Agent Harness Setup

Agent Harness Setup

This skill walks the user through installing the six-component agent harness from the MarTech AI newsletter. It is the interactive companion to the agent-harness-starter repo.

The user runs the skill, you ask the setup questions, you write the files. Charlie's directive: ask in plain prose, not the AskUserQuestion picker (feedback_no_question_picker).

Step 1 — Ask which path

Open with:

Welcome to the agent harness setup. Two paths:

1. Global harness — your personal ~/.claude/ directory (Personalisation, Memory, Skills, Agents). Recommended first time.
2. Project harness — a CLAUDE.md for a specific project you're already inside.

Which would you like?

If global → Step 2. If project → Step 7.

Step 2 — Global setup, pre-flight

Run bash skill/scripts/setup.sh (relative to wherever the repo was cloned). This creates:

~/.claude/CLAUDE.md            (empty, populated next)
~/.claude/memory/MEMORY.md     (index template)
~/.claude/skills/              (empty)
~/.claude/agents/              (empty)

If ~/.claude/CLAUDE.md already exists, ask the user before overwriting. Default: rename existing to CLAUDE.md.backup-{timestamp} and continue.

Step 3 — Personalisation interview

Ask the user, one block at a time, in plain prose:

1. Your role and what you do. "Who are you and what do you spend your time on? One paragraph."
2. Your voice. "How do you write? British or US English? Long sentences or short? Any punctuation you ban (em dashes, semicolons)? Any words you ban (leverage, tapestry, deep dive)?"
3. Your output defaults. "What format do you want answers in by default? Markdown? Plain text? Save to file?"
4. Your workflow rules. "When should I plan before executing? When should I ask vs. just go? How do I handle corrections?"
5. The audience for your work. "Who reads what you ship? Marketers? Engineers? Founders?"

After all five, draft the global CLAUDE.md based on the template at templates/CLAUDE.md with the user's answers filled in. Show it to them, ask for edits, then save to ~/.claude/CLAUDE.md.

Step 4 — Memory bootstrap

Copy templates/memory/MEMORY.md to ~/.claude/memory/MEMORY.md if it doesn't exist. Copy templates/memory/feedback_example.md to ~/.claude/memory/feedback_example.md as a reference.

Tell the user:

Memory is where every correction becomes a permanent rule. The pattern is:

1. I make a mistake.
2. You say "remember this: [rule]".
3. I write ~/.claude/memory/feedback_[name].md and add a line to MEMORY.md.
4. Next session, the rule loads automatically.

Charlie's folder has 135 of these.

Step 5 — Skills bootstrap

Copy templates/skills/post-writer-example/ to ~/.claude/skills/post-writer-example/.

Tell the user:

Skills are saved workflows you fire by name. The post-writer-example is a starter — open it, see how a SKILL.md is structured, then save your own once a workflow runs cleanly.

Step 6 — Agents bootstrap

Copy templates/agents/{strategist,builder,qa-scorer}.md to ~/.claude/agents/.

Tell the user:

Agents are specialists you delegate to. The three templates cover the most common pipeline: a strategist that decides what to build, a builder that ships it, a qa-scorer that gates quality. Rename or duplicate them for your own workflow.

Then offer:

Want me to set up an agent team for a specific workflow you run? Describe the workflow and I'll write the role files.

Step 7 — Project setup (if user chose project path in Step 1)

Ask:

1. "What's the project?" (purpose, audience, output destination)
2. "Voice or tone for this project — does it differ from your global default?"
3. "Reference files I should pull from? (brand kit, past examples, voice files)"
4. "Output format — file? Notion page? Substack draft?"
5. "Any banned words or required phrases specific to this project?"

Draft a project-level CLAUDE.md based on templates/project-CLAUDE.md. Save to <project-root>/CLAUDE.md. Show to user, ask for edits.

Step 8 — MCP connectors

Ask:

Want suggestions for which MCP connectors to wire up? I'll list the most common ones and explain how to install each.

If yes, list:

• Gmail — read + send mail
• Notion — read + write pages
• Slack — read channels + send messages
• Google Drive — read + create files
• Linear — read + create issues
• Granola — read meeting notes

For each one the user picks, explain the install path:

Tell me in this session "connect to [tool]" and I'll walk you through the OAuth setup.

Flag Charlie's note: focus on the tools the user already lives in. Five well-chosen connectors beat twenty unused ones.

Step 9 — Wrap

Show what's installed:

Your harness is live:

• Global CLAUDE.md at ~/.claude/CLAUDE.md ✓
• Memory index at ~/.claude/memory/MEMORY.md ✓
• Skills folder at ~/.claude/skills/ (with starter example) ✓
• Agents folder at ~/.claude/agents/ (with 3 starter roles) ✓

Next:

1. Read the full discipline in the MarTech AI newsletter: [link]
2. Save every correction as a memory file from here on
3. Once a workflow runs cleanly, save it as a Skill

When you mess up, tell me, and I'll route the fix to the right place.

Non-negotiables

• Plain prose questions, no AskUserQuestion picker. Charlie's preference; respect it.
• Backup before overwrite. Never silently overwrite existing ~/.claude/ files.
• Show drafts before saving. Always let the user edit before the file lands on disk.
• British English in all generated copy unless the user specifies otherwise.