AI Plugin Index
grade-iterate
Skill by alirezarezvani · updated · ★ 29k
Runs the grade and iterate loop for a Claude Managed Agent. It defines a rubric-graded outcome with a capped number of iterations, reads each verdict, and runs held-back eval cases in parallel. Built for developers building Claude Managed Agents. grade-iterate is published by alirezarezvani on GitHub.
Install grade-iterate
Built from the published package, never by AIgit clone --depth 1 https://github.com/alirezarezvani/claude-skills /tmp/claude-skills && mkdir -p ~/.claude/skills && cp -r /tmp/claude-skills/agent-launcher/skills/grade-iterate ~/.claude/skills/grade-iterateRun in your terminal. Claude Code loads skills from ~/.claude/skills automatically.
git clone --depth 1 https://github.com/alirezarezvani/claude-skills /tmp/claude-skills && mkdir -p ~/.gemini/antigravity-cli/skills && cp -r /tmp/claude-skills/agent-launcher/skills/grade-iterate ~/.gemini/antigravity-cli/skills/grade-iterateAntigravity CLI loads skills from ~/.gemini/antigravity-cli/skills, or .agents/skills for one project.
Skills can run scripts on your machine. Some need an API key: check the README.
Last commit in the past day
MIT: permissive, fine for commercial use
Clear description, installs from GitHub, 28,618 stars, original repository
No risky patterns in the files we checked
About grade-iterate
This skill covers phase 3 of building a Claude Managed Agent. It defines a CMA outcome, which requires a markdown rubric graded by an isolated grader. The outcome payload is built by outcome_builder.py, with max_iterations clamped between 1 and 20. After each run, verdict_reader.py reads the grader result and recommends the next move: ship, sharpen and re-run, escalate, or resume. Once a version passes, eval_scaffold.py generates held-back cases and a parallel run plan capped at the 25-thread CMA ceiling.
The loop is always bounded. Each iteration should move at least one rubric line from fail to pass, or the run halts at the cap and escalates. The skill is for developers who want their agent to meet a rubric without iterating forever. It is distinct from the first-launch and scheduling phases, which the source names as separate skills.
Written from the project's SKILL.md and checked against it.
What grade-iterate can do
- outcome_builder.pyBuilds the user.define_outcome payload with a required rubric and a capped max_iterations.
- verdict_reader.pyReads the grader result and recommends SHIP, SHARPEN, ESCALATE, or RESUME.
- eval_scaffold.pyGenerates held-back eval cases and a parallel run plan capped at 25 threads.
- Bounded grade-iterate loopStops at the iteration cap and escalates if no rubric line moves from fail to pass.
What you need
- Python 3
- A Claude Managed Agent build sheet in JSON
- Claude Managed Agent (CMA) setup
Weekly newsletter
Like grade-iterate? Get the best new skills every week
The best new tools, security warnings before you install something risky, and one practical tip. Free, three minute read.
Try asking
- Grade my agent against the rubric and tell me if it is good enough
- Build a user.define_outcome payload with max_iterations set to 5
- Read the last verdict and recommend the next move
Good to know
- max_iterations must be between 1 and 20
- A rubric is required for every outcome
- Parallel eval runs are capped at the 25-thread CMA ceiling
Security check
ANo download-and-run commands, hidden instructions or signs of sending keys out were found.
Checked automatically: agent-launcher/skills/grade-iterate/SKILL.md, agent-launcher/skills/grade-iterate/scripts/eval_scaffold.py, agent-launcher/skills/grade-iterate/scripts/outcome_builder.py, agent-launcher/skills/grade-iterate/scripts/verdict_reader.py. This looks for known risky patterns and is not a full audit. How grades work
Inside the skill
-
README.md997 B -
SKILL.md4 KB -
scripts/eval_scaffold.py4 KB -
scripts/outcome_builder.py4 KB -
scripts/verdict_reader.py4 KB
Alternatives to grade-iterate
All alternatives →
Use after a Codex or Claude Code feature has been implemented in Qwen Code to run the selected reference agent and Qwen Code under the same scenario, capture HTTP and terminal traces, compare request bodies, tool/function schemas, outputs, and iterate until the reproduced behavior is close enough.
Retune a skill corpus for a new model, measurement-first: mine the run archive for a baseline, establish a noise floor, audit the corpus adversarially, then cut in measured passes until a pre-registered bar clears. Requires a benchmark harness that can A/B two builds of the corpus; refuses without one.
Use when building CLI tools, implementing argument parsing, or adding interactive prompts. Invoke for parsing flags and subcommands, displaying progress bars and spinners, generating bash/zsh/fish completion scripts, CLI design, shell completions, and cross-platform terminal applications using commander, click, typer, or cobra.
Audit or update English SDK documentation against the requested implementation scope.
Creates, optimizes, and iteratively refines agent prompts, system prompts, developer prompts, and reusable prompt templates. Use when asked to improve a prompt, optimize a system prompt, rewrite an agent prompt, tune prompt wording, make a prompt more reliable, port prompts between OpenAI, Claude, or Gemini, or build prompt evals.
Main application building orchestrator. Creates full-stack applications from natural language requests. Determines project type, selects tech stack, coordinates agents.
grade-iterate: common questions
What is a CMA outcome in this skill?
An outcome is a markdown rubric graded by an isolated grader. The grader's verdict feeds failing results back for the next attempt.
Can the grade-iterate loop run forever?
No. The source says the loop is always bounded by max_iterations, which is clamped from 1 to 20.
What are held-back eval cases?
They are cases the agent never saw during iteration. They are run in parallel against the same rubric to test generalization.
How do I install grade-iterate?
Run the command above to copy the skill into ~/.claude/skills/. Claude Code picks it up automatically.
Is grade-iterate free?
The code is published under the MIT licence. The service it connects to may have its own pricing.
Is grade-iterate safe to use?
It scores 92 of 100. Grades: maintenance A, licence A, quality A. Security check: grade A, no risky patterns found in the files checked. Review the code before you install. How grades work.
What are alternatives to grade-iterate?
Similar skills: agent-reproduce-align, ce-retune, cli-developer, docs-sync.