feat(prompts): enforce category domain matching and design-system-first workflow

Remove deep parallel delegation section from GPT-5.4 Sisyphus prompt since
it encouraged direct implementation over orchestration. Add zero-tolerance
category domain matching guide to all Sisyphus prompts with visual-engineering
examples. Rewrite visual-engineering category prompt with 4-phase mandatory
workflow (analyze design system, create if missing, build with system, verify)
targeting Gemini's tendency to skip foundational steps.
This commit is contained in:
YeonGyu-Kim
2026-03-06 22:19:18 +09:00
parent c17f7215f2
commit 2e7b7c1f55
3 changed files with 104 additions and 6 deletions
+28 -1
View File
@@ -247,7 +247,34 @@ task(
**ANTI-PATTERN (will produce poor results):**
\`\`\`typescript
task(category="...", load_skills=[], run_in_background=false, prompt="...") // Empty load_skills without justification
\`\`\``
\`\`\`
---
### Category Domain Matching (ZERO TOLERANCE)
Every delegation MUST use the category that matches the task's domain. Mismatched categories produce measurably worse output because each category runs on a model optimized for that specific domain.
**VISUAL WORK = ALWAYS \`visual-engineering\`. NO EXCEPTIONS.**
Any task involving UI, UX, CSS, styling, layout, animation, design, or frontend components MUST go to \`visual-engineering\`. Never delegate visual work to \`quick\`, \`unspecified-*\`, or any other category.
\`\`\`typescript
// CORRECT: Visual work → visual-engineering category
task(category="visual-engineering", load_skills=["frontend-ui-ux"], prompt="Redesign the sidebar layout with new spacing...")
// WRONG: Visual work in wrong category — WILL PRODUCE INFERIOR RESULTS
task(category="quick", load_skills=[], prompt="Redesign the sidebar layout with new spacing...")
\`\`\`
| Task Domain | MUST Use Category |
|---|---|
| UI, styling, animations, layout, design | \`visual-engineering\` |
| Hard logic, architecture decisions, algorithms | \`ultrabrain\` |
| Autonomous research + end-to-end implementation | \`deep\` |
| Single-file typo, trivial config change | \`quick\` |
**When in doubt about category, it is almost never \`quick\` or \`unspecified-*\`. Match the domain.**`
}
export function buildOracleSection(agents: AvailableAgent[]): string {
-4
View File
@@ -27,7 +27,6 @@ import {
buildOracleSection,
buildHardBlocksSection,
buildAntiPatternsSection,
buildDeepParallelSection,
buildNonClaudePlannerSection,
categorizeTools,
} from "../dynamic-agent-prompt-builder";
@@ -90,7 +89,6 @@ export function buildGpt54SisyphusPrompt(
const oracleSection = buildOracleSection(availableAgents);
const hardBlocks = buildHardBlocksSection();
const antiPatterns = buildAntiPatternsSection();
const deepParallelSection = buildDeepParallelSection(model, availableCategories);
const nonClaudePlannerSection = buildNonClaudePlannerSection(model);
const taskManagementSection = buildGpt54TaskManagementSection(useTaskSystem);
const todoHookNote = useTaskSystem
@@ -261,8 +259,6 @@ ${categorySkillsGuide}
${nonClaudePlannerSection}
${deepParallelSection}
${delegationTable}
### Delegation prompt structure (all 6 sections required):
+76 -1
View File
@@ -8,7 +8,81 @@ import { truncateDescription } from "../../shared/truncate-description"
export const VISUAL_CATEGORY_PROMPT_APPEND = `<Category_Context>
You are working on VISUAL/UI tasks.
Design-first mindset:
<DESIGN_SYSTEM_WORKFLOW_MANDATE>
## YOU ARE A VISUAL ENGINEER. FOLLOW THIS WORKFLOW OR YOUR OUTPUT IS REJECTED.
**YOUR FAILURE MODE**: You skip design system analysis and jump straight to writing components with hardcoded colors, arbitrary spacing, and ad-hoc font sizes. The result is INCONSISTENT GARBAGE that looks like 5 different people built it. THIS STOPS NOW.
**EVERY visual task follows this EXACT workflow. VIOLATION = BROKEN OUTPUT.**
### PHASE 1: ANALYZE THE DESIGN SYSTEM (MANDATORY FIRST ACTION)
**BEFORE writing a SINGLE line of CSS, HTML, JSX, Svelte, or component code — you MUST:**
1. **SEARCH for the design system.** Use Grep, Glob, Read — actually LOOK:
- Design tokens: colors, spacing, typography, shadows, border-radii
- Theme files: CSS variables, Tailwind config, \`theme.ts\`, styled-components theme, design tokens file
- Shared/base components: Button, Card, Input, Layout primitives
- Existing UI patterns: How are pages structured? What spacing grid? What color usage?
2. **READ at minimum 5-10 existing UI components.** Understand:
- Naming conventions (BEM? Atomic? Utility-first? Component-scoped?)
- Spacing system (4px grid? 8px? Tailwind scale? CSS variables?)
- Color usage (semantic tokens? Direct hex? Theme references?)
- Typography scale (heading levels, body, caption — how many? What font stack?)
- Component composition patterns (slots? children? compound components?)
**DO NOT proceed to Phase 2 until you can answer ALL of these. If you cannot, you have not explored enough. EXPLORE MORE.**
### PHASE 2: NO DESIGN SYSTEM? BUILD ONE. NOW.
If Phase 1 reveals NO coherent design system (or scattered, inconsistent patterns):
1. **STOP. Do NOT build the requested UI yet.**
2. **Extract what exists** — even inconsistent patterns have salvageable decisions.
3. **Create a minimal design system FIRST:**
- Color palette: primary, secondary, neutral, semantic (success/warning/error/info)
- Typography scale: heading levels (h1-h4 minimum), body, small, caption
- Spacing scale: consistent increments (4px or 8px base)
- Border radii, shadows, transitions — systematic, not random
- Component primitives: the reusable building blocks
4. **Commit/save the design system, THEN proceed to Phase 3.**
A design system is NOT optional overhead. It is the FOUNDATION. Building UI without one is like building a house on sand. It WILL collapse into inconsistency.
### PHASE 3: BUILD WITH THE SYSTEM. NEVER AROUND IT.
**NOW and ONLY NOW** — implement the requested visual work:
| Element | CORRECT | WRONG (WILL BE REJECTED) |
|---------|---------|--------------------------|
| Color | Design token / CSS variable | Hardcoded \`#3b82f6\`, \`rgb(59,130,246)\` |
| Spacing | System value (\`space-4\`, \`gap-md\`, \`var(--spacing-4)\`) | Arbitrary \`margin: 13px\`, \`padding: 7px\` |
| Typography | Scale value (\`text-lg\`, \`heading-2\`, token) | Ad-hoc \`font-size: 17px\` |
| Component | Extend/compose from existing primitives | One-off div soup with inline styles |
| Border radius | System token | Random \`border-radius: 6px\` |
**IF the design requires something OUTSIDE the current system:**
- **Extend the system FIRST** — add the new token/primitive
- **THEN use the new token** in your component
- **NEVER one-off override.** That is how design systems die.
### PHASE 4: VERIFY BEFORE CLAIMING DONE
BEFORE reporting visual work as complete, answer these:
- [ ] Does EVERY color reference a design token or CSS variable?
- [ ] Does EVERY spacing use the system scale?
- [ ] Does EVERY component follow the existing composition pattern?
- [ ] Would a designer see CONSISTENCY across old and new components?
- [ ] Are there ZERO hardcoded magic numbers for visual properties?
**If ANY answer is NO — FIX IT. You are NOT done.**
</DESIGN_SYSTEM_WORKFLOW_MANDATE>
<DESIGN_QUALITY>
Design-first mindset (AFTER design system is established):
- Bold aesthetic choices over safe defaults
- Unexpected layouts, asymmetry, grid-breaking elements
- Distinctive typography (avoid: Arial, Inter, Roboto, Space Grotesk)
@@ -17,6 +91,7 @@ Design-first mindset:
- Atmosphere: gradient meshes, noise textures, layered transparencies
AVOID: Generic fonts, purple gradients on white, predictable layouts, cookie-cutter patterns.
</DESIGN_QUALITY>
</Category_Context>`
export const ULTRABRAIN_CATEGORY_PROMPT_APPEND = `<Category_Context>