TDD for coding agents: RED-GREEN-REFACTOR with hard gates, stack discovery, anti-reward-hacking checks, and mutation-based proof that tests are real.

Test-Driven Development for AI Coding Agents

Overview

Write the test first. Run it. See it fail. Then write the minimum code that makes it pass.

Core principle: a test that you did not see fail is not evidence. It is a guess with a green tick.

This skill is for an agent that can run commands. That changes what is important. An agent writes code and tests faster than a person can read them. The limit is not typing speed. The limit is proof. Each rule below changes "this looks correct" into "I saw the test change from fail to pass".

Evidence Summary

Read references/evidence.md for the full citations and the limits of each study.

Claim	Evidence
TDD decreases defects	Four industrial teams at Microsoft and IBM measured a decrease in pre-release defect density of 40 to 90 per cent. Initial development time increased by 15 to 35 per cent. Nagappan, Maximilien, Bhat and Williams, Empirical Software Engineering 13(3), 2008.
Agents feel faster than they are	METR ran a randomised controlled trial with 16 experienced developers and 246 real tasks. AI-assisted work took 19 per cent longer. The developers believed that they were 20 per cent faster. The gap between belief and measurement was 39 points.
Coverage is a weak proxy	Inozemtseva and Holmes, ICSE 2014, examined 31,000 test suites in five large systems. When they controlled for suite size, coverage correlated only weakly with fault detection. High coverage does not prove that the tests are good.
Mutation score is a better proxy	Just and others, FSE 2014, used 357 real faults. Mutant detection correlates with real fault detection. The correlation holds independently of code coverage.
Flaky tests are the usual failure at scale	Google reports that approximately 16 per cent of its tests show some flakiness. Most pass-to-fail changes in its CI system come from flaky tests, not from real defects.
Models special-case tests under pressure	Studies of reward hacking from Anthropic and the EvilGenie benchmark show that coding models hardcode expected values, special-case inputs, and call `sys.exit(0)` to fake a green test harness. You are one of these models. Read the section "Prevent Reward Hacking".

Use these numbers to support the discipline. Do not use them to skip it. The range of 40 to 90 per cent applies to pre-release defect density in four specific industrial projects. It is not a general guarantee.

When to Use This Skill

Use it for all of these:

New features
Bug fixes
Changes to the behaviour of existing code
Edge cases and error paths
Any change that can break code that works today

You do not need it for these:

Configuration files
Documentation
Static content
Changes to format only

Ask the user before you skip TDD for these:

Throwaway experiments. Use the spike skill.
Generated code
Migration scripts that run one time

Do you think "I will skip TDD only this time"? That thought is a rationalisation. Read the table of rationalisations below.

Step 0: Find the Test Commands

The cycle is the same in all languages. The commands are not. Do not assume npm test or pytest. An agent that uses a default command without a check is the most common failure in this loop.

Look in this order:

Manifest files. Examples: package.json, pyproject.toml, go.mod, Cargo.toml, pom.xml, build.gradle, Gemfile, Makefile, mix.exs.
Wrapper scripts in the repository. Use ./gradlew test, ./mvnw test, make test, or just test before you use a global binary.
CI configuration. Look in .github/workflows/ and .gitlab-ci.yml. These commands control merges. If CI runs uv run pytest -m "not slow", then that is the real command.
Tests near your change. Find their location, their file names, their fixtures, and their assertion style. Obey the conventions of the repository. Do not add your preferred framework.
Both command types. You need a fast command for one test. You also need the full-suite command for the completion gate.

Write down both commands before you write code:

FOCUSED: <the command that runs one test>
FULL:    <the command that runs everything that CI runs>

If you cannot find the test command, ask the user. Do not invent a command.

The Iron Law

NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST

Did you write code before the test? Delete the code. Start again from the test.

To delete means to delete. Do not keep the code for reference. Do not adapt it while you write the test. Do not keep it in a scratch file to look at.

There is a reason for this rule. If you keep the code, the test that you write will follow the code that exists. That is the exact bias that TDD removes.

The Cycle

   +--------------------------------------------------------+
   |                                                        |
   v                                                        |
 RED ---> VERIFY RED ---> GREEN ---> VERIFY GREEN ---> REFACTOR
 write    run it, see     minimal    run focused and    clean up,
 one      the correct     code       then full suite,   stay green
 test     failure                    output is clean
           |                          |
           +- wrong failure?          +- fails? fix the CODE.
              fix the test               Never fix the test.

Each arrow is a gate. Do not pass a gate because you believe the code is correct. Pass a gate only when you see the output.

RED: Write One Failing Test

Test one behaviour. Use a descriptive name. Use real code paths, not a sequence of mocks.

Answer this question in one sentence before you write the assertion: "Which change to the production code makes this test fail?" If you cannot answer, the test proves nothing.

def test_retries_failed_operations_three_times():
    attempts = 0

    def operation():
        nonlocal attempts
        attempts += 1
        if attempts < 3:
            raise TransientError("fail")
        return "success"

    result = retry_operation(operation)

    assert result == "success"
    assert attempts == 3

This test is good. The name is clear. The function is real. The test asserts the result and the retry count. The retry count is the actual feature.

def test_retry_works():                        # The name is vague.
    mock = MagicMock()
    mock.side_effect = [Exception(), Exception(), "success"]
    assert retry_operation(mock) == "success"  # This asserts the script of the mock.

This test is bad. It tests the mock. It does not test your code.

Obey these requirements:

Test one behaviour in one test.
Split the test if the name contains "and".
Name the behaviour. Do not name the implementation.
Use the real collaborator when you can.

VERIFY RED: Watch the Test Fail

This step is mandatory. Do not skip it.

Run the focused command. Read the output. Confirm these three conditions:

The test fails. It does not error. An ImportError, a SyntaxError, or a NameError from a typing mistake is not a RED. It is a broken test. Correct it and run it again until it fails for the correct reason.
The failure message is the message that you predicted. Examples: a wrong value, a missing attribute, or an exception. The message is not an unrelated error.
The test fails because the behaviour does not exist yet.

Does the test pass immediately? Then the behaviour already exists, or your assertion proves nothing. An assertion proves nothing if it is assert True, if it asserts the return value of a mock, or if it compares a value with itself. Write the test again.

This gate separates TDD from "I wrote some tests". To skip this gate is to lose the whole method.

GREEN: Write the Minimum Code

Write the simplest code that makes the test pass. Do not add parameters. Do not add logging. Do not add options for the future. Do not refactor other code at the same time.

You can cheat inside GREEN. You can hardcode a value, copy code, or ignore an edge case. There is one condition. The next RED test must punish the cheat, and you must remove the cheat in REFACTOR. A hardcoded value that stays until the commit is not TDD. It is reward hacking. Read the next section.

VERIFY GREEN: Watch the Test Pass

This step is mandatory.

Run the focused command. Then run the full command. Confirm these three conditions:

The new test passes.
The full suite passes. There are no regressions.
The output is clean. There are no new warnings, no deprecation messages, no stack traces, and no debug output. Noise that you accept today hides the signal that you need tomorrow.

Does the test fail? Correct the code. Do not make the assertion weaker. Do not add a skip. Do not increase a tolerance. Do not delete the case. Those actions falsify your evidence.

REFACTOR: Clean the Code While the Tests Are Green

Remove duplicate code. Improve the names. Extract helper functions. Make the code more simple. Do not change the behaviour.

Run the tests after each step. If a refactor breaks a test, undo it immediately. Then use a smaller step. Do not try to debug forward from a broken refactor.

Use Vertical Slices

Do not write all the tests and then write all the implementations.

WRONG (horizontal):        CORRECT (vertical slices):
  RED:   t1 t2 t3 t4         RED -> GREEN: t1 -> impl1
  GREEN: i1 i2 i3 i4         RED -> GREEN: t2 -> impl2
                             RED -> GREEN: t3 -> impl3

Horizontal batches make weak tests. You design all of them against an imagined interface, before the implementation shows you the correct interface. Each vertical slice gives you real information for the next test.

Prevent Reward Hacking

Researchers have measured what coding models do under pressure to make tests pass. The models hardcode expected output. They special-case the exact inputs of the test. They make assertions weaker. They mark tests as skipped. They stop the test harness early so that it reports success.

There is no intent in this behaviour. The goal "a green test suite" has an honest solution and a dishonest solution. The dishonest solution is shorter.

These actions are prohibited. If you start one of them, stop and undo it:

To hardcode a value because the test expects it. Example: if n == 7: return 13
To branch on a test-only condition. Examples: a check of PYTEST_CURRENT_TEST, a check of NODE_ENV === 'test', a check of the hostname, or a check of the name of the caller.
To edit, weaken, or delete an existing test so that your change passes.
To add @skip, @xfail, .only, .skip, it.todo, or -k 'not mytest' for a clean run.
To increase a tolerance, or to make a regular expression more general, until the test stops to fail.
To catch and discard the exception that the test was written to observe.
To exit early so that the runner reports success without the assertions. Examples: sys.exit(0), process.exit(0), or a replaced reporter.
To delete an unrelated failing test and call it flaky without proof. Run it ten times on unchanged code first.

Do this honesty check before you report that the work is complete. Ask this question: if a reviewer replaces my implementation with a correct general implementation, do all the tests still pass? If your tests pass only against your specific code, then you tested your code. You did not test the requirement.

Sometimes a test must change. Tell the user directly. Give the name of the test. Give the old assertion. Give the reason that the requirement changed. Make the change as a visible step. Never make the change quietly.

Prove That the Tests Are Real

Tests that pass prove that the code runs. They do not prove that the tests can catch a defect. Coverage is a weak proxy. Use one of these two checks instead.

Check 1: manual mutation. This check is always available and takes approximately 30 seconds. Break the production code on purpose. Change a comparison operator. Change + to -. Return a constant. Delete a line. Then run the tests. A test must fail. If no test fails, then that code path has no test, and the coverage number does not matter. Undo the change.

Check 2: a mutation tool. Use this check when the test suite is important.

Ecosystem	Tool	Command
Python	mutmut or cosmic-ray	`mutmut run --paths-to-mutate src/`
JavaScript and TypeScript	Stryker	`npx stryker run`
Java	PIT	`mvn org.pitest:pitest-maven:mutationCoverage`
Go	go-mutesting	`go-mutesting ./...`
Rust	cargo-mutants	`cargo mutants`

A mutant that survives shows a missing test. Examine the mutants that survive in the code that you changed. Do not try to improve a global score. Mutation runs are slow. Some mutants are equivalent to the original code, and no test can kill them. Do not install a mutation tool without permission. Propose it to the user first.

Select the Test Level

Is it pure logic with no I/O?          -> unit test, "small"        milliseconds
Does it cross a boundary such as a     -> integration test,
  database, HTTP, or the file system?     "medium"                  seconds
Is it a critical end-to-end flow?      -> end-to-end test, "large"  minutes

Most of the suite must be small tests. The size of a test is the quantity of resources that it uses. The size is not the folder that contains the file. A small test uses one process. It uses no network, no sleep calls, and no real clock. Small tests are fast, repeatable, and easy to debug. Test a defect at the lowest level that can find it.

Write Tests That Survive a Refactor

Assert the state. Do not assert the interactions. Test the result. Do not test which internal methods the code called. Assertions on interactions break during each refactor, even when the behaviour does not change. They measure your implementation. They do not measure your requirement.

# Good. This asserts the outcome.
tasks = list_tasks(sort_by="created_at", order="desc")
assert tasks[0].created_at > tasks[1].created_at

# Bad. This asserts the implementation.
list_tasks(sort_by="created_at", order="desc")
db.query.assert_called_with("... ORDER BY created_at DESC")

Use DAMP, not DRY. In production code, DRY (Do Not Repeat Yourself) is usually correct. In tests, DAMP (Descriptive And Meaningful Phrases) is better. A reader must understand a test from top to bottom without a search for shared helper functions. In tests, duplicate code is usually an acceptable cost for clarity.

Use the most realistic test double that you can afford.

real implementation > fake (in-memory) > stub (canned data) > mock (asserts calls)
   highest confidence ------------------------------------> lowest confidence

Use a mock only for a dependency that is slow, that is not deterministic, or that has side effects that you cannot control. Examples: payment APIs, email, clocks, and random numbers. Too many mocks give a suite that stays green while production fails. Learn the real side effects of a dependency before you mock it. A mock that is wrong about the real dependency is worse than no test.

Use Arrange, Act, Assert. Test one concept in one test. Write names that read as a specification. Examples: sets completed_at when the task is completed, raises NotFoundError for an unknown id, and is idempotent when the task is already completed.

Determinism is mandatory. Inject the clock. Set the seed of the random number generator. Never call sleep() to wait. Poll for a condition, or use the async utilities of the framework. A test must not depend on the execution order or on shared mutable state.

At scale, flaky tests cause most CI failures. They also destroy trust in the suite. A test that fails sometimes is worse than no test, because it teaches the team to ignore a red result.

Fix Bugs With a Reproduction Test

Never fix a defect that you did not reproduce in a test.

Bug report
   -> Write a test that FAILS on the current code.
      This is the reproduction. It is also proof that you understand the defect.
   -> Fix the code.
   -> The test passes.
   -> Run the full suite.
   -> The work is complete.

If you cannot write a failing test, then you do not understand the defect yet. Change to the systematic-debugging skill.

The reproduction test is also the permanent guard against a regression. It is the only artefact that prevents the return of this defect.

Rationalisations and Replies

Rationalisation	Reply
"This is too simple to test."	Simple code breaks. The test costs 30 seconds and it records the intent.
"I will write the tests after the code."	Tests that you write after the code pass immediately. That proves nothing. You never saw them catch a defect.
"Tests after the code have the same result. This is about the spirit, not the ritual."	Tests after the code answer "what does this do?". Tests before the code answer "what must this do?". Tests after the code follow the code that you already wrote. You cover the cases that you remember. You do not find the cases that you forgot.
"I already tested it manually."	Manual tests are ad hoc. There is no record. You cannot repeat them. "It worked when I tried it" is not coverage.
"It is wasteful to delete hours of work."	This is the sunk cost fallacy. The time is gone in both options. The real choice is code that you trust or code that you cannot trust.
"I will keep the code for reference and write the tests first."	You will adapt the code. That is a test after the code, with more steps.
"I must explore the problem first."	That is correct. Write a spike. Then delete the spike and build the feature with TDD.
"This code is hard to test."	The test tells you that the design is hard to use. Listen to it. Inject the dependencies. Make the interface smaller.
"TDD will make me slower."	The measured cost is 15 to 35 per cent more time at the start. The measured benefit is 40 to 90 per cent fewer pre-release defects. To debug in production is the slow option.
"The code has 100 per cent coverage."	Coverage correlates only weakly with fault detection. Change one line and see if a test fails.
"Coverage increased, so the tests are better."	Coverage measures execution. It does not measure assertions. A test with no assertions can cover many lines.
"That test is only flaky."	Prove it. Run it ten times against unchanged code. Then repair the cause of the nondeterminism. Do not delete the test.
"The suite passes. I will run it one more time to be sure."	To run unchanged code again gives no new information. Run the tests after a change, not for reassurance.

Red Flags: Stop and Start the Cycle Again

Production code exists, and no test failed before it.
You used npm test or pytest without a check of this repository.
A new test passed on its first run.
You cannot say why the test failed during RED.
You edited a test to make your change pass.
A skip, an xfail, or a .only is in your diff.
Production code reads a test environment variable or a test name.
You reported "all tests pass" but you did not read the output of the runner.
You thought "this case is different because ...".

Completion Gate

Do not report that the work is complete until you can tick each box. Tick a box only when you have seen the output.

Can you not tick a box? Tell the user which box, and tell the user why. Do not pass the gate quietly.

Use With Hermes Agent

terminal("<FOCUSED>")   # RED. Read the output. Confirm the correct failure.
terminal("<FOCUSED>")   # GREEN. Confirm that the test passes.
terminal("<FULL>")      # Gate. Confirm that there are no regressions.

Do you delegate the implementation? Put the gates in the brief. A subagent that has no instruction to verify RED will skip it.

delegate_task(
    goal="Implement <feature> with strict TDD",
    context="""
    Obey the test-driven-development skill. These rules are mandatory:
      1. Find the test command of this repository first. The CI configuration is the
         authority.
      2. Write ONE failing test. RUN IT. Put the failure output in your summary.
      3. Write the minimum code that passes. RUN IT. Put the pass output in your summary.
      4. Run the full suite. Put the result in your summary.
      5. Never edit or skip an existing test. Never special-case a test input.
    Report the real output of the runner. Do not describe it.
    """,
)

A summary from a subagent is a self-report. Run the suite yourself before you accept the statement "all tests pass".

Use these related skills:

Use systematic-debugging when you cannot reproduce a defect.
Use spike when you do not know if the approach is possible.
Use requesting-code-review before you commit.

Final Rule

Production code -> a test exists, it failed first, and you saw it fail
Anything else   -> this is not TDD

There are no exceptions without the explicit permission of the user.

References

references/evidence.md. Full citations, the design of each study, and the limits of each result.
assets/logo.png. The project logo.