Prompt Kata

Practice writing effective prompts by attempting Electron tasks with a single prompt.

Core Principles

• •One prompt only — User gets one shot, no revisions
• •No intervention — Sub-agent runs to completion or failure
• •Outcome-based — Challenges describe WHAT, you figure out HOW
• •Pass/Fail — Either the acceptance criteria are met, or they aren't

Flow

Step 1: Reset the challenge app

Reset to baseline:

cd /Users/mikaelweiss/code/dotfiles/learning/challenges/electron-app && git checkout -- . && git clean -fd .

Step 2: Generate a challenge

Read the user's goal from /Users/mikaelweiss/code/dotfiles/learning/README.md and their history from challenges/log.md.

Generation Rules:

1. •The challenge must test a skill relevant to the user's goal (scaling autonomous agents)
2. •Focus on skills that enable agent autonomy:
   • •Clear outcome specification
   • •Architectural constraint definition
   • •Failure behavior specification
   • •Verification criteria clarity
3. •Scale difficulty based on log.md history:
   • •First 5 katas: Simple features, test clear communication
   • •After 3 PASSes: Add coordination/persistence requirements
   • •After 6 PASSes: Add ambiguity that requires architectural guidance
   • •After 9 PASSes: Multi-concern challenges requiring comprehensive prompts
4. •Never repeat a challenge from the last 3 entries in log.md
5. •The challenge must be completable in a minimal Electron app

Challenge Structure:

• •Name: Short, memorable title
• •Outcome: What the user should be able to do (not implementation details)
• •Acceptance Criteria: 2-4 observable, testable criteria
• •Difficulty Indicator: What prompt skill this tests

For harder challenges (after 6+ PASSes), also generate:

• •Why it's hard: What decisions the agent will make arbitrarily without guidance
• •What your prompt must specify: The constraints needed for success

Step 3: Present the challenge

Tell the user:

## Challenge: [Name]

[The outcome statement - what the user should be able to do]

**Acceptance Criteria:**
[List the criteria as checkboxes]

**This tests:** [What prompt skill this challenge exercises]

Write your prompt below. This exact text will be given to a fresh Claude agent.
Remember: ONE prompt. Make it count.

For harder challenges (after 6+ PASSes), also show:

⚠️ This challenge has ambiguity

**Why it's hard:** [What the agent will decide arbitrarily]

**Your prompt should specify:** [What constraints are needed]

Step 4: Receive the prompt

Wait for user to provide their prompt. Do not offer suggestions or improvements.

Step 5: Run the sub-agent

Use the Task tool to spawn an agent:

• •subagent_type: general-purpose
• •prompt: The user's EXACT prompt, prefixed with the working directory
• •Working context: /Users/mikaelweiss/code/dotfiles/learning/challenges/electron-app

Format the prompt to the sub-agent as:

Working directory: /Users/mikaelweiss/code/dotfiles/learning/challenges/electron-app

[USER'S EXACT PROMPT HERE - do not modify]

Let it run to completion.

Step 6: Review and Grade

After the sub-agent finishes:

1. •Run git diff in the challenge app to see changes
2. •Run git status to see new files
3. •Run lint and tests (bun run lint && bun run test or equivalent)
4. •Review the code to verify acceptance criteria are met

Grading:

• •Check each acceptance criterion by reading the code
• •Mark as ✓ (pass) or ✗ (fail)
• •PASS: All criteria met, lint passes, tests pass
• •FAIL: Any criterion not met, or lint/test failures

Step 7: Log the result

Append to /Users/mikaelweiss/code/dotfiles/learning/challenges/log.md:

## [DATE] - [Challenge Name]

**Prompt:**
> [Their exact prompt]

**Result:** PASS / FAIL

**Criteria:**
- [✓/✗] Criterion 1
- [✓/✗] Criterion 2
- [✓/✗] Criterion 3

**Notes:**
- [What prompt skill was demonstrated or missing]
- [What the agent did well or struggled with]

---

Step 8: Reset and wrap up

Reset the app for next time:

cd /Users/mikaelweiss/code/dotfiles/learning/challenges/electron-app && git checkout -- . && git clean -fd .

Give brief feedback and encourage the user to continue with their build session.

Anti-Patterns

• •Do NOT suggest improvements to the prompt before running it
• •Do NOT modify the user's prompt in any way
• •Do NOT intervene while the sub-agent is working
• •Do NOT give hints about what the prompt should include
• •Do NOT run the same challenge twice in a row