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 rules/sijeeshmiziha/visionagent/tool-evaluationgit clone --depth 1 https://github.com/sijeeshmiziha/visionagentWhat 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 | $0.00395 | $0.00395 |
| Opus 5 | $0.00198 | $0.00198 |
| Sonnet 5 | $0.00079 | $0.00079 |
| Haiku 4.5 | $0.00040 | $0.00040 |
Grade A, and why
tool-evaluation 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 2d 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
Tool Evaluation and Improvement
Mindset
- Tools are contracts between deterministic systems and non-deterministic agents; design for agents, not just for APIs.
- Use evaluation to measure and improve tool effectiveness.
Process
- Prototype: Build a quick prototype (e.g. MCP server or tools passed into the API); test locally and collect user feedback.
- Evaluate: Run a programmatic evaluation—one agentic loop per task, each with a single task prompt and your tools.
- Analyze: Read transcripts and reasoning; identify where agents get stuck or misuse tools.
- Collaborate: Feed evaluation transcripts into an agent (e.g. Claude Code) to refactor tools and descriptions; use a held-out test set to avoid overfitting.
Evaluation Tasks
- Ground tasks in real-world use; prefer multi-step, multi-tool scenarios over trivial sandbox prompts.
- Pair each prompt with a verifiable outcome (exact match or LLM-as-judge). Avoid verifiers that reject correct answers due to formatting or phrasing.
- In evaluation system prompts, ask agents to output reasoning or feedback blocks (or use extended thinking) to diagnose why tools are or aren’t used.
Metrics
- Track accuracy plus: tool-call count, token consumption, tool errors, runtime.
- Redundant tool calls may suggest pagination or token limits; many invalid-parameter errors may suggest clearer descriptions or examples.
Analysis
- Read evaluation agents’ reasoning and raw transcripts (including tool calls and responses); “read between the lines”—agents don’t always say what they mean.
- Use metrics to spot improvement areas: e.g. rightsizing pagination, clarifying parameter descriptions, consolidating common workflows.
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.
- 2d ago First seen · 35 lines · 395 tokens per session scan A 8ee0d3fa7fdd
tool-evaluation is a cursor rule published in the GitHub repository sijeeshmiziha/visionagent (2 stars, last pushed 5mo ago), licensed MIT. It adds 395 tokens to every session, about $0.0020 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 cursor rules, from other repositories
legacy-letter
Write a letter to the people you love, to be read later — the things you'd want them to know, the stories only you hold, the permission and the love that outlive you. Use when someone says 'help me write a letter to my kids/partner', 'legacy letter', 'ethical will', 'something for them to have when I'm gone', or is…
mechanic-quote-decoder
Read a garage quote or invoice like someone who can't be padded — which line items connect to your actual symptom, which are while-we're-in-there additions, the questions that make soft lines disappear, when a second opinion pays for itself, and the scripts for declining work without souring the relationship. Use when…
the-org-simulator
Stress-test a proposed org change before announcing it — simulate who gains, who loses, who blocks, where friction erupts in the first 90 days, and run the memo leak test: how does this land when it leaks before you announce it? Use when planning a reorg, changing reporting lines, merging or splitting teams, moving a…
api-for-yourself
Publish 'how to work with me' as a literal API spec — endpoints (what to ask me for and what you'll get back), rate limits (meeting and interrupt tolerance), error codes (what happens when you surprise me Friday 5pm), auth (how to earn trust), and a changelog. Use when onboarding to a new team, when a new manager or…
java-springboot-jpa-cursorrules-prompt-file
description: "Cursor rules for Java development with Springboot and JPA integration." globs: / alwaysApply: false.
react
Cursor rule "react" from ItamarZand88/awesome-agent-conventions, covering react best practices, component structure, hooks, state management and performance.