feat: Auto-generate narrative presentation for every Sundai project using Narrative Engine

[natea]

Goal

Every Sundai.club project page should link to a compelling slide presentation that tells the story of why the app exists — the problem, the insight, the solution, and the hero's journey of the user. Not a feature list. A narrative that makes someone care about the project in 2 minutes.

Narrative Engine is a Claude Code skill that transforms content into narrative-driven presentations using proven storytelling frameworks. It could be integrated into the sundai-project-pipeline to auto-generate this for every project.

The problem

Today's Sundai project pages have:

• A title and one-line description
• A "Full Description" (usually a feature list or technical summary)
• A demo URL
• A GitHub link

What they're missing: the story. Why does this app exist? What pain did someone feel? What's the "aha" moment? Why should I click the demo link? The best Sundai demos at meetups tell a story in 5 minutes — but the project page captures none of that narrative.

What Narrative Engine does

A Claude Code skill (21 stars, by @nraford7) that runs a 4-phase pipeline:

Phase 1: Discovery

Interactive questions about audience, purpose, tone, and content type. For Sundai projects, we'd auto-fill most of this from the OpenSpec proposal:

• Audience: AI builders, technical but curious
• Purpose: Explain why this exists and make them want to try it
• Tone: Derived from project personality
• Content type: Slide presentation

Phase 2: Framework Matching

Scores content against 17 narrative/communication frameworks and picks the best fit:

Narrative Arcs	Communication Frameworks
Prestige (transformation)	Pyramid (conclusion-first)
Mystery Box (curiosity-driven)	SCQA (situation-complication-question-answer)
Heist (assembling the plan)	AIDA (attention-interest-desire-action)
Time Machine (before/after)	PAS (problem-agitate-solve)
+ 6 more arcs	+ 3 more frameworks

For Sundai projects, likely top matches:

• Time Machine — "Before this app: pain. After: delight."
• PAS — "Here's the problem. Here's why it's worse than you think. Here's the fix."
• Mystery Box — "What if you could... [demo reveal]"

Phase 3: Build Brief & Generation

A subagent generates the actual slides with:

• Auto-derived voice matched to audience + tone
• Emotional arc calibration — pacing for technical audiences
• Opening/closing strategies from a curated rhetorical library
• Source attribution tags: [DIRECT], [PARAPHRASE], [ELABORATED], [GENERATED] — shows what came from the user vs. AI

Phase 4: Review Panel

Six parallel subagents assess the output:

• Audience Advocate, Comms Specialist, Visual Designer, Critic, Content Expert, Originality Agent
• Catches generic AI patterns and ensures the narrative actually lands

Integration with the Sundai pipeline

New Step: between Steps 5 (Edit Details) and 6 (Required Defaults)

Step 5: Edit project details (GitHub URL, Demo URL, descriptions)
Step 5.5 [NEW]: Generate narrative presentation
Step 6: Required defaults (team, thumbnail)

Input to Narrative Engine (auto-extracted from OpenSpec artifacts):

• proposal.md → the problem, the "why", the insight
• design.md → technical decisions and tradeoffs (for the "how we built it" section)
• specs/*/spec.md → key capabilities and scenarios (for the "what it does" demo flow)
• Phase 1 ideation output → the pain points, market gaps, creative lens that generated the idea
• Voice interview transcript (if feat: Voice-based project ideation — interview users instead of text chat #6 is implemented) → the user's own words about the problem

Output:

• A hosted HTML slide deck (deployed alongside the app or as a static page)
• URL added to the Sundai project page as "Project Story" or "Why This Exists"
• Could also generate the Full Description text for the Sundai page itself

Example narrative for cook-with-ai

Slide 1 — The Hook (Mystery Box)
"You're staring into your fridge at 7pm. Chicken thighs, half a lemon, some garlic. You know there's a meal in there. But the 47 recipe sites you'd have to search don't know what's in YOUR fridge."

Slide 2 — The Pain
"Recipes assume you'll go shopping. Cooking blogs bury the instructions under 2000 words of backstory. And none of them adapt to what you actually have."

Slide 3 — The Insight
"What if you could just tell an AI what's in your kitchen and get back exactly what to do — step by step, with timing, with tips, with no assumptions?"

Slide 4 — The Solution (Demo)
[Screenshot of cook-with-ai in action]

Slide 5 — Try It
[Link to demo]

Relationship to other issues

• feat: Voice-based project ideation — interview users instead of text chat #6 (Voice ideation): The voice interview transcript is incredible input for Narrative Engine — the user's own words about the problem become the emotional core of the narrative
• feat: Visual brainstorming during ideation — wireframe mockups with Superpowers visual-companion #7 (Visual brainstorming): Wireframes could become slides in the presentation showing the "design journey"
• feat: Add Impeccable design skills to VM and integrate into sundai-project-pipeline #3 (Impeccable): Could polish the slide deck's visual design after generation

Open questions

• What format should the slides be? Options: HTML (self-hosted), Google Slides (via API), PDF, reveal.js
• Where to host? Alongside the demo on the same deploy, or a separate static site?
• Should the presentation be auto-generated or require a human review pass?
• How long? 5-8 slides feels right — enough to tell the story, short enough to hold attention
• Should the narrative also generate the Sundai Full Description text? (Two birds, one stone)
• Can we run the 6-agent review panel within our token budget, or should we trim it?

Tasks

• Install Narrative Engine skill into the pipeline workspace
• Create a Sundai-specific discovery profile (auto-fill audience, purpose, tone from OpenSpec)
• Build OpenSpec→Narrative Engine input pipeline (extract problem/solution/insight from artifacts)
• Choose and implement slide output format (recommend reveal.js for self-hosted HTML)
• Add Step 5.5 to sundai-project-pipeline — generate presentation after project details are set
• Deploy slides alongside the app (same Cloud Run service or Coolify instance — see explore: Alternative hosting to Cloud Run — Coolify, Kamal 2, and self-hosted options #5)
• Add "Project Story" link to Sundai project page via API
• Test end-to-end: OpenSpec artifacts → narrative → hosted slides → linked from Sundai page
• Optional: also generate Full Description text from the narrative for the Sundai page

References

• Narrative Engine: https://github.com/nraford7/Narrative-Engine
• Superpowers 5 blog post: https://blog.fsck.com/2026/03/09/superpowers-5/
• Author: @nraford7 (same author as Boswell in feat: Voice-based project ideation — interview users instead of text chat #6)

[Whatsonyourmind]

This is a great use case — narrative-driven project presentations are exactly the kind of thing that should be automated. A few architectural thoughts:

The Narrative Engine handles the content side well (storytelling structure, frameworks). The piece that's often missing is the rendering — turning that narrative into actual slides with proper layout, charts, and visual hierarchy.

One pattern that works well: Narrative Engine generates structured JSON (slide type, elements, content for each slide) → a rendering API turns it into .pptx with constraint-based layout. The rendering handles things like text overflow, chart embedding, and theme consistency automatically.

For the sundai-project-pipeline integration, the flow could be:

1. Pipeline triggers on new project creation
2. Narrative Engine extracts problem/insight/solution/journey from project README + metadata
3. Output is a JSON intermediate representation (slide types: title, problem statement, key insight, solution, demo, metrics, team, CTA)
4. Rendering API produces the .pptx with the selected theme

This keeps the narrative logic decoupled from the PPTX generation complexity. The MCP ecosystem has tools for step 3-4 that can be called from any agent pipeline.

[nraford7]

Hey Sundaiclaw/Hacker, exciting you're using this and thanks for the feedback. Unrelated, have you had much success with turning them into PPTX? I've taken the route of creating HTML templates, rendering to those, then exporting to PDF. Does your rendering API produce the PPTX well? All my experiments have been meh. Kind Regards, *Noah Raford* MANAGING PARTNER Level 1, Office 124, Building 9 Dubai Media City, Dubai U.A.E. P.O. Box 503014 Phone: +971 4 558 8197 EMIR WhatsApp: +971 52 149 9013 Email: ***@***.*** www.emirintelligence.com For complimentary insights, follow EMIR: <https://hubs.li/H0-mgCn0> <https://hubs.li/H0-mgF60> <https://hubs.li/H0-mgCs0> <https://hubs.li/H0-mgF_0> <https://hubs.li/H0_7PZ00> <https://hubs.li/H0_7PZ00>

…

On Fri, 3 Apr 2026 at 19:00, Whatsonyourmind ***@***.***> wrote: *Whatsonyourmind* left a comment (sundaiclaw/hacker#8) <#8 (comment)> This is a great use case — narrative-driven project presentations are exactly the kind of thing that should be automated. A few architectural thoughts: The Narrative Engine handles the *content* side well (storytelling structure, frameworks). The piece that's often missing is the *rendering* — turning that narrative into actual slides with proper layout, charts, and visual hierarchy. One pattern that works well: *Narrative Engine generates structured JSON* (slide type, elements, content for each slide) → *a rendering API turns it into .pptx* with constraint-based layout. The rendering handles things like text overflow, chart embedding, and theme consistency automatically. For the sundai-project-pipeline integration, the flow could be: 1. Pipeline triggers on new project creation 2. Narrative Engine extracts problem/insight/solution/journey from project README + metadata 3. Output is a JSON intermediate representation (slide types: title, problem statement, key insight, solution, demo, metrics, team, CTA) 4. Rendering API produces the .pptx with the selected theme This keeps the narrative logic decoupled from the PPTX generation complexity. The MCP ecosystem has tools for step 3-4 that can be called from any agent pipeline. — Reply to this email directly, view it on GitHub <#8?email_source=notifications&email_token=BQXFFPB647VTP44JAIJ6UQL4T6KVLA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJYGMYDANZXHAZ2M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4183007783>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BQXFFPGNHWTDTVJPLR2PF734T6KVLAVCNFSM6AAAAACXIHLX3GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCOBTGAYDONZYGM> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[Whatsonyourmind]

@nraford7 Great question — PPTX generation is one of those things that looks simple but has a lot of edge cases.

In my experience, the HTML-to-PDF route you are using is actually more reliable for visual fidelity. Native PPTX generation (via python-pptx or similar libraries) struggles with:

1. Text overflow — PPTX does not auto-resize text boxes the way HTML/CSS does. You need to pre-calculate line heights and truncate or resize fonts manually.
2. Chart embedding — Most PPTX libraries produce static images for charts rather than native PPTX chart objects, so you lose editability.
3. Theme consistency — Slide masters and layouts in PPTX are surprisingly fragile. Programmatically generated slides often drift from the intended theme.

The approach I have found works best is a hybrid:

• Generate the structured content as JSON (slide type, bullet points, chart data, speaker notes)
• Use a template-based PPTX engine (python-pptx with pre-designed .pptx templates that have placeholder shapes) rather than building slides from scratch
• For anything visually complex (charts, diagrams), render to SVG/PNG and embed as images

If your main output is PDF anyway, I would honestly stick with the HTML rendering path — it gives you far more control over layout and typography. PPTX is only worth the pain if your users specifically need to edit the slides afterward.