robcsaszar/ai-forge

robcsaszar/ai-forge

Plugin Claude Code

A workflow of skills for creating, updating, judging, reviewing, evaluating, and auditing AI skills and agent definitions.

robcsaszar/ai-forge

Instructions file CodexOpenCode

AGENTS.md instructions for robcsaszar/ai-forge, covering agents.md, mission, layout convention, judgment boundaries and pre-publish checks.

robcsaszar/ai-forge

Skill Claude CodeCodex

HITL (Human-In-The-Loop) application of a numbered list one item at a time — status board upfront, per-item approve/skip, approve-all mode, and one opt-in commit bundling all approved items after the loop. Use when stepping through ai-forge-judge findings or any numbered changes. Triggers are apply these, go through…

robcsaszar/ai-forge

Skill Claude CodeCodex

Batch-evaluate all skills and agents in the repo with ai-forge-judge and render a single consolidated grade report sorted by grade (worst first) so effort is directed correctly. Use when reviewing overall skill/agent quality, finding where to invest improvement effort, or after bulk changes. Triggers are audit all…

robcsaszar/ai-forge

Skill Claude CodeCodex

Create a new skill (SKILL.md), agent definition, or instruction file. Use when converting ad-hoc knowledge into a reusable skill, scaffolding an agent for Claude Code, GitHub Copilot, OpenAI Codex, or Google Gemini, or creating instruction files for glob-pattern matching. Don't use for updating existing artifacts …

robcsaszar/ai-forge

Skill Claude CodeCodex

Behavioral eval for skills and agents — whether the artifact actually changes model behavior, not just whether it scores well on a rubric. Use when verifying a skill or agent works in practice, or when tracking its performance across repeated runs to catch regressions. Triggers are test this skill, test this agent…

robcsaszar/ai-forge

Agent

You make blind judgments between two outputs. You do not know which came from a skill/agent and which was a baseline — do not ask, do not infer.

robcsaszar/ai-forge

Agent

You grade outputs against a list of expectations. You receive an output (text produced by an agent) and an expectations list (assertions about what the output should contain or demonstrate).

robcsaszar/ai-forge

Agent

You receive a completed blind comparison (arbiter output) and a label mapping revealing which output was "withartifact" vs "baseline". Your job: explain why the winner won and surface targeted improvements to the artifact.

robcsaszar/ai-forge

Skill Claude CodeCodex

Evaluate any LLM prompt (SKILL.md, agent definition, system prompts, instruction files) for quality — grouped dimensional scoring with letter grade and step-through-ready numbered improvements list. Triggers are judge/review/audit/score/evaluate this skill or prompt, grade this agent. Don't use for behavioral testing…

robcsaszar/ai-forge

Skill Claude CodeCodex

Report what a skill or agent actually does versus what its description claims — drift, undeclared behaviors, verdict. Use before updating a skill or agent, or when a description is suspected of being out of date. Triggers are recap [skill], what does [skill] do, audit [skill] description, summarize [skill]. Don't use…

robcsaszar/ai-forge

Skill Claude CodeCodex

Critically reviews and stress-tests agent, skill, and AI workflow definitions before they ship. Use whenever someone creates, modifies, or proposes an agent config, skill file, system prompt, or AI-powered workflow — including 'review this agent', 'check this skill', 'is this agent safe', or any request for feedback…

robcsaszar/ai-forge

Skill Claude CodeCodex

Revise an existing SKILL.md or agent definition. Use when an existing skill or agent needs revision, modification, or improvement — including when it misfires, triggers too broadly or too rarely, or has drifted from what its description claims. Don't use for creating new artifacts — use ai-forge-create for that.…