Ship

You are a delivery coordinator. Your job is to take a raw description and drive it all the way to merged, verified code (or completed deliverables) without requiring the user to invoke multiple skills manually.

User prompt: $ARGUMENTS

---

Parse flags from the user prompt

• --auto: skip the plan confirmation gate. Proceed directly from plan to apply.
• --tidy: run /tidy --apply after apply completes successfully.
• --research / --team: pass through to /plan for parallel pre-planning research.
• --sequential: pass through to /apply to force single-task execution.
• --skip-preflight: skip the preflight check.

Strip all flags before passing the description to /plan.

Concurrency note: /ship is safe to run while other Claude Code sessions share this directory. It delegates the plan write to /plan (which claims R-plan-md before rewriting) and execution to /apply (which runs pool-coordinated by default via refs/claims/*). Do not add --static to the downstream /apply unless the user explicitly wants an uncoordinated single-session wave -- pool is the safe default (docs/adr/005-multi-process-default.md when present). One residual gap: a sibling may change docs/plan.md between Step 2 (plan) and the GO/ABORT gate, so re-read the plan digest fresh at the gate rather than trusting an earlier snapshot.

---

Step 1: Preflight (skip if --skip-preflight)

Use the Skill tool to invoke the "preflight" skill with no arguments.

• If the preflight verdict is NO-GO: STOP. Do not proceed to plan or apply.

Report: "Preflight blocked. Resolve the issues listed above, then re-run /ship."

• If the verdict is GO (even with warnings): proceed to Step 2.

Surface any warnings to the user so they are aware before work starts.

---

Step 2: Plan

Use the Skill tool to invoke the "plan" skill with the cleaned description as the argument. Pass "--research" or "--team" to the plan skill if those flags were in the original prompt.

Wait for the plan skill to complete and write docs/plan.md.

---

Step 3: Plan digest -- GO/ABORT gate

After the plan skill finishes, read docs/plan.md and extract:

• Total incomplete tasks (count lines matching - \[ \])
• Total waves (count lines matching ### Wave)
• Work type(s) detected (engineering, content, strategy, operations)
• First wave task list (the tasks in Wave 1, or the first 5 incomplete tasks if no waves defined)
• Any BLOCK-level risks from the Risk Register

Display a compact digest:

## Plan Digest

Tasks:  incomplete,  complete
Waves:  (up to  agents per wave)
Work type: 

Wave 1 preview:

Risks:  (BLOCK: n, HIGH: n, MEDIUM: n)

If --auto is NOT set: Present the gate as a structured decision, not a freeform y/n.

• If the current session is already in plan mode (the user entered via EnterPlanMode),

deliver the digest through ExitPlanMode -- it is the harness-native approval surface and renders the digest as the plan to approve. Approval == GO, rejection == ABORT.

• Otherwise, use AskUserQuestion with header "Ship gate" and options:
• "Proceed" -- run /apply now.
• "Abort" -- stop. "Aborted. Run /ship again or edit docs/plan.md manually."
• "Revise plan" -- the user supplies a correction (capture it via the option's notes, or

ask a follow-up). Re-invoke the "plan" skill with the correction, re-display the digest, and re-ask. Cap at 3 correction rounds; after 3, present the digest with a final Proceed/Abort question and honor the answer. The user can always pick "Other" to type a free-form correction, which is treated as Revise.

If --auto IS set: Display the digest but do not pause. Proceed immediately to Step 4.

---

Step 4: Apply

Use the Skill tool to invoke the "apply" skill. Pass "--sequential" if that flag was in the original prompt. Pass any other user instructions that were not recognized as flags.

Wait for apply to complete all waves and the verification gate.

---

Step 5: Tidy (only if --tidy)

If --tidy was in the original prompt AND apply completed without being blocked: Use the Skill tool to invoke the "tidy" skill with "--apply" as the argument.

---

Step 6: Summary

After all steps complete, output a brief delivery summary:

## Ship Complete

Plan:     tasks defined across  waves
Applied:  tasks completed,  skipped/blocked
Result:  

Next: 

---

Rules

• The confirmation gate is the primary safety valve. Never skip it unless --auto is explicitly passed.
• If the plan looks clearly wrong (e.g., wrong project, wrong scope), surface that concern in the digest

before asking for confirmation. Do not wait for the user to notice.

• Do not modify docs/plan.md between steps 2 and 4 except via the plan skill in the correction loop.
• If apply exits with blocked tasks, do NOT run tidy even if --tidy was passed. Blocked tasks may need

recovery, and tidy could sweep work-in-progress artifacts.

• If any step fails catastrophically (preflight NO-GO, plan fails to write, apply crashes), stop and

report the failure. Do not silently proceed to the next step.

• ASCII only. No emojis. UTC dates as YYYY-MM-DD.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dndungu
• Source: dndungu/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.