build-reliable-agents

A collection of skills for building reliable LLM agents throughout their development lifecycle.

build-reliable-agents

An engineering operating system for building reliable LLM agents.

15 skills for Claude Code covering the full agent development lifecycle. Each skill was extracted from building a production LangGraph agent with a Critic-driven retrieval loop. Framework-agnostic — the principles apply to any agent architecture.


Why This Exists

These are real failure modes that led to the skills in this collection.

Critic produced confident but wrong judgments. Input was organized per sub-question. When the answer appeared in a different sub-question's results, the Critic concluded "not found" — the input structure primed per-section reasoning. Fix was restructuring to a flat evidence pool, not changing the prompt. → critic-judge-design

Prompt change improved one case, regressed three others. No baseline existed. No regression batch ran before committing. The process failure wasn't the prompt edit — it was the lack of a safety net. → prompt-change-management, regression-testing

Agent looped because tool observations were too lossy. Search returned 50 raw rows. The LLM couldn't find the signal, so it retried with rephrased queries — indefinitely. The tool's output format made the correct next action unrecoverable. → harness-design

Verdict-before-reasoning caused post-hoc justification. The output schema put the boolean verdict before the reasoning field. LLMs generate left-to-right — the model committed to a verdict first, then rationalized. Moving one field fixed it. → critic-judge-design


What's Included

Skills activate automatically based on what you're doing.

SkillWhen to use
ai-system-designHave an idea but don't know where to start
problem-explorationFacing a problem with multiple possible approaches
agent-architectureDeciding architecture — workflow vs agent, single vs multi
prompt-designWriting a new prompt, or diagnosing one producing wrong output
experiment-driven-developmentStarting any implementation task
prompt-change-managementAbout to change a prompt
regression-testingComparing two system versions
critic-judge-designDesigning a Judge, Critic, or Evaluator component
harness-designAgent misbehaves — wrong tools, missed data, loops
agent-integrationConnecting agent to a web app, API, or third-party platform
database-designDesigning a schema (general or AI-specific)
model-selectionChoosing a model, API vs local, fine-tuning decisions
memory-systemDesigning how an agent remembers across sessions
devopsDeploying to production, CI/CD, monitoring
code-reviewReviewing code with structured severity levels

All skills read from and write to PROJECT.md — a shared state contract at the repo root that keeps decisions, progress, and known issues in one place across sessions.


Philosophy

Treat every change as an experiment, not a fix. Every change needs before/after measurement. Regressions are expected.

Structure beats instructions. Input organization and output schema shape LLM reasoning more than prompt rules. Fix structure before adding rules.

Evidence before code. Start with real system output. Write down the failure mode and hypothesis before touching any file.

One job per component. If it does two things, split it.


Installation

Claude Code

/plugin marketplace add Victoriakaey/build-reliable-agents
/plugin install build-reliable-agents@build-reliable-agents

claude.ai (Manual)

git clone https://github.com/Victoriakaey/build-reliable-agents.git
cp -r build-reliable-agents/skills/ /mnt/skills/user/build-reliable-agents/

Update

/plugin update build-reliable-agents

Contributing

Contributions welcome. Each skill should be grounded in real observed failures, not theoretical best practices. Follow the existing SKILL.md structure.


License

MIT License. See LICENSE for details.