Atoms

Full fleet snapshot of all *-atoms.com catalog sites — deploy health, DNS, HTTPS reachability, atom_types per catalog with item counts, schema compliance, /ai/index.json, Terraform backend, sidenav, brand asset uniformity, builder script, federation field, and registry presence.

Explore user intent, requirements, and design space before writing any code. Surfaces constraints, alternatives, and tradeoffs. Must run before any creative work — features, components, new functionality, or behavior modifications. Prevents building the wrong thing.

Capture a complete, resumable snapshot of in-flight work to ~/.ai/checkpoints/<owner>__<repo>/CHECKPOINT.md. Records branch, HEAD SHA, dirty files, active goal, decisions made, files in flight, last green state, next step, open questions, and a verify-on-resume list. Use before /clear, at context limits, or at phase boundaries.

Review a code diff for correctness, security, and maintainability. Surfaces findings at HIGH / MEDIUM / LOW severity with file:line citations. Follows the grounding requirement from Code.md §11.5: every medium-or-higher finding includes evidence from the actual diff, not from inference.

Five-phase systematic debugging: reproduce → isolate → root-cause → fix-with-regression-test → verify. Follows Code.md §11.4. Never proposes a fix before root cause is confirmed. Fixes address the root cause, not the surface symptom. Every fix begins with a failing test that reproduces the bug.

Decompose a multi-task problem into independent units and dispatch subagents in parallel. Use when facing 2+ independent tasks with no shared state or sequential dependencies. Each agent gets a self-contained prompt with full context.

Guides completion of a development branch: verify all tests pass, run the adversarial self-pass, open PR, wait for CI green, merge. Presents structured options for merge, PR, or cleanup. Use when implementation is complete.

Create distinctive, production-grade frontend interfaces with high design quality. Generates creative, polished code that avoids generic AI aesthetics. Use when building web components, pages, or applications.

Purge merged branches, worktrees, and stashes across the repo and all submodules. Aggregate surviving in-flight work into a single dated branch. Hand off to make build. Operates on primary repo and all git submodules with GitHub remotes.

Full delivery cycle: recover from checkpoint, ingest the issue queue, partition by domain, spawn Tech Lead agents per batch, run TDD Writer → Coder → Adversarial Tester → PR → merge pipeline for each issue. Use when starting a sprint, resuming in-flight work, or executing a multi-issue delivery run.

Discover and run the project's test suite. Report failures with file:line citations. Supports Go (go test), Node (jest/vitest/mocha), Python (pytest), and Astro (astro check). Quarantine flaky tests identified on repeated runs.

Process incoming code review feedback with technical rigor. Verify each finding against the actual code before agreeing. Distinguish valid findings from misreadings. Implement only confirmed findings. Do not perform blind, sycophantic agreement — that's as harmful as ignoring review.

Prepare and request a code review after completing implementation. Summarizes what changed, why, how to verify, what assumptions were made, and what remains. Run before merging any non-trivial change.

Create new skills, modify and improve existing skills, and measure skill performance. Guides through the full skill lifecycle: capture intent → research → write SKILL.md → test with evals → iterate. Use when creating a skill from scratch, editing an existing skill, running evals, or optimizing a skill's triggering description.

Execute implementation plans by spawning one subagent per independent task. Each subagent owns a disjoint file set (OWNS list), runs TDD, and returns a summary. Run code review between batches. Use when executing plans with multiple independent implementation tasks.

Five-phase debugging protocol: reproduce → isolate → root cause → fix with regression test → verify. Use when encountering any bug, test failure, or unexpected behavior. Never proposes a fix before root cause is confirmed. Enforces the 3-cycle local cap and 5-cycle total cap to prevent thrashing.

Write a failing test before writing implementation code. The test must be red before any production code is written. Implementation makes the test green with the minimal change. Refactor only with tests green. Apply when implementing any feature or bugfix.

Ensure an isolated workspace exists for feature work via git worktrees. Creates a worktree at the canonical location before any implementation. Single-repo work: <repo>/.worktrees/<name>/. Cross-repo or persistent: ~/.ai/worktrees/<name>/. Use at the start of any non-trivial feature or before executing implementation plans.

Run verification commands and confirm output before claiming any work is complete, fixed, or passing. Evidence before assertions always. Use before committing, opening PRs, or reporting success. Never cite your own reasoning as proof — run the command and cite the output.

Produce a Conventional Commits message from a staged diff or change description. Enforces Code.md §11.2: correct type prefix, 72-char subject line, imperative mood, body explaining why not what, one logical change per commit.

Produce a structured implementation plan before touching code on any medium or non-trivial task. Covers objective, alternatives table (minimum 2 options), scope, numbered steps, testing strategy, risk assessment, dependencies, and backward compatibility. Required for 2+ file changes.

Write or edit SKILL.md files for Claude Code skills. Apply the three-level loading principle: metadata (name + description) always in context, SKILL.md body loaded on trigger, bundled resources on demand. Use when creating new skills, editing existing ones, or verifying skills work before deployment.