agentic-eval-first-development
Architect, execute, and iterate on AI evaluations using the Data-Task-Score framework. Treats evals as the modern, quantifiable version of a PRD. Use when the user asks to "build an eval," "improve model quality," "test an agent workflow," "quantify product intuition," "move beyond vibe checks," "measure AI output," "score LLM responses," "benchmark a prompt," or "set up evaluation infrastructure." Also triggers on phrases like "how do I know if this is working," "is the model getting better," or "eval-driven development."
What this skill does
# Agentic Eval-First Development
Evals are infrastructure, not afterthoughts. Define success criteria *before* writing prompts or task logic. The eval becomes the spec.
## Framework: Data → Task → Scores
Every eval has exactly three components:
1. **Data** — Golden dataset of inputs (the test cases)
2. **Task** — The operation being evaluated (LLM call, agent workflow, MCP pipeline)
3. **Scores** — Categorical rubric that maps outputs to normalized 0–1 values
## Step 1: Define the PRD (Data & Scores)
### Build the Golden Dataset
Collect or generate **10–20 representative inputs** covering the full range of expected usage.
- Use a high-reasoning model to autogenerate diverse test cases if manual examples are unavailable
- **Intentionally include inputs expected to fail** — these map current model limitations
- Store as JSON or JSONL for reproducibility. See [references/golden-dataset-template.md](references/golden-dataset-template.md) for the format
### Define the Scoring Rubric
Use **categorical scoring** (Options A/B/C) rather than asking for raw numbers. Raw numeric scores drift across evaluators and models.
- Every score must include a written **rationale** explaining the grade
- All scores normalize to **0–1** for cross-model comparison. See [references/scoring-rubrics.md](references/scoring-rubrics.md) for rubric templates
- Run `scripts/normalize_scores.py` to convert categorical results to normalized values
**Example categorical scorer:**
```
A (1.0) — Fully correct, well-structured, addresses all aspects
B (0.5) — Partially correct or missing key elements
C (0.0) — Incorrect, off-topic, or harmful
```
## Step 2: Configure the Task (The Harness)
The task is the operation under evaluation.
1. **Tool Pruning** — If using MCP, limit available tools to only what's necessary. Models select incorrect tools when overwhelmed with options
2. **System Prompt** — Define initial instructions based on success criteria from Step 1 (e.g., "don't ask clarifying questions," "respond in JSON")
3. **Isolation** — Each eval run must be independent. No shared state between test cases
## Step 3: Execute the Flywheel Loop
```
┌─────────────────────────────────────────┐
│ OFFLINE: Run golden dataset locally │
│ → Identify gaps → Refine prompt/tools │
└──────────────┬──────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ ONLINE: Deploy scorers to production │
│ → Monitor real user logs │
└──────────────┬──────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ CLOSE THE LOOP: Production failures │
│ → Add back to golden dataset │
└─────────────────────────────────────────┘
```
1. **Offline iteration** — Run experiments locally against the golden dataset. Iterate on prompts, tools, and model selection until scores stabilize
2. **Online validation** — Deploy scorers to production monitoring real user logs
3. **Close the loop** — When online score (e.g., 0.3) < offline score (e.g., 0.75), identify production failures and add them to the golden dataset
## When to Stop Iterating
- Offline scores plateau across 3+ consecutive runs
- Online/offline gap is < 0.1
- Remaining failures are edge cases outside the product's scope
## Troubleshooting
| Symptom | Likely Cause | Fix |
|---------|-------------|-----|
| All scores are 0 | Scorer criteria too strict | Do a manual vibe check — if you disagree with the scorer, update the rubric |
| Scores are always 1.0 | Scorer criteria too lenient or test cases too easy | Add adversarial inputs and tighten rubric |
| Online ≪ Offline | Golden dataset doesn't represent real usage | Add production failure cases to dataset |
| Scores vary wildly between runs | Non-deterministic task or scorer | Pin temperature=0, add more specific rubric criteria |
## Key Principle
The eval is the **durable asset**. Models change, prompts evolve, agent frameworks get replaced — but a well-built eval survives all of it. When switching models, re-run the eval; don't re-do the product thinking.
Related in AI Agents
skill-development
IncludedComprehensive meta-skill for creating, managing, validating, auditing, and distributing Claude Code skills and slash commands (unified in v2.1.3+). Provides skill templates, creation workflows, validation patterns, audit checklists, naming conventions, YAML frontmatter guidance, progressive disclosure examples, and best practices lookup. Use when creating new skills, validating existing skills, auditing skill quality, understanding skill architecture, needing skill templates, learning about YAML frontmatter requirements, progressive disclosure patterns, tool restrictions (allowed-tools), skill composition, skill naming conventions, troubleshooting skill activation issues, creating custom slash commands, configuring command frontmatter, using command arguments ($ARGUMENTS, $1, $2), bash execution in commands, file references in commands, command namespacing, plugin commands, MCP slash commands, Skill tool configuration, or deciding between skills vs slash commands. Delegates to docs-management skill for official documentation.
reprompter
IncludedTransform messy prompts into well-structured, effective prompts — single or multi-agent. Use when: "reprompt", "reprompt this", "clean up this prompt", "structure my prompt", rough text needing XML tags and best practices, "reprompter teams", "repromptception", "run with quality", "smart run", "smart agents", multi-agent tasks, audits, parallel work, anything going to agent teams. Don't use when: simple Q&A, pure chat, immediate execution-only tasks. See "Don't Use When" section for details. Outputs: Structured XML/Markdown prompt, quality score (before/after), optional team brief + per-agent sub-prompts, agent team output files. Success criteria: Single mode quality score ≥ 7/10; Repromptception per-agent prompt quality score 8+/10; all required sections present, actionable and specific.
adaptive-compaction
IncludedAdaptive add-on policy and recovery layer that decides WHEN to compact, prune, snapshot, or fork -- replacing fixed-percent auto-compaction across Claude Code, Codex, and MCP-capable hosts. Trigger on auto-compact timing or damage: "when should I compact", "is it safe to compact now or start a fresh session", "auto-compact fires too early/mid-task", "switching to an unrelated task but the window still has space", "context rot", "answers get worse the longer the session runs", "the agent forgot the plan or my decisions after it summarized", "add a layer on top that manages context without changing the agent", raising autoCompactWindow to give the policy room, or installing/tuning a cross-tool compaction policy or PreCompact hook -- even when "compaction" is never said but the problem is context-window pressure or post-summarization memory loss. Do NOT use to summarize a conversation, build RAG, write a summarization prompt (decides WHEN not HOW), or answer max-context-length trivia.
agent-skill-creator
IncludedCreate cross-platform agent skills from workflow descriptions. Activates when users ask to create an agent, automate a repetitive workflow, create a custom skill, or need advanced agent creation. Triggers on phrases like create agent for, automate workflow, create skill for, every day I have to, daily I need to, turn process into agent, need to automate, create a cross-platform skill, validate this skill, export this skill, migrate this skill. Supports single skills, multi-agent suites, transcript processing, template-based creation, interactive configuration, cross-platform export, and spec validation.
llm-wiki
IncludedUse when building or maintaining a persistent personal knowledge base (second brain) in Obsidian where an LLM incrementally ingests sources, updates entity/concept pages, maintains cross-references, and keeps a synthesis current. Triggers include "second brain", "Obsidian wiki", "personal knowledge management", "ingest this paper/article/book", "build a research wiki", "compound knowledge", "Memex", or whenever the user wants knowledge to accumulate across sessions instead of being re-derived by RAG on every query.
skill-master
IncludedAgent Skills authoring, evaluation, and optimization. Create, edit, validate, benchmark, and improve skills following the agentskills.io specification. Use when designing SKILL.md files, structuring skill folders (references, scripts, assets), ingesting external documentation into skills, running trigger evals, benchmarking skill quality, optimizing descriptions, or performing blind A/B comparisons. Keywords: agentskills.io, SKILL.md, skill authoring, eval, benchmark, trigger optimization.