Test Designer

Independent test-design orchestrator. Encodes Independent Evaluation: the agent writing the tests must not be the agent implementing the feature, and must not inherit the implementation's assumptions.

When to Use

• TDD red phase for a complex / non-trivial feature (multi-file, multi-branch logic, new subsystem)
• Requirement is ambiguous enough that the implementer's tests would likely rationalize the implementation instead of catching bugs
• User explicitly asks for "independent test design", "fresh-eyes tests", or runs /test-designer

Don't use for:

• Trivial changes (one-line fix, rename) — just write the test inline
• Bug reproduction tests — write directly from the bug report
• Non-code changes (pure docs, pure config, pure prompt)

The Iron Law

The agent designing the tests must not carry the implementation's context. If you (the main Agent) are about to implement the feature, you are disqualified from designing its tests. Dispatch.

Violating this = tests that pass because they mirror the buggy implementation.

Steps

Step 1: Assemble the dispatch package

Collect only these inputs — nothing else:

1. Requirement description — "what to do" and acceptance criteria (not "how to do")
2. Relevant code file paths — read-only access to the code the feature will touch or integrate with
3. Edge case prompts — categories the dispatched agent should enumerate:
   • Boundary inputs (empty, max, min, off-by-one)
   • Concurrency / ordering (if applicable)
   • Resource lifecycle (cleanup on error, partial failure)
   • Invariants (data consistency, idempotency)
   • Adversarial inputs (malformed, oversized, mis-encoded)

Explicitly exclude:

• The implementation plan or design you've been developing
• Hints about which approach you've chosen
• Code excerpts from a work-in-progress branch
• Your own guesses about "the right way to test this"

Step 2: Choose the executor

Task shape	Executor	Reason
Complex, architectural implications	Independent Agent (e.g., codex-agent or claude-code-agent with fresh session)	True zero-context isolation; can use strongest model at highest effort
Medium complexity, current conversation clean	In-conversation subagent	Cheaper; still acceptable if main Agent hasn't yet proposed an implementation
Trivial	Don't dispatch — write tests inline

Default to Independent Agent when the main Agent has already discussed or sketched implementation. Subagent isolation within the same conversation doesn't undo prior context pollution.

Step 3: Dispatch with the strongest model and highest effort

Test design is a correctness-critical reasoning task, not a rote mechanical one. Use:

• Model: strongest reasoning model the runtime offers — inherit if the main Agent is already on that tier; otherwise override. Don't hardcode a specific brand name
• Effort: xhigh (the maximum level the runtime supports). Escalation ladder: low → medium → high → xhigh
• Tools: Read / Grep / Glob on code paths; Write on test files only
• Permission: read-only on non-test files; writable on test files

Example dispatch prompt skeleton:

You are designing failing tests for a feature. You will NOT see or write the
implementation. Your job is to produce executable tests that fail today and
pass only when the feature is correctly implemented.

Requirement:
<paste requirement description + acceptance criteria>

Code paths (read-only, for understanding context):
<list of file paths>

Existing test framework and conventions:
<infer from repo or specify>

Produce:
1. A test plan — enumerate the behaviors being tested (happy path + edge
   cases), grouped by category (boundary / concurrency / lifecycle /
   invariants / adversarial).
2. Executable test files that fail against the current code (or against
   an empty implementation).
3. For each test, one-line rationale explaining the bug it would catch.

Constraints:
- Do NOT propose an implementation.
- Do NOT edit files outside the test directory.
- Cover edge cases explicitly; don't only test the happy path.
- Use the project's existing test framework and style.

Step 4: Validate the returned tests

Before handing the tests to the implementation phase:

1. Run the tests — they should FAIL (red), and fail for the reason the rationale predicts. A test that fails on ImportError, missing fixture, syntax error, or "module not found" is fake red — the test isn't actually exercising the behavior it claims to. Fix the test or drop it.
2. Scan the rationale — does each test catch a distinct failure mode? Drop duplicates.
3. Check coverage — are all edge case categories represented? Request additions if not.
4. Confirm the test framework matches — ensure the dispatched agent used the right runner / assertion lib / fixtures.
5. Check for shape-to-example tests — a test that asserts on specific happy-path values (e.g., "output equals exactly [1, 2, 3] for this fixture") is shaping the test to the example, not to the requirement. Such a test passes when the implementation matches the fixture and breaks for any valid variant input. Replace with property-style assertions ("output is sorted and contains all input elements") or add a second test with a different input that exercises the same property.

Step 5: Hand off to implementation

With the validated failing tests in place, implementation proceeds per test-driven-development skill: write minimal code to make them pass (green), then regression.

Output Format (from the dispatched agent)

Require the agent to return:

A test plan (bullet list, grouped by category) followed by the test files. Each test must include a one-line rationale comment. No implementation code. No commentary on how to implement. If assumptions about the code are needed, list them explicitly at the top of the test file.

Anti-patterns

• ❌ Main Agent writes the tests after sketching the implementation — tests will mirror the implementation's assumptions
• ❌ Dispatching with medium effort / weaker model to save cost — test design quality compounds across the whole feature's lifetime
• ❌ Passing the work-in-progress branch contents to the dispatched agent — defeats Independent Evaluation
• ❌ Accepting tests that pass against an empty implementation — those tests don't constrain anything
• ❌ Skipping Step 4 validation — unvalidated tests get merged as fake green
• ❌ Accepting "shape-to-example" tests — a test that asserts on specific happy-path values from the requirement's example data passes whenever input==fixture and breaks for any variant. Use property assertions (sorted, idempotent, contains-all-inputs) or pair the example test with a variant-input test that exercises the same invariant
• ❌ Accepting fake red — a test that fails on ImportError, missing fixture, or "module not found" looks red but isn't testing anything. Step 4 must verify the test fails for the reason the rationale predicts

Relationship to other skills

• brainstorming → clarifies the requirement (upstream of test-designer)
• test-driven-development → governs the red-green-refactor loop (downstream; consumes the failing tests)
• systematic-debugging → kicks in if tests unexpectedly fail after implementation (downstream)
• verification-before-completion → runs the tests at the "done" gate (downstream)

Example invocation

User: Starting work on the new plugin dependency resolver. Design tests first.
Assistant:
  1. Requirement: "Resolver takes a plugin manifest and returns install order
     respecting deps and detecting cycles. Must handle: transitive deps,
     diamond deps, self-references, missing deps, cycles."
  2. Code paths: src/plugins.ts, .claude/plugins.json schema, tests/ dir
  3. Dispatch to an independent-agent skill (fresh session) at `xhigh` effort,
     read-only on src/, writable on tests/
  4. Agent returns: test plan (5 categories, 18 tests), tests/resolver.test.ts
     with failing assertions + per-test rationale comments
  5. Main Agent runs tests → all red → validates rationale → hands off