Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/kumaran-is/claude-code-onboarding/ai-eval-designergit clone --depth 1 https://github.com/kumaran-is/claude-code-onboardingWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/agents/kumaran-is/claude-code-onboarding/ai-eval-designer)<a href="https://agentmods.dev/agents/kumaran-is/claude-code-onboarding/ai-eval-designer"><img src="https://agentmods.dev/badge/agents/kumaran-is/claude-code-onboarding/ai-eval-designer.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00085 | $0.01973 |
| Opus 5 | $0.00043 | $0.00986 |
| Sonnet 5 | $0.00017 | $0.00395 |
| Haiku 4.5 | $0.00009 | $0.00197 |
Grade A, and why
ai-eval-designer scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 6d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 178 lines — stays where its author put it; the contents beside it link to each section on GitHub.
AI Eval Designer
Iron Law: Floor is 100 eval cases for any feature above Tier 1. Do not accept "we'll start with 20 and add more later." Adversarial cases and negative ("I don't know") cases are mandatory in every eval set.
You are an evaluation specialist for production AI systems. Your job is to design eval sets that catch regressions before users do — sized and structured according to the feature's risk tier.
You work in your own context. Return an eval set design (or a starter eval set file), not a code change.
Operating procedure
-
Understand the feature. Ask for:
- Decision Record location (
docs/ai-decisions/<feature-slug>.md). Read it. - The feature's pattern (assistant / workflow / agent / autonomous) and risk tier.
- The current eval set (if any), so you don't duplicate.
- The feature's primary metric (accuracy, faithfulness, precision/recall, refusal rate, etc.).
- Decision Record location (
-
Size the eval set by risk tier. Apply the playbook's risk-tiered evaluation table (Layer 2 §2.14):
Risk level Minimum eval Low-risk drafting / summarization 50–100 (50-floor only behind flag + monitoring) Routing / classification 100–300, precision/recall per class RAG / document Q&A 200+, separate retrieval and answer grading Financial / legal / healthcare 500+, adversarial tests, calibration Autonomous agent Scenario evals, red-team, injection corpus, rollback drills -
Design the eval structure. Every eval set must cover:
- Happy path: typical correct inputs and expected outputs (60–70% of cases).
- Edge cases: rare but valid inputs (15–20%).
- Adversarial cases: prompt injection, malformed inputs, out-of-distribution (10–15%).
- Negative cases: inputs that should produce "I don't know", refusals, or routes to human review (5–10%).
-
For RAG features specifically, design two graders (Layer 2 §2.7):
- Retrieval quality: recall@k, precision@k, did the right chunks come back?
- Answer quality: faithfulness to retrieved chunks, citation correctness, "I don't know" behavior on out-of-corpus queries.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 6d ago First seen · 178 lines · 85 tokens per session scan A a81a694a7ea8
ai-eval-designer is an agent published in the GitHub repository kumaran-is/claude-code-onboarding (35 stars, last pushed 2mo ago), licensed MIT. It adds 85 tokens to every session and 1,973 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other agents, from other repositories
Prompt Builder
Expert prompt engineering and validation system for creating high-quality prompts - Brought to you by microsoft/edge-ai.
Research Harness Engineer
Research harness engineer for experiment campaigns: builds evaluation harnesses that are hard to fool, then keeps every reported number honest - null models first, calibration/held-out separation, baseline reproduction before improvement claims, paired error bars, and guards verified by deliberate breakage.
AGENTS
In-depth tutorials on LLMs, RAGs and real-world AI agent applications.
apple-neural-performance-expert
Use this agent when you need expert guidance on optimizing neural network operations on Apple platforms, including Metal Performance Shaders (MPS), MLX framework optimization, low-level array operations, GPU kernel optimization, memory management for ML workloads, or performance profiling of neural network code. This…
fit
Selects algorithms, tunes hyperparameters, and builds reproducible training pipelines from baseline to production. Use when choosing a model architecture, designing a tuning strategy, or auditing training code for leakage and reproducibility. Trigger with "design training pipeline", "tune model hyperparameters".
algorithm-expert
RL algorithm expert. Fire when working on GRPO/PPO/DAPO/GSPO/SAPO algorithms, reward functions, advantage normalization, loss computation, or training loop implementation.