Instruction Budget — Development skill for Claude Code
A public repo of an evaluations research project.
How to install Instruction Budget
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open Arize-ai/instruction-budget and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Instruction Budget does
A public repo of an evaluations research project.
Alternatives in Development
- GDScript VS C# — Skill Instruction Comparison — GDScript vs C# — Skill Instruction Comparison skill 2.8k ★
- 04 Cyber Risk Instruction — Cyber Risk Instruction 2.3k ★
- Claude Code Plugins — jeremylongshore Instruction-template plugins and MCP plugin packs 1.7k ★
README
instruction-budget
IFScale benchmark replication on the May 2026 frontier models, instrumented with Arize AX.
The full plan is in [IFSCALE_REPLICATION.md](IFSCALE_REPLICATION.md). Read it before running anything.
Quickstart
python -m venv .venv && source .venv/bin/activate
pip install -e .
# 1. Generate the (model, density, seed) -> keywords sampling matrix
python -m scripts.generate_samples
# 2. Run the smoke grid against Claude Opus 4.7 (week-1 self-validation)
python -m src.runner --model claude-opus-4-7 --densities 10,100,250,500 --seeds 1
Required env vars
In `.env` (already gitignored):
ANTHROPIC_API_KEY=...
OPENAI_API_KEY=...
GEMINI_API_KEY=...
DEEPSEEK_API_KEY=...
# Arize AX — required. The runner will refuse to start without these.
ARIZE_SPACE_ID=...
ARIZE_API_KEY=...
Layout
See §5.1 of [IFSCALE_REPLICATION.md](IFSCALE_REPLICATION.md). Reference code from the paper authors lives in [ifscale-original/](ifscale-original/).
Related Skills
Research Field Onboarding
A reusable research-onboarding skill for Claude, ChatGPT, Codex, and other instruction-following agents.
Lab Nt
The discoverer — shape a research idea into a contract (question · falsifiable metric · fence · budget), then
Skill Evaluations Bottom 10
Bottom 10 Skills: Skill-Judge Evaluations with TDD Improvements
App Eval
/app-eval — run deterministic agent evaluations
Harbor
Harbor is a framework for running agent evaluations and creating and using RL environments.
Claude Keysmith
Managed Claude Code instruction deployment with safe CLI recovery and an unsigned macOS/Windows desktop beta.
Related Agents
Research Paper Writer
Use this agent when you need to write academic papers, research reports, or technical manuscripts based on exp
Lead Researcher
Use this sub-agent to conduct deep web research on a batch of B2B leads. Spawn one instance per batch of 5 lea
Staff Engineer Architect
Use proactively for complex technical problems requiring deep architectural analysis, system design decisions,