undefined

undefined

undefined

---

ppt-image-first/
├─ SKILL.md                          # Agent skill definition
├─ references/
│  ├─ workflow.md                    # Full stage-by-stage workflow spec
│  ├─ conversation_framework.md      # Intake + confirmation dialogue rules
│  └─ preview-flow.md                # Image preview generation logic
├─ templates/
│  ├─ content_report_reference.md    # Template: content baseline doc
│  ├─ design_spec_reference.md       # Template: visual design spec
│  ├─ slide_blueprint_reference.md   # Template: per-page blueprint
│  └─ spec_lock_reference.md         # Template: execution constraints
└─ assets/
   ├─ preview_shell/index.html       # Style comparison UI shell
   ├─ candidate_picker_shell/index.html  # Multi-candidate selection UI
   └─ review_shell/index.html        # Review & retouch UI shell

INTAKE_FIELDS = [
    "purpose",          # defense / product pitch / research report / training
    "audience",         # professor panel / investors / internal team
    "page_count_hint",  # rough number or duration ("20 slides" / "10 min talk")
    "materials",        # what the user already has
    "identity_anchor",  # school / company / lab / brand name
]

undefined

undefined

1. Overall tone: light / dark / neutral middle?
2. Direction: conventional professional / visually distinctive?
3. How many style directions to preview first? (recommend 2–3)

PREVIEW_PAGES_PER_DIRECTION = [
    "cover_page",     # Title + identity anchor
    "toc_page",       # Table of contents / agenda
    "body_page",      # Representative content page
]

STYLE_INFERENCE_CATEGORIES = {
    "must_continue":    [],  # Traits clearly present, clearly liked
    "confirm_extend":   [],  # Traits that worked here, check if wanted deck-wide
    "do_not_lock":      [],  # Accidental/contextual traits, not repeatable rules
}

PLANNING_ARTIFACTS = [
    "design_spec.md",      # Global visual rationale + continuity constraints
    "slide_blueprint.md",  # Per-page: intent, content payload, visual strategy
    "spec_lock.md",        # What CAN and CANNOT change during generation
]

Generate mode:
A) One final image per slide (faster)
B) Multiple candidates per slide, then pick (slower, more control)

REVIEW_FEEDBACK_SCHEMA = {
    "slide_index": int,
    "issue_type": "visual | content | layout | consistency",
    "description": str,
    "suggested_fix": str,  # optional
}

import openai
import base64
from pathlib import Path

client = openai.OpenAI()  # reads OPENAI_API_KEY from env

def generate_slide_image(
    prompt: str,
    slide_index: int,
    output_dir: str = "./output/slides",
    size: str = "1792x1024",  # 16:9 approximation
) -> Path:
    """Generate a single slide image using GPT Image 2."""
    response = client.images.generate(
        model="gpt-image-1",
        prompt=prompt,
        n=1,
        size=size,
    )
    
    image_b64 = response.data[0].b64_json
    image_bytes = base64.b64decode(image_b64)
    
    out_path = Path(output_dir)
    out_path.mkdir(parents=True, exist_ok=True)
    
    slide_path = out_path / f"slide_{slide_index:02d}.png"
    slide_path.write_bytes(image_bytes)
    
    print(f"[slide {slide_index}] saved → {slide_path}")
    return slide_path

from pptx import Presentation
from pptx.util import Inches, Pt
from pathlib import Path

def build_pptx_from_images(
    image_paths: list[Path],
    output_path: str = "./output/deck.pptx",
    width_inches: float = 13.33,   # 16:9 widescreen
    height_inches: float = 7.5,
) -> Path:
    """Package a list of full-page slide images into a PPTX file."""
    prs = Presentation()
    prs.slide_width = Inches(width_inches)
    prs.slide_height = Inches(height_inches)
    
    blank_layout = prs.slide_layouts[6]  # blank layout — no placeholders
    
    for idx, img_path in enumerate(image_paths):
        slide = prs.slides.add_slide(blank_layout)
        slide.shapes.add_picture(
            str(img_path),
            left=Inches(0),
            top=Inches(0),
            width=Inches(width_inches),
            height=Inches(height_inches),
        )
        print(f"[pptx] added slide {idx + 1}: {img_path.name}")
    
    out = Path(output_path)
    out.parent.mkdir(parents=True, exist_ok=True)
    prs.save(str(out))
    print(f"[pptx] saved → {out}")
    return out

import os
from pathlib import Path

def run_generation_pipeline(
    slide_prompts: list[str],
    deck_title: str = "deck",
    output_dir: str = "./output",
) -> Path:
    """
    Given a list of per-slide prompts, generate images and package into PPTX.
    slide_prompts should come from slide_blueprint.md — one prompt per page.
    """
    slides_dir = Path(output_dir) / "slides"
    image_paths = []
    
    for i, prompt in enumerate(slide_prompts):
        path = generate_slide_image(
            prompt=prompt,
            slide_index=i + 1,
            output_dir=str(slides_dir),
        )
        image_paths.append(path)
    
    pptx_path = build_pptx_from_images(
        image_paths=image_paths,
        output_path=f"{output_dir}/{deck_title}.pptx",
    )
    return pptx_path

    "Table of contents slide. Sections: 1. Background 2. Data & Methods 3. Results 4. Discussion 5. Conclusion. "
    "Same dark navy color scheme, numbered list with clear hierarchy, minimal decorative elements.",
    
    "Body slide: 'Key Findings'. Three main data points as large stat callouts: +2.3°C average temp increase, "
    "67% of monitored stations affected, 15-year trend data. Dark navy background, accent color teal, "
    "clean data-forward layout.",
]

output = run_generation_pipeline(
    slide_prompts=prompts,
    deck_title="meteorology-defense",
    output_dir="./output",
)
print(f"Done: {output}")

undefined

def generate_slide_candidates(
    prompt: str,
    slide_index: int,
    n_candidates: int = 3,
    output_dir: str = "./output/candidates",
) -> list[Path]:
    """Generate N candidate images for one slide for user selection."""
    paths = []
    for c in range(n_candidates):
        path = generate_slide_image(
            prompt=prompt,
            slide_index=slide_index,
            output_dir=f"{output_dir}/slide_{slide_index:02d}",
        )
        # rename to include candidate index
        new_path = path.parent / f"candidate_{c + 1}.png"
        path.rename(new_path)
        paths.append(new_path)
        print(f"[candidate {c + 1}/{n_candidates}] slide {slide_index}")
    return paths

undefined

undefined

undefined

undefined

undefined

---

undefined

undefined

import subprocess
import json
from pathlib import Path

def launch_preview_shell(preview_images: dict[str, list[Path]]) -> None:
    """
    Write preview manifest and open the preview shell in browser.
    preview_images: {"direction_A": [cover, toc, body], "direction_B": [...]}
    """
    manifest = {
        direction: [str(p) for p in pages]
        for direction, pages in preview_images.items()
    }
    
    shell_dir = Path("assets/preview_shell")
    manifest_path = shell_dir / "preview_manifest.json"
    manifest_path.write_text(json.dumps(manifest, indent=2))
    
    # Open in default browser
    subprocess.run(["open", str(shell_dir / "index.html")])  # macOS
    # subprocess.run(["xdg-open", str(shell_dir / "index.html")])  # Linux

def feedback_to_retouch_prompt(
    original_prompt: str,
    feedback: dict,
) -> str:
    """Convert structured review feedback into an updated generation prompt."""
    issue = feedback["description"]
    fix = feedback.get("suggested_fix", "")
    
    retouch_instruction = f"REVISION: {issue}"
    if fix:
        retouch_instruction += f" Fix: {fix}"
    
    return f"{original_prompt}\n\n{retouch_instruction}"

undefined

undefined

undefined

undefined

Root cause: prompt drift — each slide prompt diverges from the locked spec.
Fix:
  1. Prepend spec_lock.md's "Locked" section to EVERY slide prompt
  2. Use a prompt prefix template:

STYLE_PREFIX = """
[STYLE LOCK] Dark navy #0A1628 background. Heading font: Inter Bold.
Logo bottom-left. Slide number bottom-right. Teal accent #2DD4BF on highlights.
DO NOT add gradients, textures, or decorative borders not in this spec.
"""

full_prompt = STYLE_PREFIX + slide_specific_prompt

This is a workflow sequencing error.
content_report.md is a SOURCE document, not a script.
The slide_blueprint.md should ADAPT content into slide-appropriate payloads.
Each slide blueprint entry must go through:
  content_report → narrative selection → slide payload → generation prompt
Never pipe content_report text directly into image generation prompts.

undefined

---