forge

name	forge
version	1.0.0
description	Self-evolving AI agent skill for UI/UX design, code intelligence, and design system generation. Activates contextual rules based on task type — only loads what's needed. Use for any frontend design, component creation, page layout, design system, or code architecture task. Supports React, Vue, Svelte, HTML/CSS. Self-learning enabled: collects feedback and evolves rules over time.
triggers	["design","UI","UX","component","layout","theme","page","frontend","CSS","Tailwind","React","Vue","Svelte","redesign","audit","a11y","accessibility"]

FORGE — Contextual Design Intelligence

Activation Protocol

1. Read DESIGN.md if it exists in project root or .skills/
2. Read memory/activeContext.md for session continuity
3. Classify the task using the matrix below
4. Load ONLY the modules the task requires
5. Generate output following loaded module rules
6. Run Output Audit (8-point checklist) before delivery
7. Append feedback to evolution/feedback.csv

Task Classification Matrix

Signal Keywords	Domain	Complexity	Modules to Load	Est. Tokens
fix, adjust, tweak, change, update	component-fix	low	code-scaffold (subset)	0.3-0.8K
style, color, font, theme, palette	style-adjust	low	design-intelligence (dials)	0.6-1.0K
create, build, new component, add	component-create	med	code-scaffold + a11y-enforcer	2.5-3.5K
redesign, revamp, rebuild, overhaul	page-redesign	high	design-intel + code-scaffold + DESIGN.md	4-6K
design system, brand, theme setup	system-gen	high	design-system-generator + CSVs	5-8K
refactor, restructure, reorganize	code-restructure	med	codegraph-rules + code-scaffold	3-5K
audit, review, check, evaluate	quality-audit	low	output-auditor + a11y + anti-patterns	1.5-2.5K

Priority rule: If multiple signals match, use the highest complexity. Override: User can request full rules with /forge:full or minimal with /forge:min.

Module Loading Sequence

1. DESIGN.md → always (if exists)
2. Classified module → from matrix above
3. CodeGraph → only if task involves existing codebase
4. Memory → read activeContext.md at session start
5. Output audit → always before delivery
6. Feedback → always after output

Context Loading Rules

• Never load all modules simultaneously. Each module is independent.
• design-intelligence.md: Design taste, aesthetics, dials, tones, anti-slop
• code-scaffold.md: Implementation patterns, CVA, a11y, component architecture
• design-system-generator.md: CSV data queries, theme workflow, DESIGN.md generation
• codegraph-rules.md: When/how to query code structure (existing codebases only)
• memory-rules.md: When to read/write memory tiers
• token-rules.md: When to compress, cache, prune context
• guardrails.md: Safety and input validation
• a11y-enforcer.md: Accessibility enforcement (BLOCKING on fail)
• evaluator.md: Quality metrics and regression testing
• feedback-collector.md: Post-output feedback logging
• pattern-learner.md: Periodic feedback analysis
• rule-refiner.md: Rule change proposals
• output-auditor.md: 8-point pre-delivery audit
• regression-detector.md: Quality regression detection
• quality-scorer.md: Composite quality scoring
• rule-versioning.md: Version control for rule changes
• benchmark-runner.md: Automated quality benchmarks
• evolution-engine.md: Full evolution cycle orchestration

Output Audit (MANDATORY)

Before delivering any design/code output, run this 8-point checklist:

#	Check	Action on Fail
1	Anti-Pattern Scan	Check against data/anti-patterns.csv — auto-fix or flag
2	DESIGN.md Compliance	Verify tokens match DESIGN.md — auto-correct
3	Accessibility Gate	WCAG 2.1 AA — BLOCK delivery, force fix
4	Consistency Lock	One accent, one radius scale, consistent headings — auto-normalize
5	AI Slop Test	No cream bg, no purple glow, no generic hero — regenerate
6	Dial Compliance	Output matches declared dials — adjust
7	Framework Check	No framework anti-patterns — auto-correct
8	Component Completeness	All states present (hover, focus, disabled, error) — add missing

Check #3 is non-optional. All others produce warnings.

Feedback Collection (MANDATORY)

After every output, append one row to evolution/feedback.csv:

timestamp,task_type,modules_loaded,token_count,output_accepted,user_correction,correction_type,correction_detail,model_used,session_id

• output_accepted: yes / no / partial
• correction_type: color | spacing | typography | layout | animation | accessibility | code-quality | other
• This is passive and non-blocking — never interrupt the user's workflow

Three Dials (Quick Reference)

Every design task starts by inferring dial values:

Dial	Range	Default	Low =	High =
DESIGN_VARIANCE	1-10	5	Conservative, safe	Bold, experimental
MOTION_INTENSITY	1-10	3	Static, subtle	Animated, dynamic
VISUAL_DENSITY	1-10	5	Spacious, minimal	Dense, information-rich

Brief Inference: Before coding, declare your interpretation of the design intent and dial values. Show your Design Read.

Commands

Command	Action
/forge:full	Load all modules (override classification)
/forge:min	Load minimum modules only
/forge:audit	Run full output audit on current output
/forge:analyze	Run pattern-learner on accumulated feedback
/forge:evolve	Run full evolution cycle (learn → critique → evolve)
/forge:benchmark	Run 10-task benchmark suite
/forge:rollback v=X.Y.Z	Rollback rules to version X.Y.Z

Per-Model Defect Awareness

Model	Known Defect	Override
Gemini	Image hover effects break	Avoid hover-based image transforms
Codex	Typographic ceilings	Force explicit font-size/line-height
GPT-4o	Purple/indigo bias	Strict accent-color validation
AI Agent	Cream/beige background bias	Enforce tinted neutrals rule

Token Budget Discipline

• SKILL.md router: ≤ 4K tokens (this file)
• Individual modules: 0.3-2.5K each
• Total for simple task: < 1K
• Total for complex task: < 8K
• NEVER load all modules at once (18K+ total — defeats the purpose)