Become a superworker. Someone who works much more productively by using AI tools that can carry out tasks and act on their behalf. — Cambridge Dictionary
Your AI agent is capable. But without the right guardrails, it guesses at architecture, skips tests, and ignores platform conventions. AI Agent Skills solves that — a battle-tested library of skills and rules that give Claude Code, Cursor, and compatible agents the discipline, judgment, and domain knowledge of a senior engineer, encoded once and applied automatically to every prompt, every file, every project.
Out of the box, AI coding agents are capable but undisciplined. They generate code that works in isolation but violates architecture boundaries, skips test coverage, or ignores platform conventions. This repo fixes that by encoding hard-won engineering standards into two complementary layers:
- Skills — active behaviors: slash commands that give agents a structured process to follow for complex tasks (debugging, multi-agent orchestration, etc.)
- Rules — passive guardrails:
.mdcfiles loaded by Cursor that enforce architecture, naming, and testing patterns on every code generation
Together they turn a capable AI into a reliable engineering partner.
Skills live in skills/ and install as Claude Code slash commands. Each skill is a SKILL.md file that defines a structured process the agent follows when invoked.
mkdir -p .claude/skills/<skill-name>
cp skills/<skill-name>/SKILL.md .claude/skills/<skill-name>/SKILL.mdProblem it solves: Agents jump to guesses. This skill forces a disciplined root-cause process before touching any code.
/debug [paste error or description]
The agent follows a fixed four-step pipeline:
- Gather — collects structured context: problem · expected · actual · code · error · env · tried
- Diagnose — identifies the root cause in one sentence, not a symptom
- Fix — shows a minimal diff; no rewrites, no style changes, no scope creep
- Verify — tells you exactly how to confirm the fix worked (command, assertion, or observable behavior)
Rules the agent must obey: never guess without labeling it a hypothesis; never propose more than one fix per response; never refactor beyond what directly fixes the bug.
Problem it solves: Complex coding tasks need more than one agent. /friday is a coordination loop that drives tasks end-to-end through independent agents with human approval gates and evidence-based quality checks.
brainstorm → plan → approve → implement → review → smoke-test → ship
The orchestrating session only coordinates — it never writes feature code itself. Each phase is delegated:
| Phase | Agent | Responsibility |
|---|---|---|
| Plan | Fable (or Opus 4.8 fallback) | Explores repo, writes a self-contained plan.md |
| Implement | Codex gpt-5.5 at xhigh |
Edits files from the plan — never commits |
| Review | Codex adversarial-review | Independent non-Claude reviewer; finds issues or approves |
| Smoke test | Opus | Proves the change actually runs with captured evidence |
| Ship | Orchestrator | Commits and pushes only after all gates pass |
Two modes:
| Mode | Trigger | When to use |
|---|---|---|
| Full loop | "friday" / "run the friday loop" |
You have a task description; want the full pipeline including brainstorm + plan |
| Execute mode | "friday this plan" / "friday <file>.md" |
You already have a plan file; skip to implement → review → ship |
Quality gates (non-negotiable):
- No implementation without explicit plan approval
- No PASS on smoke test without shown evidence
- No self-review — review must be an independent agent
- No committing orchestration artifacts
Rules live in .cursor/rules/ and are loaded automatically by Cursor IDE based on globs in their frontmatter. They are passive — they apply to every matching file without any invocation.
git clone https://github.com/nphausg/aiagent.skills.git
cp -r aiagent.skills/.cursor/rules/ your-project/.cursor/rules/Or manually copy any rule block and prepend it to your prompt.
| Rule file | Enforces |
|---|---|
android_clean_architecture.mdc |
Domain / Data / Presentation layers; all business logic in UseCase |
android_di_hilt_patterns.mdc |
@HiltViewModel, @Inject, @Module — consistent Hilt wiring |
android_general_rules.mdc |
MVVM: ViewModel + StateFlow + UiState sealed classes |
android_naming_conventions.mdc |
File, class, and function naming aligned with Kotlin idioms |
android_unit_testing_bdd.mdc |
BDD-style names (given_when_then), MockK/Mockito, runTest |
android_integration_testing.mdc |
Integration test structure and Hilt test component setup |
android_ui_testing.mdc |
Compose / Espresso UI testing patterns |
android_workmanager_best_practices.mdc |
WorkManager constraints, chaining, and retry strategies |
| Rule file | Enforces |
|---|---|
react_hooks.mdc |
Functional components · useEffect cleanup · useCallback / useMemo stability |
| Rule file | Enforces |
|---|---|
general_clean_code_principles.mdc |
Single Responsibility, DRY, lazy initialization, avoid reflection |
general_performance_tips.mdc |
Lazy loading, memory allocation, recomposition avoidance |
Coming soon — contributions welcome.
Pull requests are welcome. The highest-value area to contribute right now is iOS rules — the directory is empty and ready.
Before submitting a new rule, verify:
| Check | Criteria |
|---|---|
| Naming & location | File name is descriptive; placed in the correct category folder |
globs accuracy |
Patterns match only the intended file types |
| Clarity | Every rule is actionable and unambiguous |
| Conciseness | No redundancy; target under ~100 lines |
| Consistency | No conflicts with existing rules |
| Token efficiency | Run python scripts/analyze_token_usage.py — flag if > 1500 tokens |
| AI compliance | Run OPENAI_API_KEY=... python scripts/run_ai_prompt_tests.py |
Use the PR template in .github/pull_request_template.md.
MIT © nphausg