Digitaliko Skills

Reusable AI agent skills for Claude Code. Stack-agnostic, battle-tested on real production projects.

What are skills?

A skill is workflow orchestration + domain knowledge:

• Workflow (w-something.md) — the steps, gates, and handoffs. Tells the agent WHAT to do.
• Playbook (playbook.md) — the framework, patterns, and theory. Teaches the agent WHY.
• Templates (templates/) — starter files to copy into new projects.

Agent = Model + Harness. These skills are part of the harness.

Available Skills

Workflow Chain

/w-plan → /w-task → /w-implement → /w-review → /w-pr

Skill	Purpose	Playbook
w-task	Entry point — parse task, score complexity, create branch	Complexity scoring rationale
w-plan	Structured planning for L/XL tasks	When planning helps vs hurts
w-implement	Search-first implementation with quality gates	Common AI mistakes, self-review protocol
w-review	Independent code review via subagent	Generator/evaluator separation
w-pr	Structured PR creation	Good PR anatomy

Meta Skills

Skill	Purpose	Playbook
w-init	Bootstrap the full skill chain for a new project	Harness maturity model
w-harness	Audit and improve the harness itself	Harness engineering framework

Quick Start

Option 1: Bootstrap everything

# Clone this repo
git clone https://github.com/Digitaliko/skills.git /tmp/skills

# Copy workflow skills into your project
mkdir -p your-project/.claude/commands
for skill in w-task w-implement w-review w-pr w-harness; do
  cp /tmp/skills/skills/$skill/$skill.md your-project/.claude/commands/
done

# Copy templates
cp /tmp/skills/skills/w-harness/templates/* your-project/

# Customize CLAUDE.md and REVIEW_RULES.md for your stack

Option 2: Pick what you need

Each skill is self-contained. Copy just the .md file you want:

cp skills/w-review/w-review.md your-project/.claude/commands/w-review.md

Option 3: Use /w-init

Copy the init skill and let it bootstrap everything:

cp skills/w-init/w-init.md your-project/.claude/commands/w-init.md
# Then in Claude Code:
/w-init

Harness Engineering Framework

These skills are built on the harness engineering frameworks from Thoughtworks and OpenAI.

The 2x2 Model

              Feedforward (Guides)          Feedback (Sensors)
            ┌───────────────────────┬───────────────────────────┐
Computa-    │ Linter rules, type    │ CI pipeline, tests,       │
tional      │ checker, import bans  │ build checks, dead code   │
            ├───────────────────────┼───────────────────────────┤
Inferen-    │ CLAUDE.md, rules/,    │ AI code review, /w-review │
tial        │ REVIEW_RULES.md       │ subagent, human review    │
            └───────────────────────┴───────────────────────────┘

The Steering Loop

PR review comment appears → (repeats?) → REVIEW_RULES.md → (lintable?) → ESLint rule
                                                                              ↓
                                                            /w-harness verifies coverage

The Escalation Ladder

1. Impossible — linter error, type constraint (strongest)
2. Inadvisable — linter warning, review rule
3. Documented — CLAUDE.md, rules file
4. Implicit — hope the agent figures it out (weakest)

Always push rules UP the ladder.

Philosophy

• Impossible > inadvisable > documented. Linter enforcement beats documentation.
• Context budget matters. Always-loaded rules stay under 800 lines.
• Generator/evaluator separation. The agent that writes code must NOT review it.
• Closed-loop improvement. Every repeating review comment is a harness bug.
• Start minimal, grow from data. 10 starter rules → 20 after real PRs → linter coverage increases.

Skill Structure

skills/
└── w-something/
    ├── w-something.md     # Workflow: steps, gates, output format
    ├── playbook.md        # Domain knowledge: why, framework, theory
    └── templates/         # Starter files (optional)

References

• Thoughtworks: Harness Engineering for Coding Agent Users (April 2026)
• OpenAI: Harness Engineering (February 2026)
• Anthropic: Claude Code Best Practices
• Addy Osmani: Agent Harness Engineering

Contributing

This is open source. If you've built skills that work well across projects, PRs are welcome.

License

MIT

---

Built by Digitaliko — AI-native software development studio.