Victoriakaey build-reliable-agents
Overview
This skill collection captures patterns battle-tested in production LLM system development. Each sub-skill is a self-contained process for one recurring problem. They can be used independently or in combination.
When to Use Which Sub-Skill
| Situation | Sub-Skill |
|---|---|
| Have an idea but don't know where to start | ai-system-design |
| Facing a problem with multiple possible solutions | problem-exploration |
| Previous approach failed, need to reconsider | problem-exploration |
| Deciding agent architecture (single vs multi, tool-use vs nodes) | agent-architecture |
| Writing a new prompt from scratch | prompt-design (Mode 1) |
| Existing prompt producing unstable or wrong outputs | prompt-design (Mode 2) |
| Starting any implementation task | experiment-driven-development |
| About to change a prompt | prompt-change-management |
| Need to compare two system versions | regression-testing |
| Designing a Critic, Judge, or Evaluator component | critic-judge-design |
| Connecting agent to a web app, API, or third-party platform | agent-integration |
| Designing or improving agent tools, observations, and evaluation signals | harness-design |
| Agent picks wrong tool, ignores data, or loops without progress | harness-design (diagnosis ladder) |
| Designing a database schema | database-design |
| Choosing a model or deciding API vs local | model-selection |
| Designing how an agent remembers information | memory-system |
| Deploying to production or setting up CI/CD | devops |
| Reviewing code — Claude reviews or guided self-review | code-review |
| Starting a new LLM project from scratch | ai-system-design → agent-architecture → model-selection → database-design → experiment-driven-development |
| System behaving inconsistently across runs | regression-testing → prompt-design (Mode 2) |
| Critic/Judge producing unstable judgments | critic-judge-design → prompt-design (Mode 2) |
| Agent works locally but fails when integrated | agent-integration → devops |
| Agent misbehaves — wrong tools, missed data, silent failures | harness-design → experiment-driven-development |
Core Principle Across All Sub-Skills
LLM behavior is non-deterministic. Treat every change as an experiment, not a fix.
This means:
- Every change has a before/after measurement
- Regressions are expected and must be detected early
- Prompt changes have blast radius beyond the component they touch
- Deterministic components (validators, routers, formatters) are always preferable to LLM components when both can do the job
Sub-Skills
ai-system-design/SKILL.md— Go from a vague idea to a full system design, guided by conversationproblem-exploration/SKILL.md— Exhaust the solution space before committing to any approachagent-architecture/SKILL.md— Agent architecture decision frameworkprompt-design/SKILL.md— Write or diagnose prompts for any LLM componentexperiment-driven-development/SKILL.md— 10-step process from evidence to git commitprompt-change-management/SKILL.md— Safe prompt iteration with regression detectionregression-testing/SKILL.md— Controlled comparison protocol for LLM systemscritic-judge-design/SKILL.md— Designing reliable LLM-as-Judge componentsagent-integration/SKILL.md— Connect agents to full-stack apps, APIs, and third-party platformsdatabase-design/SKILL.md— Schema design for general and AI-specific systemsmodel-selection/SKILL.md— API vs local, model selection, fine-tuning decisionsmemory-system/SKILL.md— Design how agents remember and retrieve informationdevops/SKILL.md— Deploy to production, CI/CD, monitoring, rollbackharness-design/SKILL.md— Design action space, tool boundaries, observation format, and evaluation signalscode-review/SKILL.md— Claude reviews your code or guides you through reviewing