version: "1.0.0" name: documenting-test-suites description: 'Writes the documentation a test suite needs to be usable by someone who did not build it: run instructions, environment setup, tag glossary, ownership, fixture and data notes, and architecture decision records. Use when onboarding someone into a suite, when "how do I run these" keeps being asked, when a test architecture decision needs recording, or when a suite is inherited with no documentation.' argument-hint: 'Test suite path, its stack and runners, the CI setup, and who the readers are' user-invocable: true

Documenting Test Suites

Use this skill when a test suite is understandable only to the person who wrote it.

Test documentation has one job: get a competent stranger from a fresh clone to a passing local run, and then let them change something without breaking a convention they could not have known about. Everything that does not serve that is decoration.

The discipline that keeps it useful is knowing what not to write. A document that restates what the code already says goes stale within a sprint and then actively misleads. Document what the reader cannot find by looking: the unwritten convention, the reason behind a choice, the gotcha no config file confesses.

When to Use

• a new joiner needs to run the suite and there is no path from clone to green
• "how do I run just the API tests" is a recurring question
• a suite has been inherited and nobody knows why it is shaped the way it is
• a test architecture decision was made and will be re-argued in six months
• tags exist and their meanings live in one person's head
• fixtures and seed data have rules that only surface when you break them

Operating Principles

• Write what the environment cannot tell you. A list of npm scripts belongs in package.json; the reason two of them exist belongs in the README.
• Optimize for the first hour. The most valuable page gets a stranger to a green run. Everything else is secondary.
• Every command in the docs is copy-pasteable and has been run. An untested command is a broken promise with a trust cost.
• Decisions get a record, not a comment. Why page objects and not fixtures, why this runner, why the suite is split this way. Those get re-litigated without a record.
• Ownership is named. A suite with no owner has no maintainer, and the documentation is where that becomes visible.
• State what is not covered. A reader who assumes coverage that does not exist is worse off than one who knows the gap.

Workflow

Phase 0: Identify the readers

Different readers need different documents. Name them before writing:

A single document serving all five serves none. Split by reader, and link.

Phase 1: Write the quickstart first

The highest-value section, and the one to validate hardest.

It contains, in order:

1. Prerequisites with exact versions: Node, package manager, Docker, browsers
2. Install, one command
3. Environment setup: which file to copy, which values are needed, where non-secret values come from and who to ask for the secrets
4. Start dependencies: the app under test, the database, the mock server
5. Run, one command
6. Expected result: how many tests, roughly how long, what a green run looks like
7. First troubleshooting entries: the three failures a newcomer actually hits

Point six is skipped almost always and it is what tells a stranger whether their run went right.

Validation gate for this phase: follow your own quickstart on a clean clone, in a fresh shell, with no environment variables set, and record every place you had to know something the document did not say. That list is the actual content of the section.

Phase 2: Document the conventions

Only the ones a reader would otherwise get wrong.

• where a new test goes, by type
• naming conventions for files, suites, and test titles
• the locator or selector policy, and what is banned
• fixture and helper rules: what belongs where, what must not be shared
• data rules: seeding, cleanup, whether tests may share an account, whether parallel-safe
• assertion conventions the team has settled on
• what a pull request touching tests is expected to include

Each entry answers a question a reviewer has asked more than once. If no one has ever asked, it probably does not need writing.

Phase 3: Write the tag glossary

Tags are a control surface. When their meanings are undocumented, they decay into personal shorthand and CI filters start excluding things nobody intended.

./resources/tag-glossary-template.md gives one row per tag: what it means, who runs it, where it runs, and who may add it.

Phase 4: Record the decisions

For each significant choice, write a short record using ./resources/test-adr-template.md:

• the decision
• the context that forced it
• the alternatives that were genuinely considered
• the consequences, including the ones the team is unhappy about
• the conditions that would justify revisiting

Candidates worth a record: runner choice, page objects versus fixtures, mocking policy, parallelization strategy, environment strategy, why a suite was split, why a whole approach was abandoned.

The alternatives section is the load-bearing part. A record that lists only the chosen option gets overturned by the next person who thinks of an alternative, because nothing tells them it was already considered and rejected.

Phase 5: Document the CI reality

What someone needs when a run is red and they were not the one who broke it:

• which workflows run, when, and against what
• how the suite is sharded and how to read a merged report
• where artifacts live: traces, videos, screenshots, coverage, and how long they are kept
• which secrets exist, what they are for, and who administers them
• how to reproduce a CI failure locally, including the container command if the environments differ
• the retry policy and what a retried pass means in the report

Phase 6: State the gaps

A short, honest section:

• areas with no coverage, and whether that is deliberate
• known flaky and quarantined tests, with their expiry dates
• tests that are skipped and why
• what the suite does not attempt to verify

This is the section that stops the suite from being trusted for more than it does. It also feeds assessing-release-readiness directly.

Phase 7: Prune

Documentation rots. Before finishing, delete:

• anything the environment already states, unless the lookup is genuinely expensive
• steps for a tool the suite no longer uses
• commands that have not been run since they were written
• history that is not a decision record

Then add a line saying when the document was last verified against a real run, and by whom.

Common Failure Modes

• a README that lists every npm script and none of the reasons any of them exist
• a quickstart written from memory by the person who set the environment up two years ago
• commands that no longer work, which teach a newcomer that the docs cannot be trusted
• documenting the framework rather than this suite, which the framework's own docs already do better
• decisions recorded as "we use page objects" with no context and no alternatives
• a tag glossary that is a list of tags with no meanings
• no ownership, so nobody updates it
• no gaps section, so readers assume the suite covers more than it does

Resource Map

• ./resources/test-readme-template.md - full README structure with a worked example and the validation gate
• ./resources/test-adr-template.md - decision record format, worked examples, and the list of decisions worth recording
• ./resources/tag-glossary-template.md - tag table, naming rules, and the CI filter mapping
• ./resources/onboarding-path.md - a staged first-week path for someone joining the suite, with checkpoints

Related Skills

• designing-test-automation-architecture (planned) - when the structure being documented should be redesigned first
• creating-instructions - when a convention should be enforced by agent instructions rather than only written down
• analyzing-quality-metrics - when metric definitions need a permanent home the team can cite
• assessing-release-readiness - which consumes the coverage gaps section directly
• testing-api-contracts - when spec gaps found during contract testing need writing up
• handling-sensitive-test-data - when the data rules being documented involve personal data

Definition of Done

This skill is complete when:

• the readers are named and each has a document aimed at them
• the quickstart was followed on a clean clone and every missing assumption it exposed is now written down
• every command in the documentation has been executed as written
• the expected result of a successful run is stated, including rough count and duration
• conventions cover only the things a reader would otherwise get wrong
• every tag has a meaning, a runner, and an owner
• significant decisions have records that include the rejected alternatives
• coverage gaps, skips, and quarantines are listed honestly
• the suite has a named owner and the document has a last-verified date