القائمة
This skill should be used when generating images for educational content. It provides multi-turn reasoning partnership methodology with Gemini for professional quality visual generation.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | image-generator |
| description | This skill should be used when generating images for educational content. It provides multi-turn reasoning partnership methodology with Gemini for professional quality visual generation. |
| version | 1.0.0 |
Visual generation converges toward accepting first output ("looks good enough") and following technical specifications rigidly. This produces generic aesthetics and misses Gemini 3's reasoning capabilities.
This skill provides multi-turn reasoning partnership methodology with professional quality standards.
Receive from visual-asset-workflow v5.0:
## The Story
[Narrative about what's visualized]
## Emotional Intent
[What it should FEEL like]
## Visual Metaphor
[Universal concept for instant comprehension]
## Subject / Composition / Action / Location / Style / Camera / Lighting
[Official Gemini 3 prompt structure]
## Color Semantics
Blue (#2563eb) = Authority (teaches governance)
Green (#10b981) = Execution (teaches work)
## Typography Hierarchy
Largest: Key insight (information importance drives sizing)
Medium: Supporting components
Smallest: Context
## Pedagogical Reasoning
[Why these choices serve teaching]
DO NOT convert briefs back to pixel specs - Use AS-IS to activate reasoning.
Principle: Creative briefs activate Gemini's reasoning about HOW to achieve intent visually
CRITICAL: Use gemini.google.com ONLY (NOT Google AI Studio, NOT other image generators)
Initialize once:
For EACH visual:
.playwright-mcp/Gemini-Generated-Image-*.png to robolearn-interface/static/img/part-{N}/chapter-{NN}/{filename}.pngPrinciple: New chat per visual prevents cross-contamination; immediate verification catches issues early; immediate embedding prevents orphans
ALL must pass before download:
Gate 1: Spelling Accuracy (99% standard)
Gate 2: Layout Precision
Gate 3: Color Accuracy
Gate 4: Typography Hierarchy
Gate 5: Teaching Effectiveness (<5 sec concept grasp)
Gate 6: Uniqueness Validation (NEW)
{filename}.prompt.md exists and visual matches itDecision:
Principle: Explicit criteria prevent "good enough" false positives; uniqueness check prevents duplicate rework
When: Batch mode with >8 visuals OR continuation session
Condensation strategy:
Example condensation:
ORIGINAL (250 tokens):
"Top Layer shows the Coordinator at center top with label 'Orchestrator'
featuring a conductor icon, with the role description 'Strategic oversight,
contract validation', rendered in Gold color (#fbbf24) as a Large hexagon..."
CONDENSED (80 tokens):
"Top Layer - Coordinator: Center top: 'Orchestrator' (conductor icon),
Role: 'Strategic oversight, contract validation', Gold (#fbbf24), Large hexagon."
Success metric: 60-70% token reduction while maintaining 100% first-attempt success rate
Principle: Efficiency without sacrificing reasoning activation
After copying image to destination, BEFORE starting next visual:
Determine lesson file:
Chapter and Lesson metadatarobolearn-interface/docs/[part]/[chapter]/[lesson-file].mdFind insertion point:
Insert reference:
Verify no code block interruption:
Why this matters: Completes the work immediately; prevents orphan images
Result: Each visual is generated → validated → placed → verified before moving to next
Principle: Immediate embedding prevents disconnect between generation and integration
Avoid: Accepting first output without evaluation
Prefer: Teaching Gemini your standards through iteration
Iteration Pattern:
Turn 1: Initial Generation
Turn 2: Principle-Based Feedback (if gates fail)
Gate 4 FAILED: Typography hierarchy incorrect
The largest text is "$100K" (supporting detail) but should be "$3T"
(key insight students must grasp).
Pedagogical reasoning: Information importance drives sizing. $3T is
the insight, not the starting value. Visual weight teaches what matters.
Increase '$3T' to dominant size (largest element). Reduce '$100K' to
supporting size. This teaches magnitude through visual hierarchy.
Turn 3: Validation
Principle: You teach Gemini (principle-based feedback), Gemini teaches you (reveals understanding), Co-evolve toward quality
Why it matters: Gemini learns your pedagogical standards across iterations
Avoid: Permission-asking between visuals
Prefer: Autonomous batch execution
When invoked with: "generate all visuals" or "batch generate"
Execute WITHOUT STOPPING:
For EACH visual in approved list:
A. NEW CHAT (context isolation)
B. Generate image (paste condensed creative brief)
C. Verify quality (6 gates including uniqueness)
D. Iterate if needed (max 3 tries)
E. Download when pass (organized directory)
F. Embed in lesson immediately (no orphans)
G. Log progress ("✅ Generated N/M")
H. IMMEDIATELY next visual (NO STOPPING)
NEVER ask:
Only report summary at END:
BATCH COMPLETE
Total: 18 visuals
✅ Generated: 16 (2K, avg 2-3 iterations)
⚠️ Deferred: 2 (quality issues after 3 tries)
Time: ~45 min
Location: robolearn-interface/static/img/part-{N}/chapter-{NN}/
Principle: Autonomous execution without interruption = efficient batch processing
From visual-asset-workflow, enforce during generation:
A2 Beginner Limits:
B1 Intermediate:
C2 Professional:
Validation during generation: "Does this visual's complexity match proficiency from creative brief?"
Principle: Complexity mismatch = pedagogical failure
AFTER completing batch, analyze systematically:
Success patterns:
Failure analysis:
Improvement opportunities:
Output: history/visual-assets/reflections/chapter-{NN}-reflection.md
Principle: Systematic reflection → Continuous improvement
If session ends mid-batch (token limit, context overflow):
Create checkpoint file: history/visual-assets/checkpoints/part-{N}-checkpoint.md
## Batch Status: Part {N}
**Date:** 2025-11-24
**Status:** INTERRUPTED at {X}/{Y} images
### Completed:
- ✅ Image 1: {filename} (2 iterations, embedded in lesson-01.md)
- ✅ Image 2: {filename} (1 iteration, embedded in lesson-02.md)
...
### Remaining:
- ⏳ Image 8: {filename} (not started)
- ⏳ Image 9: {filename} (not started)
...
### Quality Stats (so far):
- Success rate: {X/Y} production-ready
- Avg iterations: {N}
- Gate failures: Gate 1: {n}, Gate 2: {n}...
### Continuation Instructions:
1. Read this checkpoint
2. Start at Image {next}
3. Continue autonomous batch mode
4. Update checkpoint after each image
On continuation:
Principle: Seamless recovery from interruptions maintains momentum
Never:
visuals/ directory (use part/chapter organization)Even if it seems reasonable:
You tend to accept visuals after 1 iteration even with minor issues. Push for quality:
Professional content creators iterate. You should too.
You'll know this skill is working when:
Result: Professional-quality visuals with distinctive aesthetics, generated autonomously with systematic quality control, embedded immediately, and recoverable from interruptions.
Produces a complete structural analysis of the IntelliStack codebase — modules, API surface, database schema, auth flows, and RAG pipeline — as a human-readable export map that any developer can use as a migration blueprint regardless of their target framework.
Migrates the IntelliStack Better-Auth OIDC authentication system to any target auth framework, preserving OAuth flows, session management, role/permission model, and JWT claim structure.
Migrates IntelliStack FastAPI routers, services, and Pydantic schemas to any target backend framework while preserving every business rule, auth guard, and error contract.
Converts IntelliStack's SQLAlchemy 2.0 models and Alembic migrations to any target ORM or database toolkit, preserving the full schema, indexes, relationships, and cascade rules.
Generates typed API clients and frontend integration code for any framework consuming the IntelliStack REST API, including auth flows, learning path UI, chatbot UI, and admin dashboards.
Re-implements the IntelliStack RAG pipeline (Qdrant + Cohere + OpenAI SSE) in any target language or framework, preserving hybrid retrieval, reranking, citation architecture, and stage-based access control.