Produce a small, durable instruction file for the agents that automatically receive it. This skill authors guidance; it does not implement the application or perform its QA.

A statement belongs in AGENTS.md only when it passes all three tests:

A statement failing audience goes in a task-specific skill. Detailed explanations/evidence go in documentation; execution state goes in the project tracker. A stable documentation entry point can stay in AGENTS.md; a task-by-task document/skill catalog cannot. Do not invent skill names or create additional skills without authorization.

TDD, verification, commit policy, and honest completion are standing obligations for every contributor. They may distinguish runtime from documentation work without becoming scenario runbooks. Keep those obligations; put commands, fixture recipes, particular flows and recovery procedures in skills.

These are the two canonical templates. They preserve the user's TDD and commit wording; project/client/brand details, technology and human-language choices are generalized. Do not load the other workflow unless the project has both surfaces. For mixed projects use one minimal shared root and surface-specific nested guidance only where distinct enduring rules justify it. Do not force CLI/library-only projects into graphical-app workflows; clarify that boundary.

Preserve the seven TDD steps and test-layer definitions. E2E and Manual QA use the fully working application with zero mocks, including real LLM calls, supporting services and persisted results. Unit/integration tests may isolate dependencies when useful. Before each creation/refresh of an endpoint mock, make a fresh real endpoint/database query and save the complete response verbatim to a file; derive the mock from that file, never documentation or invented objects. Use synthetic authorized test data; never commit private responses or credentials. Read the endpoint mock policy when adapting that contract. Missing access/budget blocks required real checks, not permission to fake a pass.

Aim for 200–400 words at root and 40–120 words nested, smaller when fewer invariants exist. These are review targets, not permission to weaken requirements or add filler.

Keep root to purpose/scope, agreed language policy, non-obvious universal constraints, one documentation entry point, and shared quality/commit/completion gates. Nested files need distinct folder-wide rules; do not create one for every standard source/test/resources directory.

Adapt nested copies by removing root/project sections and ancestor quality gates already in force. Retain the subtree purpose and unique invariant instructions. Do not repeat parents, explain automatic instruction loading, or link other AGENTS.md files. Do not promote folder-only rules to root.

An internal framework, hidden compatibility boundary, or nonstandard root can qualify when every agent in scope needs it. Enforced versions belong in manifests/documentation; refer to a stable compatibility policy only when necessary.

Treat as an internal draft, not the delivered result. Complete adaptation, validation and publication/reconciliation as within the same authoring task, before presenting final files or opening a PR. Commit the finished only; remove your candidate. If interrupted, report the task as incomplete.

Show files actually changed, template copied, adaptations/relocations and checks. Leave application code untouched unless authorized. Preserve unrelated work and tool-specific instructions. Create a CLAUDE.md import only when requested or established; do not duplicate guidance or overwrite tool-specific content.

Require affected instructions and documentation to be updated in the same logical change when enduring facts/policies change. Do not claim enforcement against an agent ignoring the skill: the helper proves the starting copy; final scope and meaning require review.