already authorized. Anything new, such as accepting, rejecting, merging, granting an exception or changing scope, is prepared as a single question with options, a recommendation, impact and evidence, and waits for the maintainer's signature.

The skill states that it never merges.

• Verifies every claim before deciding what to do with it
• Reuses prior signed decisions instead of re-asking the same question
• Prepares a docket of pending items for the maintainer
• Records its work as comments, labels and links in the tracker

What do you need before installing Harbor?

Harbor is designed for a specific setup. The skill file lists three requirements, and it only works with a remote GitHub repository as the tracker in this version.

• GitHub CLI (gh) installed, with access to the target repository
• A remote GitHub repository used as the issue tracker
• A tracker label vocabulary recorded by the companion skill /oh-my-claudecode:drydock

How do you install the Harbor skill?

Download the ZIP from the skill's page on this directory and unzip it into one of two places. Use ~/.claude/skills/ to make the skill available in your user account, or use a project's .claude/skills/ folder to make it available only in that project.

You can also add it through the Claude apps: open Settings, go to Capabilities, then Skills, and upload the skill there.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into .claude/skills/ inside the project
• App install: Settings → Capabilities → Skills, then upload

How do you trigger Harbor with natural requests?

Harbor is set to run only when you invoke it, so the agent will not start triage on its own. The skill's argument hint lists the four ways to call it. Phrase your request close to one of these forms.

Each request maps to a specific job: a sweep looks over what has changed, a look at a numbered item inspects one issue or PR, a sign command records a decision you have made, and a readiness question asks what is ready for you.

• sweep: review what has changed in incoming work
• look at #N: inspect one issue or pull request by its number
• sign ...: record your decision on a pending proposal
• what's ready?: list items waiting for you

What does a Harbor run look like in practice?

Before anything else, Harbor checks that the remote tracker responds and that it is the expected repository. It does this with gh repo view on the recorded or implicit remote. If the remote cannot be reached, Harbor stops and reports the blocker instead of writing to local files.

Every action leaves one of four records in the tracker: verification evidence, a proposal, a decision, or an execution result. A proposal names one explicit question and includes a recommendation, the alternative, the scope and non-goals, and linked evidence. When you sign a decision, the record names who authorized it and where that authorization came from.

• Step 0 checks the remote tracker is reachable and correct
• Verification records the command or method used and what was not verified
• Proposals contain exactly one main question
• Changes to a signed proposal produce a new snapshot rather than an edited one

What are the safety limits and notes on scripts and network access?

The skill file states that it runs scripts and uses the network. Its network use is to reach the GitHub repository through the GitHub CLI. Harbor will not write to local markdown files as if they were a tracker, so dispositions end up where reporters and maintainers can see them.

Authority is limited. An API token, a reporter saying 'approved', or the model's confidence does not count as a disposition. Standing rules must be explicitly signed by the maintainer, and a rule only allows an action when the covered facts are verified, the rule permits the action and the host permits it.

Only one harbor writer should run per repository. If exclusivity cannot be confirmed or another writer is detected, Harbor produces drafts only and makes no tracker writes.

• Runs scripts and uses the network, so review the files before installing
• Never merges; merge decisions remain with the maintainer
• Standing rules are the only way to let the agent act without a new signature
• Use one writer per repository to avoid stale writes

Tips for getting good results from Harbor

Ask for a rule when you see the same report pattern again and again. The skill is built so that one signed standing rule replaces many repeated signatures. A good rule names the scope it covers, the allowed actions, the facts that must hold, and the exceptions, such as regressions, security reports or an invalidated parent issue.

Keep your label vocabulary recorded before the first run, since Harbor relies on the tracker labels it sets up and reads.

• Write standing rules with explicit exceptions
• Check the label set exists before the first run
• Expect a closed parent issue to not automatically make new reports closeable