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 instructions/slbug/claude-ruby-grape-rails/eval-reviewgit clone --depth 1 https://github.com/slbug/claude-ruby-grape-railsWhat 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 | $0.01586 | $0.01586 |
| Opus 5 | $0.00793 | $0.00793 |
| Sonnet 5 | $0.00317 | $0.00317 |
| Haiku 4.5 | $0.00159 | $0.00159 |
Grade A, and why
claude-ruby-grape-rails eval-review.instructions.md 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.
How it starts
The opening of the file, as written. The whole thing — 139 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Eval Framework Review Rules
Audience: Agents, Not Humans
Imperative-only.
Eval Definitions (lab/eval/evals/*.json)
- Each JSON defines dimensions: completeness, accuracy, conciseness, triggering, safety, clarity
- Each dimension has a weight (float, all weights should sum to ~1.0) and an array of checks
- Check types must match functions in matchers.py MATCHERS dict
- Skill-specific checks override the default_eval() fallback in scorer.py
- Eval definitions should not exceed 3KB
Trigger Corpora (lab/eval/triggers/*.json)
- Each JSON has: skill, should_trigger, should_not_trigger, hard_should_trigger, hard_should_not_trigger
- Minimum counts: 4 should_trigger, 4 should_not_trigger, 2 hard_should_trigger, 2 hard_should_not_trigger
- Hard prompts must have an
axisfield with >= 2 distinct values across hard_should_trigger (typically "confusable" and "multi_step") - Prompts must not contain skill names (routing contamination)
- No duplicate prompts within a file (normalized comparison)
Context Budget (lab/eval/context_budget.py)
- Advisory checks for CLAUDE.md size and aggregate routing-prompt char budget (8,000-char hard ceiling on the skill listing per CC routing budget)
- Zero API cost — file reads and frontmatter scanning only
- Wired into
--changed,--all, and--cimodes in run_eval.sh
Module Map
The canonical module list is ls lab/eval/*.py. Re-derive on every
review pass; do not rely on a frozen enumeration here.
When reviewing a change under lab/eval/:
- Read the module's docstring and
importblock — they declare its role and whether it touches an LLM provider. - Check whether
make eval-ci-deterministicreaches it transitively (seelab/eval/run_eval.sh --cibody and the allowlist inlab/eval/tests/test_eval_ci_determinism.py::DETERMINISTIC_PATH_FILES). Modules on that path MUST stay LLM-free. - LLM-bearing modules (
behavioral_scorer.py,epistemic_suite.py,trigger_scorer.py --semanticpath) are intentionally OFF the deterministic path. Confirm any new transport import goes into one of those, not into the deterministic path.
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 · 139 lines · 1,586 tokens per session scan A 082851a01a47
claude-ruby-grape-rails eval-review.instructions.md is an instructions file published in the GitHub repository slbug/claude-ruby-grape-rails (7 stars, last pushed 3d ago), licensed MIT. It adds 1,586 tokens to every session, about $0.0079 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 instructions, from other repositories
autotask-mcp dev_workflow.instructions.md
Guide for using Taskmaster to manage task-driven development workflows.
ai-workflow AGENTS.md
Instructions for cunhaax/ai-workflow, covering ai workflow template, rules — non-negotiable, project overview, commands and architecture.
trackly-cli CLAUDE.md
Claude Code instructions for trackly-app/trackly-cli, covering trackly-cli, tech stack, backend production source of truth, directory structure and key commands.
foggy-data-mcp-bridge-python CLAUDE.md
Instructions for foggy-projects/foggy-data-mcp-bridge-python, covering foggy data mcp bridge — python, 快速启动, 安装依赖, 运行测试 and 启动 mcp 服务(连接 docker mysql).
McpServer AGENTS.md
AGENTS.md instructions for sharpninja/McpServer, covering agent instructions, session start, rules, byrd test gate and where things live.
AmbyKit CLAUDE.md
Instructions for ambystechcom/AmbyKit, covering claude code — notes for the ambykit repo and claude-specific.