Prompt Architect

A Claude skill that helps you write better prompts — the kind that work the first time, not the fifth.

Built from Anthropic's Prompting 101 workshop (Code w/ Claude, May 2025). Same 10-block structure the Anthropic Applied AI team teaches internally.

---

What is this, in plain English?

You know how sometimes Claude gives you a great answer, and sometimes it goes off in a weird direction? The difference is almost always the prompt — what you typed in.

This skill is like a writing coach for prompts. You tell it what you want Claude to do (in regular language), and it builds you a properly-structured prompt that works.

You don't need to know anything technical. You just need to:

1. Install the skill (one-time, see Install below).
2. Ask Claude something like "build me a prompt to summarize my weekly emails into a one-page brief".
3. The skill will ask 3–4 quick questions, then hand you a polished prompt to copy-paste.

---

Why does this exist?

Most people write prompts like text messages — short, vague, and hoping for the best. The Anthropic team showed in their workshop that the same task can go from "Claude thinks this is a skiing accident" to "Claude correctly identifies who's at fault" just by structuring the prompt properly.

This skill packages that structure into something you can use without having to memorize it.

---

What it does

When you ask Claude to write or improve a prompt, this skill:

1. Asks you a few questions — what task, what input, what output, what's the worst-case failure.
2. Picks a template — extraction, classification, generation, analysis, agent, or generic.
3. Builds the prompt using the 10-block framework: role, tone, background, task steps, examples, output format, reminders, and prefill.
4. Hands you three things:
   • The prompt itself, ready to copy.
   • A short note explaining each block (so you can tweak it later).
   • A test plan — three example inputs to try, including an edge case.

---

The 10 blocks (one-liner each)

1. Task and tone — who Claude is, what success means.
2. Background — static info that doesn't change between runs.
3. Detailed steps — the procedure, in order.
4. Examples — input → output pairs, especially edge cases.
5. Conversation history — prior turns, if relevant.
6. The actual ask — the live input for this run.
7. Think step by step — explicit reasoning instruction.
8. Output format — exact shape of the answer.
9. Final reminder — re-state the most important rules at the end.
10. Prefill — start Claude's response for it, to lock the format.

Full detail with examples in references/framework.md.

---

Install

Two install paths. Most people want Path A (Cowork desktop app). Pick the one that matches what you use.

---

Path A — Claude Cowork (desktop app) · recommended

Click-only. No Terminal. Two clicks worth of attention: Sync (step 7) registers the marketplace, then a separate "+" click (step 8) actually installs the plugin. Both are needed — Sync alone won't install.

Step-by-step

1. Open the Claude desktop app and switch to the Cowork tab.

2. Get to Customize. Two ways depending on your version:

   • If you see "Customize" in the left sidebar: click it.
   • If you don't: click Settings → Connectors — you'll see a message "Connectors have moved to Customize." Click the underlined Customize.

3. In Customize, find "Personal plugins" in the left column. Click the small "+" button next to it.

4. A menu pops up. Hover over Create plugin, then click Add marketplace in the submenu.

5. A dialog opens with a URL field labeled "A GitHub owner/repo or a git repository URL."

6. Paste this URL:

   https://github.com/hosanxiv/prompt-architect
   

7. Click Sync. The dialog closes silently. The marketplace catalog is now registered with Cowork — but the plugin isn't installed yet.

8. Now find and install the plugin. Click Plugins in the Customize left sidebar (you'll see Skills / Connectors / Plugins as three options). At the top, click the Personal tab. Scroll until you see prompt-architect listed with a small "+" button next to it. Click that "+" to install.

   ⚠️ This is the step everyone misses. Sync just registers the catalog. The plugin isn't active until you click this "+". If the plugin entry isn't visible after Sync, refresh the panel or restart Claude and check again.

9. Quit Claude fully. Press Cmd+Q on Mac. Closing the window is not enough — Claude keeps running in the background and won't refresh the slash menu. Reopen Claude.

10. Verify. Open a new chat, type / in the message box. prompt-architect should appear in the dropdown.

To use it, either click prompt-architect from the dropdown, or just type a request like "build me a prompt for X" — the skill auto-triggers based on its description matching what you're asking.

What you should see at each step

After step	What's visible
7 (clicked Sync)	Dialog closes silently. Nothing else changes immediately — this is correct.
8 (clicked + on prompt-architect)	The "+" disappears or changes to a checkmark — plugin is installed.
9–10 (Cmd+Q, reopen, /)	prompt-architect appears in your slash dropdown AND under "Personal plugins" in the Customize sidebar.

If any step doesn't match — check the troubleshooting table below.

---

Path B — Claude Code (CLI) · for Terminal users

Two commands:

claude plugin marketplace add hosanxiv/prompt-architect
claude plugin install prompt-architect@the-ai-burrow

Restart your claude session. Type /prompt-architect — appears in the slash dropdown.

---

Updating later

When the plugin gets new versions, refresh:

• Cowork: Customize → Personal plugins → click the-ai-burrow → click Update (or it auto-updates on launch).
• Claude Code: claude plugin marketplace update the-ai-burrow

Uninstalling

• Cowork: Customize → Personal plugins → click prompt-architect → Remove.
• Claude Code: claude plugin uninstall prompt-architect@the-ai-burrow

Troubleshooting

Symptom	What's wrong	Fix
Synced, dialog closed, but nothing in my Personal plugins list	You stopped at Stage 1. Sync only adds the marketplace — you still need to install the plugin.	Go to Customize → Plugins (left sidebar) → Personal tab → Local uploads. Click the "+" next to prompt-architect. Then quit Claude (Cmd+Q) and reopen.
Clicked "+" but /prompt-architect doesn't show when I type /	The slash menu didn't refresh.	Quit Claude fully — Cmd+Q on Mac. Closing the window is not enough; the app keeps running in the background. Then reopen.
"Failed to add marketplace" when clicking Sync	URL or path issue.	Use the full URL https://github.com/hosanxiv/prompt-architect. Cowork doesn't accept local file paths in this dialog.
"This repository isn't a marketplace"	Cowork couldn't find .claude-plugin/marketplace.json in the repo.	Refresh the page and try again. If it persists, open an issue.
Can't find "Customize" in the left sidebar	Cowork moved Customize behind Settings.	Open Settings → Connectors. You'll see "Connectors have moved to Customize." — click the underlined Customize.
Two /prompt-architect entries in Claude Code (CLI)	Duplicate from a stray manual install.	Delete ~/.claude/skills/prompt-architect/. The plugin install (via claude plugin install) is the one to keep.

Repo structure

prompt-architect/                       ← this repo = MARKETPLACE
├── .claude-plugin/
│   └── marketplace.json                ← marketplace catalog
├── plugins/
│   └── prompt-architect/               ← THE PLUGIN
│       ├── .claude-plugin/
│       │   └── plugin.json             ← plugin manifest
│       └── skills/
│           └── prompt-architect/       ← skill name
│               ├── SKILL.md            ← skill instructions
│               ├── references/         ← deep-dive docs
│               ├── templates/          ← task-type templates
│               └── examples/           ← worked before/after
├── README.md
├── INSTALL.md
├── CHANGELOG.md
├── LICENSE
└── .gitignore

Roadmap

• Templates for RAG, evals, code-generation tasks.
• A --quick mode that drops to 2 questions for fast prompt drafts.
• More worked before/after examples beyond the insurance case.
• Localization of the README.

---

Use

The canonical trigger: /prompt-architect

Type the slash command in your Claude session:

/prompt-architect

Or describe what you need in plain language — these all trigger the skill too:

build me a prompt that turns my meeting transcripts into action items

improve this prompt: [paste your prompt]

I want Claude to read product reviews and tell me which ones are angry — give me a prompt for that

The skill activates, asks 3–4 multiple-choice questions (rendered as clickable cards in Cowork; plain Q&A in the CLI), then hands you a fully-structured prompt.

What you'll get back

Roughly this shape:

**Your prompt:**

```
SYSTEM:
<role>...</role>
<task>...</task>
...

USER:
<input>
...your real content goes here...
</input>
```

**What each block does:**
- <role> sets up who Claude is (block 1).
- <task> is the step-by-step procedure (block 3).
- ...

**Test it on these 3 inputs first:**
1. A clean example (should work perfectly).
2. An edge case (ambiguous, missing data).
3. A hostile input (junk, off-topic — Claude should refuse to guess).

Copy the prompt, paste it where you'd normally chat with Claude, and you're done.

---

What's in this repo

prompt-architect/
├── SKILL.md              ← the skill file Claude reads (the brain)
├── README.md             ← this file
├── LICENSE               ← MIT
├── INSTALL.md            ← detailed install steps
├── references/           ← deep-dive docs the skill loads when needed
│   ├── framework.md      ← all 10 blocks explained
│   ├── checklist.md      ← pre-flight checklist
│   ├── xml-patterns.md   ← XML tag conventions
│   ├── anti-patterns.md  ← common mistakes
│   └── prefill-and-thinking.md
└── templates/            ← starting points by task type
    ├── _generic.md
    ├── extraction.md
    ├── classification.md
    ├── generation.md
    ├── analysis.md
    └── agent.md

You can read the references/ and templates/ files directly if you want to learn the framework yourself — they're written as docs, not just config.

---

Contributing

PRs welcome. Especially:

• New templates for task types not covered (RAG, code-gen, eval).
• Real before/after examples.
• Translations of the README.

Style: short, evidence-based, no fluff. Match the tone of the existing files.

---

Credits

Author / maintainer Hosan Swee — Founder, The AI Burrow · LinkedIn

The AI Burrow is an applied AI collective based in Singapore — consultancy, workshop studio, and builder community. We help teams move past the hype: documenting what actually works in production, what breaks, and what to do differently. This skill is one of the open-source artifacts that comes out of that work.

Source material Workshop content adapted from Hannah Moran and Christian Ryan, Applied AI team at Anthropic — "Prompting 101", Code w/ Claude, May 2025.

This skill packages their framework into a reusable artifact. All conceptual credit to Anthropic; any errors in the repo packaging are mine.

License

MIT. See LICENSE.