Claude Code Skills

A collection of Claude Code skills for structured software delivery, built around the Ship pipeline.

Quick Reference

Skill	Description	Install
/ship	Pipeline dispatcher — runs the next stage from the manifest	npx skills@latest add dfollent/skills/ship
/ship-generate-questions	Repo scoping + intent-neutral research questions (stage 1)	npx skills@latest add dfollent/skills/ship-generate-questions
/ship-research-codebase	Objective codebase research from questions (stage 2)	npx skills@latest add dfollent/skills/ship-research-codebase
/ship-solution-design	Solution design with grill-me walk + antagonistic review (stage 3)	npx skills@latest add dfollent/skills/ship-solution-design
/ship-work-breakdown	High-level phase breakdown for human review (stage 4)	npx skills@latest add dfollent/skills/ship-work-breakdown
/ship-create-plan	Detailed tactical plan, hardened by antagonistic review (stage 5)	npx skills@latest add dfollent/skills/ship-create-plan
/ship-implement-plan	Phase-by-phase plan execution with verification (stage 6)	npx skills@latest add dfollent/skills/ship-implement-plan
/grill-me	Stress-test a plan or design	npx skills@latest add dfollent/skills/grill-me
/project-orient	Load project context from Confluence	npx skills@latest add dfollent/skills/project-orient
/project-update	Write findings back to Confluence	npx skills@latest add dfollent/skills/project-update
/project-lint	Audit Confluence doc tree for quality	npx skills@latest add dfollent/skills/project-lint

Install all skills at once:

npx skills@latest add dfollent/skills

Required: subagent + shared assets

The Ship skills depend on two things the skills installer does not copy (they have no SKILL.md). Install them manually:

# 1. The antagonistic-reviewer subagent (used by stages 3 and 5).
#    Subagents live in agents/, NOT inside skill folders — Claude Code discovers
#    them from ~/.claude/agents/ (personal) or .claude/agents/ (project).
cp agents/antagonistic-reviewer.md ~/.claude/agents/

# 2. The shared manifest/coverage references every ship skill reads by path.
cp -R _ship-shared ~/.claude/skills/_ship-shared

Skills and subagents are separate mechanisms: a skill cannot bundle a subagent. The ship skills invoke antagonistic-reviewer by name (subagent_type: antagonistic-reviewer), which only resolves if the agent file is installed under an agents/ directory. The ship skills reference the shared files at ~/.claude/skills/_ship-shared/, so that folder must live there.

The Ship pipeline

Skills that break delivery into focused stages. Each stage runs in its own context window, produces a static markdown artefact in a per-ticket <TICKET>-ship/ folder, and feeds into the next. Human review concentrates where it has the most leverage.

The /ship dispatcher is an on-rails orchestrator: it resolves the active per-ticket manifest and runs exactly one stage per invocation. Every ship- skill is also independently runnable on its own.

The flow

1. /ship-generate-questions — Scopes which repos in a multi-repo workspace the ticket touches (it legitimately sees the ticket), then writes one intent-neutral research-questions file per repo plus the per-ticket manifest.

2. /ship-research-codebase — Takes only the questions (never sees the ticket). Researches each repo in-thread, or fans out a Workflow (one agent per question theme). Outputs an objective research-<repo>.md per repo. No opinions.

3. /ship-solution-design — Takes research + ticket. Produces a solution design covering the coverage map, driven by an interactive grill-me decision-tree walk and hardened by tiered antagonistic review. First review point.

4. /ship-work-breakdown — Takes the design. Produces a high-level phase breakdown: order of changes, what each phase delivers, testing strategy. Second review point.

5. /ship-create-plan — Takes work breakdown + design + research. Produces a detailed tactical plan the agent executes, hardened by a design-lens antagonistic-reviewer pass. Skim, no full review needed.

6. /ship-implement-plan — Execute the plan, phase by phase, with verification and a pause for manual checks between phases.

Why it works

• You review the design and breakdown, not 1000 lines of plan or code. That's where leverage is highest.
• Each skill stays under 40 instructions. Smaller prompts = fewer skipped steps.
• Research is unbiased. The questions stage sees the intent; the research stage doesn't. Prevents cherry-picking facts that support what the model already wants to build.
• Every stage produces a standalone markdown doc in the per-ticket folder. Resume from any point in a fresh session.
• Antagonistic review stress-tests the design and plan through adversarial single-lens passes before any code is written.

Skipping steps

Not every task needs all six stages. Use your judgement:

• Small bug fix? Skip straight to implementation.
• Well-understood change? Start at /ship-solution-design or /ship-create-plan.
• Complex feature in unfamiliar code? Run the full flow via /ship.

Project Skills (Confluence)

Skills for project work in Confluence. One reads from the project surface, the others write back to it or audit it.

/project-orient

Load context from Confluence without a meeting or handover.

• "What projects do we have?" — lists all projects in a space
• "Orient me to project X" — project tree, open questions, blockers, unvalidated assumptions, recent decisions
• "I want to work on the API task" — loads the target doc + parent + project root, surfaces anything that might affect the work

/project-update

Write findings back so the project surface stays current.

• Appends to existing docs or creates child pages, with provenance (who, when, which tool)
• Inline tags for open questions, assumptions, blockers, decisions, actions
• Tags get resolved in-place: [ASSUMPTION -> CONFIRMED], [BLOCKER -> CLEARED]
• Captures implementation write-backs: PR links, branch names, learnings, rejected assumptions

/project-lint

Audits a project's Confluence document tree for quality, consistency, and structure. Finds stale items, contradictions, orphan pages, and missing tags.

Why this matters

• No one needs a handover meeting to pick up work. /project-orient self-serves into any project.
• No one writes status reports. /project-update captures state as a side effect of doing the work.
• AI sessions can scan the tags programmatically. The project surface is machine- and human-readable.

Utility Skills

/grill-me

Stress-test a plan or design. Claude interviews you relentlessly about every aspect, walking down each branch of the decision tree and resolving dependencies one by one. Best used after /ship-solution-design or /ship-work-breakdown.

Attribution: Matt Pocock (MIT License).

Acknowledgements

The coding workflow is informed by Dex Horthy's CRISPY methodology and the evolution from Research-Plan-Implement (RPI). Several skills are derived from or inspired by HumanLayer's Claude Code commands (Apache 2.0). See NOTICE for full attribution.