| name | create-agent |
| description | "Generates a validation agent .md file for any workflow step. The agent scores outputs 1-100 across 5 dimensions, provides actionable rewrites, and enforces hard violations. Use when adding quality gates to new or existing skills. Trigger on: 'create a validation agent', 'build a QA agent', 'I need a quality gate for [step]'. Output is a standalone agent-[name].md file ready for Project Knowledge upload." |
Create Agent
Before Starting
Confirm before starting:
This meta-skill generates validation agents for any workflow step. Each agent is a standalone .md file that can be uploaded to Project Knowledge and used as a quality gate before any output moves downstream.
Input: Step name + validation focus
Output: agent-[name].md — ready for Project Knowledge upload
Phase 1: Define the Validation Scope
Ask the user:
- What step does this agent validate? (e.g. "creative brief", "hook bank", "Nano Banana prompt")
- What does "good" look like? What are the non-negotiable quality markers?
- What are the most common failure modes? What goes wrong when this step is rushed?
- Who consumes the output? What downstream skill or person needs this to be right?
- What frameworks apply? Which Scale Academy frameworks are relevant?
Relevant frameworks to consider:
- Schwartz — 5 awareness levels (Unaware, Problem Aware, Solution Aware, Product Aware, Most Aware)
- LF8 — Whitman's 8 innate desires (survival, food, freedom from fear, sexual companionship, comfortable living, superiority, care for loved ones, social approval)
- C1-C4 matrix — Demonstration / Education / Pain / Identity angles
- Cialdini — 7 principles (reciprocity, commitment, social proof, authority, liking, scarcity, unity)
- Creative mechanics — M1-M8 (Implied Answer, Reframe, Trojan Horse, etc.)
Phase 2: Design the 5 Scoring Dimensions
Each dimension must be:
- Specific to the step being validated (not generic quality metrics)
- Observable — the evaluator can check it without subjective judgment
- Graduated — clear difference between a 60, 80, and 95
For each dimension, define:
- Name and description
- What a 95 score looks like (example)
- What an 80 score looks like (example)
- What a 60 score looks like (example)
The final score is the average of all 5 dimensions. Nothing ships below 90.
Phase 3: Define Hard Violations
Hard violations are instant fails regardless of score. They represent errors so fundamental that no amount of quality elsewhere compensates.
Good hard violations are:
- Objective — anyone can verify them (yes/no, not a judgment call)
- Critical — they break the output for the downstream consumer
- Non-compensable — a perfect score on everything else doesn't fix them
Typical categories:
- Missing required data (empty fields, absent sections)
- Framework misapplication (wrong awareness stage, no angle justification)
- Accuracy failures (paraphrased verbatims, invented data)
- Compliance breaches (banned words, grammar errors in final copy)
Phase 4: Write Calibration Examples
For each of the 5 dimensions, write concrete examples at three score levels:
- 95 — excellent, ships immediately
- 80 — good but needs a specific fix
- 60 — significant issues, requires rework
These examples are the calibration tool. Without them, scoring is subjective.
Phase 5: Define the Review Process
Write the step-by-step process the agent follows:
- Load the output to validate
- Check hard violations first (if any → instant fail, stop here)
- Score each dimension 1-100 with reasoning
- Calculate overall score (average of 5 dimensions)
- List what works and why
- List what does not work and why
- For each issue: provide a specific rewrite ("change X to Y because Z")
- Final verdict: PASS (90+) / REVISE (70-89) / FAIL (<70 or hard violation)
Phase 6: Generate the Agent File
Use this template to produce the final agent-[name].md:
# Agent: [Name]
> Validation agent for [what it validates]. Score minimum: 90/100.
## Focus
[What this agent validates — 2-3 sentences]
## Scoring Dimensions (each scored 1-100)
### 1. [Dimension Name]
[Description + what good looks like]
- **95:** [concrete example]
- **80:** [concrete example]
- **60:** [concrete example]
### 2. [Dimension Name]
[Same structure]
### 3. [Dimension Name]
[Same structure]
### 4. [Dimension Name]
[Same structure]
### 5. [Dimension Name]
[Same structure]
## Hard Violations (instant fail)
- [Violation 1 — objective, verifiable]
- [Violation 2]
- [Violation 3]
## Review Process
1. Load the output to validate
2. Check hard violations first — if any triggered, FAIL immediately
3. Score each dimension 1-100 with written reasoning
4. Calculate overall score (average of 5 dimensions)
5. List what works and why (be specific)
6. List what does not work and why (be specific)
7. For each issue: specific rewrite — "change X to Y because Z"
8. Final verdict: PASS (90+) / REVISE (70-89) / FAIL (<70 or hard violation)
## Output Format
[Agent name] — [Output being validated] — [Date]
**Hard Violations:** NONE / [list]
| Dimension | Score | Key reasoning |
|---|---|---|
| [Dim 1] | [X/100] | [1-line reasoning] |
| [Dim 2] | [X/100] | [1-line reasoning] |
| [Dim 3] | [X/100] | [1-line reasoning] |
| [Dim 4] | [X/100] | [1-line reasoning] |
| [Dim 5] | [X/100] | [1-line reasoning] |
| **Overall** | **[X/100]** | |
**What works:**
- [Point 1]
- [Point 2]
**What does not work:**
- [Issue 1] → Rewrite: [specific fix]
- [Issue 2] → Rewrite: [specific fix]
**Verdict:** PASS / REVISE / FAIL
Save as agent-[name].md in the relevant category folder.
Deliver
Present the generated agent file and confirm:
"Here's your validation agent for [step]. It scores across 5 dimensions, enforces [N] hard violations, and requires a minimum of 90/100 to pass. Upload it to Project Knowledge to activate it as a quality gate."
Related Skills
- All skills in the repository — any skill output can have a validation agent
01-audit/brand-research — example: agent-brand-research.md validates the brief
03-strategy/creative-brief — example: agent-creative-brief.md validates the brief
04-production/static-production — example: agent-static-production.md validates the prompt