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.
git clone --depth 1 https://github.com/Amey-Thakur/AI-SKILLSWrote 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/commands/amey-thakur/ai-skills/agent-eval-design)<a href="https://agentmods.dev/commands/amey-thakur/ai-skills/agent-eval-design"><img src="https://agentmods.dev/badge/commands/amey-thakur/ai-skills/agent-eval-design/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/commands/amey-thakur/ai-skills/agent-eval-design"><img src="https://agentmods.dev/badge/commands/amey-thakur/ai-skills/agent-eval-design.svg" alt="Reviewed on agentmods" width="80" 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 | $0.00028 | $0.00457 |
| Opus 5 | $0.00014 | $0.00229 |
| Sonnet 5 | $0.00006 | $0.00091 |
| Haiku 4.5 | $0.00003 | $0.00046 |
Grade A, and why
agent-eval-design 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 11d 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.
What it actually says
You were invoked as a slash command. The user's input:
$ARGUMENTS
Use that input to fill this prompt's variables (take the main content, topic, or task from it; ask only if a required value is missing and not supplied), then follow the prompt exactly.
Design an evaluation for: {system}
Known/feared failures: {failures}
Build the eval:
- Define what "good" means for this system, concretely: the dimensions that matter (correctness, groundedness, format, tone, safety, task completion) and the bar for each. You cannot evaluate what you have not defined.
- Assemble the eval set from real and hard cases: representative inputs, the failure modes above, edge cases (empty, huge, adversarial, each language), and cases mined from actual failures. Start small and honest (50 good cases beat 5000 scraped ones).
- Choose the grader per dimension, cheapest sufficient first: programmatic checks where possible (schema valid, right answer, cites the right source), LLM-as-judge only for open-ended quality, with a rubric. Calibrate any LLM judge against human labels before trusting it.
- Test behaviors, not just outputs: invariance (paraphrase in, same answer), correct refusals (should-refuse and should-not-refuse sets), multi-turn behavior, and tool-use correctness for agents.
- Make it a regression gate: run on every prompt/model change, compare per-case not just averages (an average that hides a regression on one cluster is a net loss), and grow the set as new failures appear.
Rules: build the eval before optimizing (without it, every change is a blind bet). Offline evals estimate quality, not business impact (that needs live measurement). Every production failure becomes a new eval case. If the "good" definition is unclear, that is step zero, not something to skip.
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.
- 11d ago First seen · 43 lines · 28 tokens per session scan A 9ed88f32934a
agent-eval-design is a command published in the GitHub repository Amey-Thakur/AI-SKILLS (7 stars, last pushed 5d ago), licensed MIT. It adds 28 tokens to every session and 457 once invoked, about $0.0001 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 2026-08-31.
Other commands, from other repositories
verify-work
Execute a work item's testing-plan.md after implementation — run the planned activities, record results, commit the evidence (verification phase).
regression-test
Write a failing regression test, fix the bug, verify, and check for similar issues.
merge-check
Pre-merge quality gate with parallel verification. Runs build, archive, test, and lint checks.
complete-feature
Complete a feature with full validation across build, tests, lint, and patterns. Runs the complete-feature skill.
evaluate
Evaluate a skill across model tiers using blind testing.
checklist
Generate a custom checklist for the current feature based on user requirements.