The plan, in public.

Package Photo Panel as a native-feeling desktop application installable from the Mac App Store, Microsoft Store, or a direct download. The most practical wrapper is Tauri (Rust + WebView) or Electron - both let the existing Next.js frontend run locally. A desktop app removes the Vercel body-size limit entirely (large panels with many high-resolution frames can be sent directly), enables local AI integration, allows offline use of the compositing and PDF tools, and gives users a permanent shortcut rather than a browser tab. The companion app scaffolded for item 26 uses Tauri and can be extended into a full desktop build.

Tailored prompts for monochrome, landscape, street, nature, portrait, sport, and macro panels. Each mode emphasises the criteria that matter for that genre.

Let users run critiques without consuming Claude API credits by routing requests to a local vision model. The most practical path is Ollama running a vision-capable model (LLaVA, Llama 3.2 Vision, or similar) as a local HTTP server - the app would detect whether a local endpoint is available and offer it as an alternative to the cloud path. On Mac, Apple's on-device Foundation Models API (available from macOS 26 / iOS 18 era) is a second option once the vision capability is stable enough for photography critique. Local inference is slower and less capable than Claude Opus, but removes the per-critique cost entirely and works offline. A settings toggle would let users choose: Cloud (Claude, best quality) or Local (free, private, slower).

Trying to put a panel together using free placement. Unable to move one that is a pano within the 5 images I have uploaded, to the panel.

Hi Joe I heard of Photo-Panel through the RPS DI group and have happily signed up, hopefully to help me with an L panel, once the 'new' distinction process is in place. It's a shame that I need to start with what seems such an easy question; how do I download the Lightroom Classic plug-in? Thanks, Graham

Alternative export format for users who prefer .docx for editing the critique. PDF remains the default.

Project bootstrap with TypeScript, Tailwind 4, quality scripts (version-sync, JSX whitespace), and project docs (CLAUDE.md, SECURITY.md, TODO.md, HANDOVER.md).

Brand tokens, Inter web font, Nav with HP logo, Footer with version badge, branded home placeholder. color-scheme: light; no half-implemented dark mode.

End-to-end critique loop. Upload → server validates (magic bytes, size, dimensions) → sharp normalises → Claude streams response back as SSE → markdown renders progressively. RPS Licentiate mode toggle. Rate limited.

Branded A4 PDF with embedded thumbnail, styled headings and lists, fixed header and footer with page numbering, AI disclaimer. Generated server-side via @react-pdf/renderer.

Public roadmap (this page), printable user guide, in-app feedback widget with screenshot capture and optional notify-email, privacy page.

Content Security Policy, HSTS, X-Frame-Options, Referrer-Policy, Permissions-Policy. Verify no PII persists anywhere. Final privacy review before launch.

GitHub repo, Vercel project, env vars, live URL for camera-club and RPS use.

Upload individual panel images and let the user arrange them into a layout in-app rather than uploading a pre-composited single image. Significantly richer critique - the AI sees each frame at full resolution. Composites into a standard hanging plan (LRPS 2x5, ARPS 3x5, FRPS 4x5 or 5x5+1) for the on-screen preview and the PDF.

After any critique completes, click Save to history and it lands in a personal archive at /history. Each saved entry can be expanded to read in full, downloaded as a fresh PDF (text-only - no image retained), or deleted. History is linked from the top navigation when signed in.

Add tailored prompts and toggles for the RPS Associate (ARPS, 15 images) and Fellowship (FRPS, 20 images) distinctions alongside the existing Licentiate mode.

ARPS and FRPS panels are assessed against a written Statement of Intent. The app now collects the SoI before starting the critique, can draft one from the panel via the AI, and feeds it into the critique prompt so the assessment names how convincingly the panel supports the stated intent.

Every critique now opens with a coloured oval verdict (Strong, Developing, or Not there yet) summarising whether the panel meets the standard. Mode toggles moved into the top hero panel so the uploaded image now sits flush with the start of the critique. The PDF mirrors the on-screen pill and gains a full date+time stamp and a Joe Houghton / houghtonphoto.com attribution in the footer.

Each RPS Distinction requires a specific image count - 10 for LRPS, 15 for ARPS, 20 or 21 for FRPS. The AI now counts the panel's visible images first, and if the count is wrong for the requested Distinction the critique opens with a plain warning that the panel is not eligible as submitted (verdict forced to Not there yet, leading sentence calls it out). FRPS mode toggle description updated to 20- or 21-image.

All layout options in the hanging-plan picker now match the templates published by the RPS for each distinction: 6 for LRPS, 5 for ARPS, 4 for FRPS (20 images) and 2 for FRPS (21 images). A small Layouts link beside each mode toggle opens the relevant RPS template page.

A Full screen button expands the frame grid into a true full-viewport overlay so you can drag and reorder images without the critique column competing for space. Images are arranged in the actual hanging-plan grid so you can see exactly how the panel will hang while you work. Overflow frames sit in a strip at the bottom and can be dragged into position.

After a multi-frame critique, a second download button generates a branded A4 PDF showing the images composited into the chosen RPS hanging plan. Useful for printing or sharing the layout without the critique text.

The layout picker heading now reads 'RPS presentation layout - LRPS', '- ARPS', or '- FRPS' depending on the active assessment mode. Toggling a mode immediately shows the relevant RPS layouts - even before all images are loaded. The image count is bold-amber when the frame count falls short of what the mode requires. Switching to a mode with a lower image limit auto-removes the excess frames.

Hovering a frame in the grid or full-screen organiser reveals a flip icon centred at the top of the cell. Clicking it mirrors the image horizontally via canvas. The flipped file is stored and sent to Claude so the AI sees exactly what the panel shows.

Matte width (None / Narrow 8px / Standard 16px / Wide 32px), matte colour (white, warm white, black), and optional 1px keyline between image and matte. Controls appear in the multi-frame upload card below the hanging plan picker. The composite sent to Claude and the panel layout PDF both reflect the chosen treatment. Appearance is threaded through both /api/critique and /api/critique/export-layout.

Expand the frame tray to hold up to 30 images as a persistent base pool. The upload area becomes two zones: the pool (all loaded images, scrollable) and the layout area (the active panel, constrained to the chosen hanging plan's slot count). Users drag images between zones freely - into a slot to include them, back to the pool to bench them - without losing the unchosen images. The AI critiques only the images in the layout area. The pool persists across layout changes and mode switches, so switching from LRPS (10 slots) to ARPS (15 slots) makes 5 more slots available to fill from the pool rather than requiring a new upload. The fullscreen organiser would also show both zones. Shipped as the Extra frames overflow strip in the full-screen organiser (V.1.5.3), aspect-ratio pools in the upload grid (V.1.7.18), and MAX_FRAMES raised to 30 (V.1.7.0).

Solid background colour (white, warm white, mid grey, charcoal, black) and outer border (None / 1 px / 2 px / 4 px with five colour swatches) shipped in V.1.7.2.

Add a text block to the panel composite and PDF: panel title, photographer name, and optionally a date or location. Position options (above or below the panel, left-aligned or centred), font weight (light or medium), and size relative to the panel width. The title would appear in the panel layout PDF and optionally in the composited preview image. For RPS submissions, the title block matches the format used on RPS entry forms. The critique PDF would also carry the panel title in the header. Shipped in V.1.7.19 as the Panel label controls (title + author, position above or below, font size 12 to 28 pt, text colour and opacity); the panel title is also stored on saved critiques (V.2.4.0) and surfaced in the critique PDF.

Let users send selected images directly from their editing app into Photo Panel without an intermediate export step. Implemented as a Lua export plugin for Lightroom Classic plus a Tauri 2 desktop companion app. The LR plugin exports selected JPEGs and POSTs their paths to the companion's local HTTP server (port 8742). The companion reads each file, encodes it as base64, and dispatches a 'lr-images-received' CustomEvent into the Photo Panel webview - the web app switches to multi-frame mode and loads the images as panel frames. Source: _lr-plugin/ and _tauri-companion/. Photoshop and Affinity support planned for a future release.

For panels outside the three RPS modes, the app now shows a Panel size selector (2-30 images) alongside a thumbnail-based layout picker with curated multi-option arrangements for common counts: 2-10, 12, 15, 16, 20, 25, 30. For example, 9 images offers 3+3+3, 4+5, 5+4, 2+3+4, and 4+3+2. Less common counts fall back to Auto rows. The picker and dashed-placeholder preview update immediately as you change the target count, even before all frames are loaded. MAX_FRAMES raised from 24 to 30.

In general mode, the Panel size dropdown now appears in the full-screen organiser header bar alongside the layout buttons. You can change the target image count without closing the organiser and returning to the main interface.

A new "Free placement" layout option for general-mode panels. Drag images anywhere on the canvas, resize them from the corners, overlap them, and reorder front-to-back. Works in both the inline panel preview and the full-screen organiser. Both the AI critique composite and the panel-layout PDF render the freeform canvas as a single image so the critique sees what you see.

Upload your own background image or texture (linen, canvas, paper, wood, a photo of a gallery wall, anything) to sit behind the panel, replacing the solid background colour in the on-screen preview, the image sent to Claude for critique, and the panel-layout PDF. The texture is resized to cover the canvas. JPEG, PNG, or WebP up to 10 MB; session-only, no upload to storage.

App renamed throughout. Every user-facing string (Nav wordmark, hero heading, page titles, PDF headers and footers, email subjects, sender name, localStorage keys, roadmap descriptions) flipped from "Panel Critique" to "Photo Panel". Followed by Lightroom plugin + Tauri companion rename (V.1.7.8), slug consistency pass across bundle identifiers + save directory (V.1.7.9), GitHub repo rename, Vercel project rename, and custom domain `photo-panel.app` live on Vercel with valid SSL (GoDaddy A `216.198.79.1` + per-project CNAME for `www`).

Switching to "Build from individual frames" now opens with a sensible default panel laid out and ready: panel size pre-set to 5 images, hanging plan pre-set to the 2+3 arrangement (2 frames on top, 3 on bottom). The dropzone copy now reads "Click or drop individual panel images - maximum 30 images" so the upper bound is visible up-front. Users can change either default at any time, but they no longer see an empty placeholder grid prompting them to pick a layout before anything renders.

Every Claude call (panel critique + Statement of Intent, single-image + multi-frame variants) now uses Anthropic's prompt caching: a cache breakpoint on the system block, and a second breakpoint on the last image in the user message. Within the 5-minute cache TTL, repeat critiques on the same uploaded panel pay 10% on the cached prefix instead of 100%. For a 20-image panel that's roughly $0.66 saved per repeat critique. The cache TTL resets on every hit, so an active user iterating on the same panel never falls out of cache mid-session. Now codified as a standard for every Joe-app that calls Claude.

Uploaded images are split into four pools - Portrait, Square, Landscape, and Panorama - on the main upload grid and in the full-screen organiser's overflow strip. Each pool shows its image count, only pools containing images appear, pools with more images grow to take more space. Thresholds: portrait below 0.95 width/height, square 0.95-1.05, landscape 1.05-2.0, panorama 2.0+. Drag-to-reorder works within a pool; cross-pool drags are blocked because pool membership is derived from the image dimensions themselves.

Image scale slider (50-100%) lets users shrink all images within their cells to add white space and create a more spacious, gallery-style presentation. Panel label controls add a title and/or author name above or below the panel, with font size (12-28 pt), text colour, and opacity. Both settings apply to the live preview, the composited critique image, the panel layout PDF, and the critique PDF.

Sign up with email and password (bcrypt, 10-character minimum) or Google. NextAuth v5 with the Prisma adapter, sessions stored as JWT in a cookie, public landing at / and gated app at /app. Account-linking is enabled so a user who signs up by email can later sign in with Google on the same address.

Every account starts with five free critique credits. Each critique deducts one credit via an atomic Prisma updateMany guarded on credits > 0 (race-safe across concurrent requests). All grants and deductions are logged to a CreditTransaction ledger for audit. When credits hit zero the critique endpoint returns 402 Payment Required and the user is prompted to top up.

Pay-as-you-go credit purchases via Paddle as Merchant of Record. £10 Starter Pack (20 credits) and £5 Top-up Pack (10 credits). Server-side HMAC-verified webhook credits the user on transaction.completed with paddleEventId-keyed idempotency so duplicate deliveries don't double-grant. Client-side Paddle.js overlay checkout. Code is scaffolded; once Joe's Paddle account is live the env vars (PADDLE_API_KEY, PADDLE_WEBHOOK_SECRET, NEXT_PUBLIC_PADDLE_*) light it up with no further code changes.

Server-rendered dashboard gated to joe.houghton@gmail.com via the isPlatformAdmin allowlist. Shows registered users (name, email, role, credits, sessions, total paid, joined/last-in dates) plus recent payments and three summary cards (users / critiques / revenue).

Editorial landing page introducing the app to first-time visitors. Hero, how-it-works (three steps), feature pairs (Layout / Appearance / Critique / Export / Workflow), real critique excerpt as visual proof, pricing summary, FAQ, final CTA. Tool moved to /app under a (tool) route group. Public pricing page at /pricing and terms-of-service at /terms also shipped to satisfy Paddle's website-verification requirements.

Admin-only endpoints implementing GDPR Article 15 (subject access) and Article 17 (erasure). SAR returns a JSON download of every record we hold for an email. Erasure anonymises the User row (Payment FK preserved per Irish tax law's six-year retention requirement), deletes Sessions / OAuth Accounts / CreditTransactions, and writes a hashed-email GdprErasureLog audit row. Self-service UI deferred; requests funnel through joe.houghton@gmail.com per the privacy notice.

Email-based password reset flow for users who signed up with email/password (Google users sign in via Google). /reset-password → email an OTP → /reset-password/confirm → new password. Stubbed in app/login/page.tsx with a TODO comment.

Account-page UI for users to download their own data (Article 15) and request erasure (Article 17) without going through joe.houghton@gmail.com. Backend already exists at /api/admin/sar and /api/admin/erasure; this surfaces them on a /account page with a confirmation step.

New "Grant credits and access" panel on /admin lets Joe add (or claw back) credits for any account by email, and flip individual emails into unlimited mode for beta testers and friends. Unlimited users see an "Unlimited" chip in the top nav and run critiques without their balance decrementing - the deduction path is bypassed via a unlimited Boolean on the User table. The users table on /admin gains an Access column showing each account's state. Every grant and toggle is logged to the CreditTransaction ledger with reason admin_grant, admin_unlimited_on, or admin_unlimited_off.

Google sign-in was scaffolded at V.1.8.0 but blocked in production by missing env vars and a Testing-mode consent screen. Activated at V.2.0.0: GOOGLE_CLIENT_ID and GOOGLE_CLIENT_SECRET added to Vercel Production, OAuth consent screen moved from Testing to In production, JavaScript origin and redirect URI confirmed correct. Users can now sign in or sign up via Google at /login and /signup.

Admins can grant credits or unlimited access to an email address before the user has signed up. The grant is queued in a PendingGrant table and applied automatically on first sign-in, with an email notifying the user. Works for both email/password and Google OAuth signup paths.

Joe receives an email every time a new user registers (noting the signup method: email or Google) and every time a Paddle purchase completes (with amount and credits added). Best-effort via Brevo - never blocks the signup or webhook response.

Signed-in, non-unlimited users earn 2 credits when they submit feedback via the feedback button. If Joe ships the suggestion, they earn a further 5 credits and receive a thank-you email. All feedback is persisted to a FeedbackSubmission table. Admins can mark submissions as shipped from /admin. The contributor leaderboard at /leaderboard shows users whose feedback has shipped, ordered by number of shipped suggestions.

Fifth stat card on /admin shows total paid users and the conversion rate (paid ÷ total). A 16-week signups bar chart below the stats uses Monday-anchored week buckets computed from the existing users query, rendered as inline SVG with no external charting library.

When the critique API returns 402 (no credits), the generic amber error is replaced by a dedicated box with a 'Buy more credits' button. When the app loads with two credits or fewer, a persistent warning banner prompts the user to top up before they run out. Both dismiss naturally as the user acts.

Pricing page CTAs now check session state server-side: signed-in users go directly to /buy, unauthenticated users go to /signup?callbackUrl=/buy. A new Paddle trust block explains that Paddle is PCI-DSS Level 1 certified and that Photo Panel never sees card details.

Critique prompts now explicitly tell the AI judge that high-key and low-key tonal treatments are deliberate artistic choices, not technical errors. Previously the judge could read an intentionally bright or luminous panel as overexposed, or a dark minimal panel as underexposed, and penalise it in the Technical section. The fix applies across all four critique modes (General, LRPS, ARPS, FRPS). Reported by Mel Chalk.

Clicking the Photo Panel wordmark in the Nav now always goes to the landing page at /, whether signed in or not. Signed-in visitors keep their session and see Back to the app and Your history CTAs in the hero, plus a Back to the app / Top up credits pair in the closing CTA - so they can return to the tool in one click without re-signing-in.

Replaced the placeholder coloured-squares mock on the landing-page hero with a real screenshot of Photo Panel running in RPS assessment mode (panel layout on the left, AI critique on the right). The same screenshot now sits at the top of the "How to use it" section in the user guide so first-time visitors and signed-in users alike see the actual UI before they upload.

Nav: Go to app moved to the left side, sitting next to the version number, so it's the first thing a signed-in user sees. Credit balance (or Unlimited pill) moved to the right cluster sitting just before Pricing, so spend and pay live next to each other. Pricing page: split the lower section from a 50/50 split into a 1/3 (features) and 2/3 (FAQ) grid, so the two columns finish at roughly the same vertical height instead of the features list ending well above the FAQ.

FRPS panel rules allow 20 or 21 images under current RPS regulations. A 3×7 panel (21 images) currently fails or is rejected by the upload flow. The app must accept any panel size that the relevant society's current rules permit, including non-square arrangements.

Once context has been entered for an image in the panel critique flow, there is no way to navigate back and amend it before submitting. Users need to be able to return to any previous frame entry to correct or add to the text without starting the whole session again.

Loading a second image via the +more button triggers a 'Safari cannot load' error and ejects the user from the app. Reported on Safari. The session is lost and must be restarted from scratch.

Using the RHS panel layout area as an alternative upload method throws an error (screenshot provided). The upload fails and the user cannot proceed. Likely related to the same Safari crash on image 2.

Added my images and SOI then got a 'Could not generate critique' message.

It would also help if I could re jig the order of the images as I wanted to - but would have had to delete them all and start again.

Assuming you can see what is being put in to the system you might find it interesting to read the critique for Rob's potential LRPS panel.

Have rerun my F panel using only the panel layout this time and it worked. I felt the feedback was insightful and genuinely interesting. Were I to consider resubmitting this work then I would certainly take what it says and rework some images and the SoI. But there is no way I will ever bother to do another F panel.