cd /news/artificial-intelligence/orxhestra-config-for-slide-making-lo… Β· home β€Ί topics β€Ί artificial-intelligence β€Ί article
[ARTICLE Β· art-28432] src=gist.github.com β†— pub= topic=artificial-intelligence verified=true sentiment=Β· neutral

Orxhestra config for slide making locally. Requires a skill at ./skills/ for slidemaking

A developer created an automated slide-making pipeline using Orxhestra, a local LLM agent framework. The pipeline uses a llama.cpp server with the Qwen3.5-35B-A3B model to plan, implement, screenshot, and review PowerPoint slides in a loop, following Implement Consulting Group's design system.

read7 min views1 publishedJun 15, 2026

| # Deck generation pipeline: plan β†’ implement β†’ review loop. | | | # | | | # Run: orx orx.yaml | | | # | | | # Flow: | | | # 1. planner β€” reads the brief, produces a slide-by-slide plan | | | # 2. implementer β€” builds the PPTX using python-pptx + implement-design skill | | | # 3. screenshotter β€” exports slides to PNG via LibreOffice / pptx2png | | | # 4. reviewer β€” looks at the PNGs and approves or requests changes | | | # Steps 2-4 loop until the reviewer is satisfied (max 4 iterations). | | | # Uses a llama.cp server: launch it with | | | # llama-server -hf bartowski/Qwen_Qwen3.5-35B-A3B-GGUF --reasoning-budget 1 | | | version: "0.0.21" | | | defaults: | | | model: | | | provider: openai | |

| name: bartowski/Qwen_Qwen3.5-35B-A3B-GGUF | |
| base_url: http://localhost:8080/v1 | |

| tools: | | | exit: | | | builtin: "exit_loop" | | | filesystem: | | | builtin: "filesystem" | | | shell: | | | builtin: "shell" | | | artifacts: | | | builtin: "artifacts" | | | todos: | | | builtin: "write_todos" | | | human_input: | | | builtin: "human_input" | | | skills: | | | implement_design: | |

| name: implement-design | |
| description: > | |

| Implement Consulting Group's complete design system. Covers brand colors, | | | typography (Palatino Linotype + Arial), logo placement, icon library, | | | slide layouts, and the five non-negotiables. Use for any PPTX or visual | | | output that must follow Implement's brand identity. | | | directory: "skills/implement-design" | | | agents: | | | # ── 1. Planner ────────────────────────────────────────────────────────────── | | | planner: | | | type: llm | | | description: "Reads the deck brief and produces a detailed slide-by-slide plan." | | | instructions: | | | | You are a presentation strategist. Your output is consumed by a coding | | | agent β€” NOT a human β€” so be precise and actionable. | | | Read the user's brief carefully. Then produce a numbered slide plan with: | | | - Slide number | |

| - Layout type (title, section divider, content, two-column, full-bleed, etc.) | |
| - Headline (exact text) | |

| - Body content (bullet points, data, quotes β€” exact text) | | | - Visual notes (icon suggestions, color emphasis, illustration placement) | | | Follow Implement Consulting Group's brand principles: | | | - Calm, insightful, professional β€” less is more | | | - Default to Blue-Green (#67817F) + Green-Grey (#B9C7C2) | | | - Generous white space; no layered design | | | - Palatino Linotype only as accent font (never bold, never ALL CAPS) | | | Output ONLY the slide plan. Do not write code. | | | tools: | | | - human_input | | | - artifacts | | | max_iterations: 10 | | | # ── 2. Implementer ────────────────────────────────────────────────────────── | | | implementer: | | | type: llm | | | description: "Builds the PPTX from the slide plan using the Implement design skill." | | | instructions: | | | | You are a presentation builder. You receive a slide plan and must produce | | | a complete, brand-correct PPTX file at output/deck.pptx. | | | ## Workflow | | | 1. Call list_skills β€” confirm the implement-design skill is available. | | | 2. Call load_skill for implement-design to read brand guidelines and | | | load the domain reference for PPTX (references/pptx.md). | | | 3. Call load_skill_resource for references/pptx.md to get the full | | | PPTX workflow, composition patterns, and python-pptx code examples. | | | 4. Use read_file to inspect skills/implement-design/scripts/im_slide.py | | | and the core slide template at skills/implement-design/assets/core_slide.pptx. | | | 5. Write a Python build script (output/build_deck.py) that: | | | - Uses python-pptx and im_slide.py helpers | | | - Starts from the core_slide.pptx template | | | - Implements every slide from the plan with correct layout, typography, | | | colors, and logo placement | | | - Saves the result to output/deck.pptx | | | 6. Run the build script with the shell tool. | | | 7. Verify output/deck.pptx exists; read it back to confirm non-zero size. | | | ## Brand non-negotiables (from Implement design system) | | | - Tone-in-tone color β€” one main color (max two) and its nuances | | | - No layered design β€” no element stacked on another | | | - Palatino Linotype is an ACCENT font only β€” never bold, never colored | | | - Icons only when they add genuine value; no fill colors on icons | | | - White space is a feature β€” resist the urge to fill every area | | | Never ask the user to do anything manually. Execute every step yourself. | | | tools: | | | - filesystem | | | - shell | | | - artifacts | | | - todos | | | skills: | | | - implement_design | | | max_iterations: 20 | | | # ── 3. Screenshotter ──────────────────────────────────────────────────────── | | | screenshotter: | | | type: llm | | | description: "Exports the PPTX slides to PNG files for visual review." | | | instructions: | | | | You are a screenshot agent. Convert output/deck.pptx to PNG images | | | so the reviewer agent can inspect each slide visually. | | | ## Steps | | | 1. Create the directory output/screenshots/ if it does not exist. | | | 2. Try exporting via LibreOffice headless: | | | shell_exec: libreoffice --headless --convert-to png --outdir output/screenshots/ output/deck.pptx | | | If that succeeds, rename files to slide_01.png, slide_02.png, etc. | | | 3. If LibreOffice is unavailable, try python-pptx + Pillow: | | | - Read each slide, render its thumbnail, save as output/screenshots/slide_NN.png | | | 4. List output/screenshots/ and confirm PNG files are present. | | | 5. Save the list of screenshot paths as an artifact named screenshot_paths. | | | Report success or failure clearly so the reviewer knows what to look at. | | | tools: | | | - filesystem | | | - shell | | | - artifacts | | | max_iterations: 10 | | | # ── 4. Reviewer ───────────────────────────────────────────────────────────── | | | reviewer: | | | type: llm | | | description: "Reviews slide screenshots and approves or requests targeted fixes." | | | instructions: | | | | You are a design reviewer for Implement Consulting Group presentations. | | | You will receive PNG screenshots of each slide. Evaluate them against | | | the Implement brand standard: | | | ## What to check | | | - Layout: No layered elements. Generous white space. Clean alignment. | | | - Color: Tone-in-tone. Primarily Blue-Green (#67817F) / Green-Grey (#B9C7C2). | | | No unauthorized accent colors. | | | - Typography: Palatino Linotype used sparingly as accent only (never bold, | | | never ALL CAPS, never colored). Arial for all body/headline text. | | | - Logo: IM_ logo present in correct corner position on title/section slides. | | | - Icons: Only present when adding clear value; no fill colors. | | | - Content: Slide headlines match the plan. No text overflow. | | | ## Decision | | | - If ALL slides pass: call exit_loop to finish. | | | - If issues exist: output a numbered list of specific, actionable fixes for | | | the implementer. Be precise β€” include slide number, what is wrong, and | | | what the correct behavior should be. | | | Load screenshots with load_artifact using the screenshot_paths artifact. | | | You MUST read the actual PNG files with read_file to visually inspect them. | | | tools: | | | - exit | | | - filesystem | | | - artifacts | | | max_iterations: 10 | | | # ── Loops & pipeline ──────────────────────────────────────────────────────── | | | build_review_loop: | | | type: loop | | | description: "Implement β†’ screenshot β†’ review loop until approved." | | | agents: [implementer, screenshotter, reviewer] | | | max_iterations: 8 | | | deck_pipeline: | | | type: sequential | | | description: "Full deck generation: plan first, then the build-review loop." | | | agents: [planner, build_review_loop] | | | main_agent: deck_pipeline | | | runner: | | | app_name: deck-generator | | | session_service: memory | | | artifact_service: | | | type: local | | | path: artifacts/ |

── more in #artificial-intelligence 4 stories Β· sorted by recency
── more on @orxhestra 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain β€” perfect for shipping the agent you just read about.

$git push zahid main
β†’ Live at https://your-agent.zahid.host βœ“
Get free account β†’ Pricing
from €0/mo Β· no card required
LIVE [news/orxhestra-config-for…] indexed:0 read:7min 2026-06-15 Β· β€”