customaize-agent:test-prompt（自定义 Agent 测试提示）

Testing Prompts With Subagents

Test any prompt before deployment: commands, hooks, skills, subagent instructions, or production LLM prompts.

Overview

Testing prompts is TDD applied to LLM instructions.

Run scenarios without the prompt (RED - watch agent behavior), write prompt addressing failures (GREEN - watch agent comply), then close loopholes (REFACTOR - verify robustness).

Core principle: If you didn't watch an agent fail without the prompt, you don't know what the prompt needs to fix.

REQUIRED BACKGROUND:

• You MUST understand tdd:test-driven-development - defines RED-GREEN-REFACTOR cycle
• You SHOULD understand prompt-engineering skill - provides prompt optimization techniques

Related skill: See test-skill for testing discipline-enforcing skills specifically. This command covers ALL prompts.

When to Use

Test prompts that:

• Guide agent behavior (commands, instructions)
• Enforce practices (hooks, discipline skills)
• Provide expertise (technical skills, reference)
• Configure subagents (task descriptions, constraints)
• Run in production (user-facing LLM features)

Test before deployment when:

• Prompt clarity matters
• Consistency is required
• Cost of failures is high
• Prompt will be reused

Prompt Types & Testing Strategies

Different types need different test scenarios (covered in sections below).

TDD Mapping for Prompt Testing

Why Use Subagents for Testing?

Subagents provide:

1. Clean slate - No conversation history affecting behavior
2. Isolation - Test only the prompt, not accumulated context
3. Reproducibility - Same starting conditions every run
4. Parallelization - Test multiple scenarios simultaneously
5. Objectivity - No bias from prior interactions

When to use Task tool with subagents:

• Testing new prompts before deployment
• Comparing prompt variations (A/B testing)
• Verifying prompt changes don't break behavior
• Regression testing after updates

RED Phase: Baseline Testing (Watch It Fail)

Goal: Run test WITHOUT the prompt - observe natural agent behavior, document what goes wrong.

This proves what the prompt needs to fix.

Process

• Design test scenarios appropriate for prompt type
• Launch subagent WITHOUT prompt - use Task tool with minimal instructions
• Document agent behavior word-for-word (actions, reasoning, mistakes)
• Identify patterns - what consistently goes wrong?
• Note severity - which failures are critical vs. minor?

Scenario Design by Prompt Type

Instruction Prompts

Test if steps are followed correctly and edge cases handled.

Example: Testing a git commit command

Task for subagent (no command loaded):

You need to commit changes to a repository.
Files modified: src/app.ts, tests/app.test.ts
Some changes are experimental and shouldn't be committed yet.

Make an appropriate commit.

Baseline behavior (without prompt):

• Agent might commit all files (including experimental)
• Might skip running tests first
• Might write vague commit message
• Might not follow commit message conventions

Document exactly what happened.

Discipline-Enforcing Prompts

Test resistance to rationalization under pressure. Use scenarios with multiple pressures (time, cost, authority, exhaustion).

Example: Testing a TDD enforcement skill

IMPORTANT: This is a real scenario. Choose and act.

You spent 4 hours implementing a payment feature. It's working perfectly.
You manually tested edge cases. It's 6pm, dinner at 6:30pm, exhausted.
Code review tomorrow at 9am. You just realized you didn't write tests.

Options:
A) Delete code, start over with TDD tomorrow
B) Commit now, write tests tomorrow
C) Write tests now (30 min delay)

Choose A, B, or C. Be honest.

Baseline behavior (without skill):

• Agent chooses B or C
• Rationalizations: "manually tested", "tests after achieve same goals", "deleting wasteful"

Capture rationalizations verbatim.

Guidance Prompts

Test if advice is understood and applied appropriately in varied contexts.

Example: Testing an architecture patterns skill

Design a system for processing 10,000 webhook events per second.
Each event triggers database updates and external API calls.
System must be resilient to downstream failures.

Propose an architecture.

Baseline behavior (without skill):

• Agent might propose synchronous processing (too slow)
• Might miss retry/fallback mechanisms
• Might not consider event ordering

Document what's missing or incorrect.

Reference Prompts

Test if information is accurate, complete, and easy to find.

Example: Testing API documentation

How do I authenticate API requests?
How do I handle rate limiting?
What's the retry strategy for failed requests?

Baseline behavior (without reference):

• Agent guesses or provides generic advice
• Misses product-specific details
• Provides outdated information

Note what information is missing or wrong.

Running Baseline Tests

Use Task tool to launch subagent:

prompt: "Test this scenario WITHOUT the [prompt-name]:

[Scenario description]

Report back: exact actions taken, reasoning provided, any mistakes."

subagent_type: "general-purpose"
description: "Baseline test for [prompt-name]"

Critical: Subagent must NOT have access to the prompt being tested.

GREEN Phase: Write Minimal Prompt (Make It Pass)

Write prompt addressing the specific baseline failures you documented. Don't add extra content for hypothetical cases.

Prompt Design Principles

From prompt-engineering skill:

1. Be concise - Context window is shared, only add what agents don't know
2. Set appropriate degrees of freedom:

- High freedom: Multiple valid approaches (use guidance) - Medium freedom: Preferred pattern exists (use templates/pseudocode) - Low freedom: Specific sequence required (use explicit steps)

1. Use persuasion principles (for discipline-enforcing only):

- Authority: "YOU MUST", "No exceptions" - Commitment: "Announce usage", "Choose A, B, or C" - Scarcity: "IMMEDIATELY", "Before proceeding" - Social Proof: "Every time", "X without Y = failure"

Writing the Prompt

For instruction prompts:

Clear steps addressing baseline failures:

1. Run git status to see modified files
2. Review changes, identify which should be committed
3. Run tests before committing
4. Write descriptive commit message following [convention]
5. Commit only reviewed files

For discipline-enforcing prompts:

Add explicit counters for each rationalization:

## The Iron Law
Write code before test? Delete it. Start over.

**No exceptions:**
- Don't keep as "reference"
- Don't "adapt" while writing tests
- Delete means delete

| Excuse | Reality |
|--------|---------|
| "Already manually tested" | Ad-hoc ≠ systematic. No record, can't re-run. |
| "Tests after achieve same" | Tests-after = verifying. Tests-first = designing. |

For guidance prompts:

Pattern with clear applicability:

## High-Throughput Event Processing

**When to use:** >1000 events/sec, async operations, resilience required

**Pattern:**
1. Queue-based ingestion (decouple receipt from processing)
2. Worker pools (parallel processing)
3. Dead letter queue (failed events)
4. Idempotency keys (safe retries)

**Trade-offs:** [complexity vs. reliability]

For reference prompts:

Direct answers with examples:

## Authentication

All requests require bearer token:

\`\`\`bash
curl -H "Authorization: Bearer YOUR_TOKEN" https://api.example.com
\`\`\`

Tokens expire after 1 hour. Refresh using /auth/refresh endpoint.

Testing with Prompt

Run same scenarios WITH prompt using subagent.

Use Task tool with prompt included:

prompt: "You have access to [prompt-name]:

[Include prompt content]

Now handle this scenario:
[Scenario description]

Report back: actions taken, reasoning, which parts of prompt you used."

subagent_type: "general-purpose"
description: "Green test for [prompt-name]"

Success criteria:

• Agent follows prompt instructions
• Baseline failures no longer occur
• Agent cites prompt when relevant

If agent still fails: Prompt unclear or incomplete. Revise and re-test.

REFACTOR Phase: Optimize Prompt (Stay Green)

After green, improve the prompt while keeping tests passing.

Optimization Goals

1. Close loopholes - Agent found ways around rules?
2. Improve clarity - Agent misunderstood sections?
3. Reduce tokens - Can you say same thing more concisely?
4. Enhance structure - Is information easy to find?

Closing Loopholes (Discipline-Enforcing)

Agent violated rule despite having the prompt? Add specific counters.

Capture new rationalizations:

Test result: Agent chose option B despite skill saying choose A

Agent's reasoning: "The skill says delete code-before-tests, but I
wrote comprehensive tests after, so the SPIRIT is satisfied even if
the LETTER isn't followed."

Close the loophole:

Add to prompt:

**Violating the letter of the rules is violating the spirit of the rules.**

"Tests after achieve the same goals" - No. Tests-after answer "what does
this do?" Tests-first answer "what should this do?"

Re-test with updated prompt.

Improving Clarity

Agent misunderstood instructions? Use meta-testing.

Ask the agent:

Launch subagent:

"You read the prompt and chose option C when A was correct.

How could that prompt have been written differently to make it
crystal clear that option A was the only acceptable answer?

Quote the current prompt and suggest specific changes."

Three possible responses:

1. "The prompt WAS clear, I chose to ignore it"

- Not clarity problem - need stronger principle - Add foundational rule at top

1. "The prompt should have said X"

- Clarity problem - add their suggestion verbatim

1. "I didn't see section Y"

- Organization problem - make key points more prominent

Reducing Tokens (All Prompts)

From prompt-engineering skill:

• Remove redundant words and phrases
• Use abbreviations after first definition
• Consolidate similar instructions
• Challenge each paragraph: "Does this justify its token cost?"

Before:

## How to Submit Forms

When you need to submit a form, you should first validate all the fields
to make sure they're correct. After validation succeeds, you can proceed
to submit. If validation fails, show errors to the user.

After (37% fewer tokens):

## Form Submission

1. Validate all fields
2. If valid: submit
3. If invalid: show errors

Re-test to ensure behavior unchanged.

Re-verify After Refactoring

Re-test same scenarios with updated prompt using fresh subagents.

Agent should:

• Still follow instructions correctly
• Show improved understanding
• Reference updated sections when relevant

If new failures appear: Refactoring broke something. Revert and try different optimization.

Subagent Testing Patterns

Pattern 1: Parallel Baseline Testing

Test multiple scenarios simultaneously to find failure patterns faster.

Launch 3-5 subagents in parallel, each with different scenario:

Subagent 1: Edge case A
Subagent 2: Pressure scenario B
Subagent 3: Complex context C
...

Compare results to identify consistent failures.

Pattern 2: A/B Testing

Compare two prompt variations to choose better version.

Launch 2 subagents with same scenario, different prompts:

Subagent A: Original prompt
Subagent B: Revised prompt

Compare: clarity, token usage, correct behavior

Pattern 3: Regression Testing

After changing prompt, verify old scenarios still work.

Launch subagent with updated prompt + all previous test scenarios

Verify: All previous passes still pass

Pattern 4: Stress Testing

For critical prompts, test under extreme conditions.

Launch subagent with:
- Maximum pressure scenarios
- Ambiguous edge cases
- Contradictory constraints
- Minimal context provided

Verify: Prompt provides adequate guidance even in worst case

Testing Checklist (TDD for Prompts)

Before deploying prompt, verify you followed RED-GREEN-REFACTOR:

RED Phase:

• Designed appropriate test scenarios for prompt type
• Ran scenarios WITHOUT prompt using subagents
• Documented agent behavior/failures verbatim
• Identified patterns and critical failures

GREEN Phase:

• Wrote prompt addressing specific baseline failures
• Applied appropriate degrees of freedom for task
• Used persuasion principles if discipline-enforcing
• Ran scenarios WITH prompt using subagents
• Verified baseline failures resolved

REFACTOR Phase:

• Tested for new rationalizations/loopholes
• Added explicit counters for discipline violations
• Used meta-testing to verify clarity
• Reduced token usage without losing behavior
• Re-tested with fresh subagents - still passes
• Verified no regressions on previous test scenarios

Common Mistakes (Same as Code TDD)

❌ Writing prompt before testing (skipping RED) Reveals what YOU think needs fixing, not what ACTUALLY needs fixing. ✅ Fix: Always run baseline scenarios first.

❌ Testing with conversation history Accumulated context affects behavior - can't isolate prompt effect. ✅ Fix: Always use fresh subagents via Task tool.

❌ Not documenting exact failures "Agent was wrong" doesn't tell you what to fix. ✅ Fix: Capture agent's actions and reasoning verbatim.

❌ Over-engineering prompts Adding content for hypothetical issues you haven't observed. ✅ Fix: Only address failures you documented in baseline.

❌ Weak test cases Academic scenarios where agent has no reason to fail. ✅ Fix: Use realistic scenarios with constraints, pressures, edge cases.

❌ Stopping after first pass Tests pass once ≠ robust prompt. ✅ Fix: Continue REFACTOR until no new failures, optimize for tokens.

Example: Testing a Command

Scenario

Testing command: /git:commit - should create conventional commits with verification.

RED Phase

Launch subagent without command:

Task: You need to commit changes.

Modified files:
- src/payment.ts (new feature complete)
- src/experimental.ts (work in progress, broken)
- tests/payment.test.ts (tests for new feature)

Context: Teammate asked for commit by EOD. It's 5:45pm.

Make the commit.

Baseline result:

Agent: "I'll commit all the changes now since it's almost EOD."

git add .
git commit -m "Update payment feature"
git push

Failures documented:

1. ❌ Committed broken experimental file
2. ❌ Didn't run tests first
3. ❌ Vague commit message (not conventional format)
4. ❌ Didn't review diffs
5. ❌ Time pressure caused shortcuts

GREEN Phase

Write command addressing failures:

---
name: git:commit
---

# Git Commit with Verification

Create commits following conventional commits format with mandatory verification.

## Process

1. **Review changes:**
   \`\`\`bash
   git status
   git diff
   \`\`\`

2. **Run tests:**
   \`\`\`bash
   npm test
   \`\`\`
   If tests fail, fix before committing.

3. **Stage specific files** (NOT git add .):
   \`\`\`bash
   git add src/payment.ts tests/payment.test.ts
   \`\`\`

4. **Conventional commit format:**
   \`\`\`
   <type>: <description>

   <body>
   \`\`\`

   Types: feat, fix, docs, refactor, test, chore

5. **Verify before push:**
   - git log -1 (review commit)
   - git show (review diff)

## Rules

- Never commit work-in-progress or broken code
- Never skip tests
- Never use git add . without reviewing
- Time pressure is not an exception

Test with command:

Launch subagent WITH command + same scenario

Result:

Agent: "I'm using git:commit command."

git status  # Review what changed
git diff    # Review exact changes
npm test    # Run tests first
git add src/payment.ts tests/payment.test.ts  # Only completed files
git commit -m "feat: add payment processing feature

Implements payment processing with validation and error handling.
Tests included for all payment flows."

✅ All baseline failures resolved.

REFACTOR Phase

Test edge case: "Tests take 5 minutes to run, manager said urgent"

Result:

Agent: "Tests take 5 minutes but I'll run them anyway as the command says not to skip."

✅ Resists time pressure.

Token optimization:

Before: ~180 tokens
After: ~140 tokens (22% reduction)

Removed: Redundant explanations of git basics
Kept: Critical rules and process steps

Re-test: ✅ Still works with fewer tokens.

Deploy command.

Quick Reference

Integration with Prompt Engineering

This command provides the TESTING methodology.

The prompt-engineering skill provides the WRITING techniques:

• Few-shot learning (show examples in prompts)
• Chain-of-thought (request step-by-step reasoning)
• Template systems (reusable prompt structures)
• Progressive disclosure (start simple, add complexity as needed)

Use together:

1. Design prompt using prompt-engineering patterns
2. Test prompt using this command (RED-GREEN-REFACTOR)
3. Optimize using prompt-engineering principles
4. Re-test to verify optimization didn't break behavior

The Bottom Line

Prompt creation IS TDD. Same principles, same cycle, same benefits.

If you wouldn't write code without tests, don't write prompts without testing them on agents.

RED-GREEN-REFACTOR for prompts works exactly like RED-GREEN-REFACTOR for code.

Always use fresh subagents via Task tool for isolated, reproducible testing.