Slidewright

Many companies still rely on PowerPoint presentations - slidewright just allows a frictionless design-to-ppt flow with easy editability and fewer conflicts.

Comment

Inspiration

I have used PowerPoint throughout consulting and startup work. AI could already turn an idea into a convincing slide image, but the inherited PPTX was often the real problem: uneven margins, clipped copy, silent font changes, flattened visuals, and layouts that collapsed after the first edit. Slidewright began with a simple question: can Codex generate the editable artifact with the same discipline we expect from production code?

What it does

Slidewright is an open-source Codex skill and plugin that turns an idea, brief, visual reference, or controlled template edit into native, editable PowerPoint.

  • Visible words remain native text with run-level bold, italic, color, and size.
  • Shapes, groups, charts, tables, connectors, notes, and reading order remain semantic objects.
  • Existing decks retain their fonts, colors, masters, layouts, placeholders, logos, spacing, and recurring chrome.
  • Margins and padding are symmetric by default; typography uses conventional whole-point sizes.
  • Copy is shortened or relaid out before shrinking below the quality floor.
  • Missing fonts, overflow, clipping, undeclared overlap, and package drift stop the build with an actionable error.
  • An optional executive-review mode adds native yellow partner comments to a separate review copy without changing the canonical deck.

How I built it

Codex was the primary engineering environment and implementation collaborator. I converted observed presentation failures and public user complaints into a binary goal ledger, then built a deterministic pipeline around them:

  1. A semantic specification captures the communication job, content hierarchy, design tokens, and source-deck identity.
  2. A compiler resolves grids, symmetric rims, spacing, reading order, and discrete whole-point text fitting.
  3. Native PowerPoint text, rich-text runs, shapes, tables, charts, groups, connectors, notes, and template relationships are emitted or preserved.
  4. Linting, font audits, OOXML inspection, visual comparison, PowerPoint save/reopen tests, destructive controls, and full-size rendered review decide whether the artifact can ship.
  5. Public CI reproduces the portable evidence on clean Windows, macOS, and Linux hosts.

Human decisions set the product thesis and non-negotiable quality bar: editability over screenshots, formatting integrity over superficial polish, and no binary goal credit without its named proof.

Challenges I ran into

PowerPoint is both a visual canvas and a document package. A deck can look correct while its reading order, relationships, chart workbook, group hierarchy, theme, or font semantics are broken. The hardest work was making those hidden properties testable without allowing the verifier to trust the generator's own claims.

Text fitting was another major challenge. Approximate font metrics, inconsistent insets, and arbitrary fractional sizes create slides that pass a screenshot check but fail after one edit. Slidewright instead uses explicit content budgets, conventional integer type tokens, and fail-closed rendered checks.

Finally, real PowerPoint reserializes OOXML. I added owned, time-bounded PowerPoint automation and compared the source, edited deck, save cycles, semantic inventories, and full-slide renders rather than assuming a ZIP-valid PPTX was safe.

Accomplishments that I am proud of after this week

  • 318/318 release tests and 15/15 required destructive-control tests.
  • 26/26 repair-free PowerPoint fixtures plus 13/13 repair controls.
  • Four licensed template families, 39 slides, 542 artifact receipts, and 195 hash-bound full-size reviews.
  • Native font embedding and two real PowerPoint save/reopen cycles with exact visible-style retention.
  • Lossless structural ingestion for four licensed decks covering master/layout/theme hierarchy, text runs, tables, charts, diagrams, notes, and recursive reading order.
  • 11/11 public CI jobs passing on the exact release commit.
  • 54 of 58 deliberately strict product and complaint-derived goals currently proven.

What I learned

Presentation generation is not primarily an image-generation problem. It is a compilation and document-integrity problem. The useful output is the object someone else can select, edit, restyle, regroup, and reopen tomorrow. The most important part was to give it a structure, a roadmap to follow in deconstructing and construting slides.

I also learned that visual QA needs adversarial evidence. A score is meaningful only if a deliberately broken deck fails it, and a template-preservation claim is meaningful only when unauthorized changes are independently detected.

What's next for Slidewright

The remaining binary goals are intentionally public: actual client-originated proof in Codex Desktop and VS Code, blind review by presentation experts and target users, and a six-suite interoperability matrix covering PowerPoint on Windows and macOS, Google Slides, Keynote, LibreOffice, and Canva. After those proofs, the next product layer is broader source-deck editing and a richer library of licensed page archetypes.

Try it

Built With

Share this project:

Updates