copilot-auto-training

Improve a markdown prompt file using Agent Lightning APO (Automatic Prompt Optimization). Use when the user asks to optimize or improve a markdown prompt, or starts a message with /trainer-optimize.

Own the end-to-end trainer loop for agent contract targets (*.agent.md files, custom agent definitions, and agent instruction documents). Use this whenever the caller needs to research, synthesize datasets, optimize, validate, and write back a trained candidate for an agent-type target. Prefer this specialized loop whenever the selected target defines tool routing, MCP skill configuration, agent personas, or handoff behavior rather than raw prompts, code, or skill definitions.

Own the end-to-end trainer loop for Python code targets optimized with Microsoft Trace (nodes, bundles, models, and trainable agent components). Use this whenever the caller needs to research, synthesize test-based datasets, optimize, validate, and write back a trained candidate for a code-type target. Prefer this specialized loop for any Python file or callable that benefits from deterministic, test-based or benchmark-based feedback rather than open-ended language instruction quality.

Own the end-to-end trainer loop for Python code targets optimized with Microsoft Trace (nodes, bundles, models, and trainable agent components). Use this whenever the caller needs to research, synthesize test-based datasets, optimize, validate, and write back a trained candidate for a code-type target. Prefer this specialized loop for any Python file or callable that benefits from deterministic, test-based or benchmark-based feedback rather than open-ended language instruction quality.

Own the end-to-end trainer loop for prompt-like files (*.prompt.md, *.prompty, *.instructions.md, system prompts, and other natural-language instruction artifacts). Use this whenever the caller needs to research, synthesize datasets, optimize, validate, and write back a trained candidate for a prompt-type target. Prefer this specialized loop for any file whose primary content is natural-language instructions rather than code, skill configuration, or agent contracts.

Own the end-to-end trainer loop for prompt-like files (*.prompt.md, *.prompty, *.instructions.md, system prompts, and other natural-language instruction artifacts). Use this whenever the caller needs to research, synthesize datasets, optimize, validate, and write back a trained candidate for a prompt-type target. Prefer this specialized loop for any file whose primary content is natural-language instructions rather than code, skill configuration, or agent contracts.

Own the end-to-end trainer loop contract for a prompt-like file, skill contract, or agent contract after the caller has already chosen the concrete stage capabilities. Use this whenever the current agent must set up the local trainer workspace, coordinate stage sequencing, maintain workflow state, manage steering and candidates, recover from manual follow-up mode, and decide whether a trained candidate is safe to write back.

Own the end-to-end trainer loop for agent skill targets (SKILL.md files and their supporting references, scripts, and evals). Use this whenever the caller needs to research, synthesize datasets, optimize, validate, and write back a trained candidate for a skill-type target. Prefer this specialized loop whenever the selected target is a SKILL.md file or the user wants to improve skill triggering accuracy, body content quality, or progressive-disclosure structure.

Own the end-to-end trainer loop contract for a prompt-like file, skill contract, or agent contract after the caller has already chosen the concrete stage capabilities. Use this whenever the current agent must set up the local trainer workspace, coordinate stage sequencing, maintain workflow state, manage steering and candidates, recover from manual follow-up mode, and decide whether a trained candidate is safe to write back.

Own the end-to-end trainer loop contract for a prompt-like file, skill contract, or agent contract after the caller has already chosen the concrete stage capabilities. Use this whenever the current agent must set up the local trainer workspace, coordinate stage sequencing, maintain workflow state, manage steering and candidates, recover from manual follow-up mode, and decide whether a trained candidate is safe to write back.

Own the end-to-end trainer loop contract for a prompt-like file, skill contract, or agent contract after the caller has already chosen the concrete stage capabilities. Use this whenever the current agent must set up the local trainer workspace, coordinate stage sequencing, maintain workflow state, manage steering and candidates, recover from manual follow-up mode, and decide whether a trained candidate is safe to write back.

Create, run, and manage AgentV evaluations for AI agents and skills using the AgentV CLI and AgentEvals standard EVAL.yaml format. Use this skill whenever the user wants to write evaluation files for AI agents, run evals with agentv CLI, convert existing test cases to EVAL.yaml, set up eval targets, understand the AgentEvals specification, debug failing evaluations, integrate evals into CI/CD pipelines, or compare agent runs with `agentv compare`. Also use this when the user mentions EVAL.yaml, agentevals, agentv, or wants to evaluate skill quality with a declarative format.

Create, run, and manage AgentV evaluations for AI agents and skills using the AgentV CLI and AgentEvals standard EVAL.yaml format. Use this skill whenever the user wants to write evaluation files for AI agents, run evals with agentv CLI, convert existing test cases to EVAL.yaml, set up eval targets, understand the AgentEvals specification, debug failing evaluations, integrate evals into CI/CD pipelines, or compare agent runs with `agentv compare`. Also use this when the user mentions EVAL.yaml, agentevals, agentv, or wants to evaluate skill quality with a declarative format.

Improve Python implementations with Microsoft Trace by turning prompts, helper functions, or small agent components into trainable code. Use this whenever the user wants to apply Trace or trace-opt, optimize Python behavior from tests or feedback, make a method trainable with nodes, bundles, or models, or decide how to structure a Trace training loop for code-focused work.

Improve GitHub Copilot custom agents by validating agent contracts, tightening tool and MCP skill routing, and minimizing prompt bloat while keeping handoffs bounded to real workspace agents. Use this whenever the user wants to create, debug, or refine a custom agent.

Improve broken prompts and context plans by choosing the smallest prompt-engineering technique that fits. Use this whenever the user asks how to rewrite or debug a prompt, compare prompt-design options, choose between grounding, structured output, examples, chaining, reasoning, or RAG for a prompt, or reduce prompt length by moving schemas, workflow specs, and repeated instructions into better structures.

Improve agent skills by validating structure, optimizing YAML frontmatter for triggering accuracy, and refining SKILL.md prompt content for reliable agent behavior. Use this whenever the user wants to create, improve, debug, or optimize an agent skill, fix skill triggering issues, extract deterministic instructions into scripts, restructure a skill for progressive disclosure, or reduce a bloated SKILL.md body.

Evaluate final outputs, response pairs, scored artifacts, or benchmark-style answer quality without relying on full trajectories. Use this whenever the judging task is mainly about end-state quality, final answer comparison, reference-plus-criteria scoring, or choosing the best output among candidates, even if the user only says compare responses, pick the best answer, or judge the final result.

Generate a formalized rubric for scoring, grading, or evaluation in the current domain. Use when a judging task needs locked dimensions, pass-partial-fail boundaries, evidence requirements, tie-breakers, or confidence guidance before candidate comparison.

Evaluate agent trajectories, tool-use traces, intermediate artifacts, runtime failures, and side effects when process quality is part of the verdict. Use this whenever the judging task involves agent runs, tool calls, planning quality, web or code traces, process reliability, or any comparison where the final answer alone is not enough, even if the user only says judge the run, compare traces, or evaluate the agent workflow.

Capture user corrections and reusable lessons from the active conversation, then update the right persistent artifact so the same mistake is less likely to happen again. Use this whenever the user wants a fix reflected in agent memory, `.agents/MEMORY.md`, instruction files, custom agents, skills, `AGENTS.md`, hooks, docs, evals, or tests instead of only in the current answer.

Research public datasets, benchmarks, documentation, and source material for official skill eval cases. Use this skill whenever a prompt or skill needs grounded public examples, authoritative dataset references, or a primary-source brief before synthesis or optimization.

Elect the strongest prompt or skill candidate from an existing evaluation workspace. Use this skill whenever a workflow already has multiple scored configurations and needs a separate leader-selection pass over grading, timing, or benchmark artifacts, especially when comparing optimizer outputs without pushing that selection logic back into the optimization runtime.

Build official `evals/evals.json` cases and explicit APO datasets from grounded or computed source data. Use whenever a prompt or skill needs eval rows, `train.jsonl`, or `val.jsonl`, especially when correct outputs must be derived from raw fields, business rules, or verifier-backed synthetic examples.

Improve a markdown prompt file using Agent Lightning APO (Automatic Prompt Optimization). Use when the user asks to optimize or improve a markdown prompt, or starts a message with /trainer-optimize.

Create or update a GitHub Agentic Workflow in .github/workflows using gh aw, including frontmatter, markdown instructions, optional MCP servers, compilation, and debugging. Use this whenever the user wants new repository automation, wants to turn a repeated GitHub process into an agentic workflow, needs to add external MCP tools to a workflow, or needs help validating and fixing a workflow before commit.

Generate a formalized rubric for scoring, grading, or evaluation in the current domain. Use when a judging task needs locked dimensions, pass-partial-fail boundaries, evidence requirements, tie-breakers, or confidence guidance before candidate comparison.

Elect the strongest prompt or skill candidate from an existing evaluation workspace. Use this skill whenever a workflow already has multiple scored configurations and needs a separate leader-selection pass over grading, timing, or benchmark artifacts, especially when comparing optimize outputs without pushing that selection logic back into trainer-optimize.