Guide · Agent skill
How to use the Launch Delivery Pipeline agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Launch is an agent skill for oh-my-claudecode, a framework for Claude-based coding agents. It makes the agent run a structured delivery process for a software change: check the repository's documentation, turn a mission brief into a spec and ordered tickets, work the ready tickets in parallel, verify the result and write a completion report. This guide explains what the skill does, how to install it, how to invoke it, what it needs and where a human has to step in.
Launch Delivery PipelineYeachan-Heo/oh-my-claudecode · MIT Download the skill Quick start
- 1Check the requirements. Confirm the oh-my-claudecode framework and its drydock, team and ask-navigator skills are installed, Node.js is available, and the repository has CLAUDE.md and GLOSSARY.md.
- 2Install the skill. Unzip the launch skill into ~/.claude/skills/ for all projects or into the project's .claude/skills/ for one repository. In the Claude apps, upload it under Settings → Capabilities → Skills.
- 3Write the mission brief. Write the objective and scope of the change yourself. This is the first human checkpoint and the basis for everything the pipeline does next.
- 4Invoke the skill explicitly. Start the run by passing the mission brief or a path to an existing spec, using the argument format the skill documents. Watch the yard gate result first.
- 5Review the checkpoints and the report. Approve the acceptance criteria, test seams and ticket decomposition when asked, answer any batched decisions, then review the completion report and its decision log before accepting it.
What the Launch skill makes the agent do
Launch runs a delivery pipeline in phases. It starts with a yard gate that audits the project's documentation scaffolding, then resolves the document language, then moves from a mission brief to a durable spec, decomposed vertical-slice tickets with blocking edges, parallel execution of the ready tickets through the team tooling, verification, and a report with a full decision log.
The skill is built around one test applied to each step: if the step is done wrong, can the system detect it, and can it be redone or rolled back automatically? If both answers are yes, the agent does the work. If there is no unique answer, the system cannot judge the result, or a mistake would be expensive, a human decides. The skill's stated aim is to delegate as much verifiable work as possible so that human time goes only to decisions a human has to make.
- Human checkpoints: write the mission brief, approve acceptance criteria and test seams, approve ticket decomposition, answer irreversible decisions that come up mid-run, and accept the completion report.
- Agent work: repository exploration, interview preparation, spec and ticket drafting with mechanical validation, implementation at agreed seams with builds and tests, code review and verification, and the paper trail.
Install the Launch skill from a ZIP or the Claude apps
Download the ZIP for the launch skill from the directory page. The skill is a folder whose main file is SKILL.md. Unzip it into one of two locations: your user skills folder at ~/.claude/skills/ to make it available everywhere, or a project's .claude/skills/ folder to limit it to one repository.
If you use the Claude apps instead, upload the skill under Settings, then Capabilities, then Skills. Install the surrounding oh-my-claudecode tooling as well. Launch depends on its drydock audit, team and deep-interview or ask-navigator skills, so it will not work alone.
- User-wide install: unzip into ~/.claude/skills/
- Single project install: unzip into the project's .claude/skills/
- Claude apps: Settings → Capabilities → Skills, then upload
Check the requirements before your first run
Launch has four prerequisites. Missing any of them will stop the run, usually at the first gate.
The skill is meant for teams that track specs and sub-issues in an external issue tracker. The spec must be a parent issue and each ticket a native sub-issue of it. Launch does not write local spec or ticket files, so the tracker is where the work lives.
- The oh-my-claudecode framework, with its drydock audit, team, and deep-interview or ask-navigator skills
- Node.js, because the audit runs as scripts/shipyard-audit.mjs
- An issue tracker where the spec is a parent issue and tickets are native sub-issues
- A repository containing CLAUDE.md and GLOSSARY.md. The run is blocked until both exist
Start a run with a mission brief
Launch is set to be invoked explicitly by a user, not picked up automatically by the model. You start it by passing either a mission brief (the objective and scope of the change) or the path to an existing spec. The argument hint for the skill is <mission brief | path to existing spec> with an optional --serial flag.
Because of this explicit-invocation setting, a casual natural-language request such as 'please fix the login bug' is not a reliable way to start the pipeline. Write the mission brief yourself, since the human owns that checkpoint. A brief should state the objective and the scope.
- Pass a mission brief when you have a clear objective and scope.
- Pass a path to an existing spec when the spec already exists.
- Add --serial to the argument list if you want the pipeline run serially rather than in parallel, as the argument hint describes.
<mission brief | path to existing spec> [--serial]Understand the gates and when the run stops
The yard gate runs first. It executes the full drydock check, which requires the day-one files CLAUDE.md and GLOSSARY.md. The audit also reports a missing or invalid documentLanguage tag, dead paths, and project skill triggers. A high-confidence finding blocks the run. The gate lists every finding verbatim, points you to /oh-my-claudecode:drydock, and stops without producing artifacts.
The fog gate handles unclear efforts. If the destination of the work is not clear, the run is routed to /oh-my-claudecode:ask-navigator before this pipeline starts, and no partial state is left behind. Launch does not overwrite or override a blocking finding unless you give an explicit, per-invocation override that names the finding or the throwaway scope. There is no general bypass flag.
- Yard gate: the drydock audit must pass, or be overridden with explicit intent for low-confidence findings, false positives, or a declared scratch repository
- Fog gate: unclear efforts are sent to the navigator and do not start the pipeline
- There is no automatic resume. After an interruption you must start a new explicit invocation
Safety and what the skill touches
Launch runs a script, scripts/shipyard-audit.mjs, through Node.js. The skill's metadata says it runs scripts and does not use the network. Read the script and the surrounding tooling before running it in a repository you care about.
Launch-authored files are limited to GLOSSARY.md, docs/adr/, and docs/business/, plus specific sediment slots after you approve the completion report. Spec and tickets go only to the tracker. Launch does not create its own approval records, rollback mechanism or cancellation path, so track durable decisions in the tracker and repository files.
- Runs scripts: yes, node scripts/shipyard-audit.mjs
- Uses network: no, according to the skill's metadata
- Files written: GLOSSARY.md, docs/adr/, docs/business/, and approved sediment slots only
- Task status belongs to team; Launch does not change it outside team's contract
node scripts/shipyard-audit.mjs [repoRoot]What people use it for
- Splitting a multi-ticket feature into a parent spec issue with sub-issues, then letting agents work independent tickets in parallel while you approve the plan.
- Running a repeatable, checkpointed process across a team that already uses oh-my-claudecode and an external issue tracker.
- Keeping a written decision log for a larger change so reviewers can see why each decision was made.
Frequently asked questions
Is the Launch skill free to use?
The source material for this page does not state pricing or licence terms beyond the licence listed for the item, which is MIT. Check the repository's licence file and the oh-my-claudecode framework's own terms before relying on it commercially.
What do I need before I can install the launch skill?
You need the oh-my-claudecode framework with its drydock, team and deep-interview or ask-navigator skills, Node.js for the audit script, an issue tracker that supports parent issues and sub-issues, and a repository containing CLAUDE.md and GLOSSARY.md.
Why does my run stop at the yard gate?
The yard gate blocks the run when the drydock audit finds high-confidence actionable problems, such as a missing CLAUDE.md or GLOSSARY.md, a missing or invalid documentLanguage tag, or dead paths in CLAUDE.md. Fix each listed finding, then run again. Launch lists the findings verbatim and produces no artifacts.
Can I use launch for a small one-off edit?
The skill is not designed for that. The audit gate, spec and ticket decomposition add overhead that is not worth it for small one-off changes. Use a simpler workflow for those.
What happens if my task is vague or exploratory?
The fog gate routes unclear efforts to /oh-my-claudecode:ask-navigator before the pipeline starts. No artifacts are written. Launch resumes only from the mission brief the navigator hands back.
Does launch use the internet?
According to the skill's metadata, it does not use the network. It does run a local script, scripts/shipyard-audit.mjs, through Node.js.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Launch suits teams already using oh-my-claudecode who track specs and sub-issues in an issue tracker and want a repeatable, checkpointed process for multi-ticket changes. It is a poor fit for small edits, exploratory work with no clear destination, or repositories without the required CLAUDE.md and GLOSSARY.md files, where a simpler workflow or the navigator is the better choice.
Download the skill