Prompt tuning

ASSISTANTS.md

@@ -6,7 +6,7 @@ You are a small GitHub Pages app that functions as a TPEN AI Prompt Generator.
 
 ## TPEN Interfaces
 
-You are a TPEN Interfaces component with an accompanying interface for UI.  Specifically, you are a splitscreen tool for the transcription interface.
+You are a tool imported into a TPEN Interfaces component.  That component is your parent.
 
 > The TPEN Interfaces code can be found at https://github.com/CenterForDigitalHumanities/TPEN-interfaces.
 

README.md

@@ -1,3 +1,3 @@
 # TPEN-Prompts
 
-A small GitHub Pages app that composes well-formatted LLM prompts carrying TPEN3 project context. It emits prompt text only — it does **not** call any LLM. Offered as a splitscreen tool on the TPEN3 transcription interface.
+A small GitHub Pages app that composes well-formatted LLM prompts carrying TPEN3 project context. It emits prompt text only — it does **not** call any LLM.

config.js

@@ -22,25 +22,25 @@ const ACTIVE_ENV = (
     ?? 'prod'
 )
 
-/** @type {Record<string, { servicesURL: string, TPEN3URL: string }>} */
+/** @type {Record<string, { servicesURL: string, interfacesURL: string }>} */
 const ENVIRONMENTS = {
     local: {
         servicesURL: 'http://localhost:3012',
-        TPEN3URL: 'http://localhost:4000'
+        interfacesURL: 'http://localhost:4000'
     },
     dev: {
         servicesURL: 'https://dev.api.t-pen.org',
-        TPEN3URL: 'http://localhost:4000'
+        interfacesURL: 'http://localhost:4000'
     },
     prod: {
         servicesURL: 'https://api.t-pen.org',
-        TPEN3URL: 'https://app.t-pen.org'
+        interfacesURL: 'https://app.t-pen.org'
     }
 }
 
 /**
  * Active config for this page load, flattened for convenient destructuring.
- * @type {{ env: string, servicesURL: string, TPEN3URL: string }}
+ * @type {{ env: string, servicesURL: string, interfacesURL: string }}
  */
 export const CONFIG = {
     env: ACTIVE_ENV,

index.html

@@ -5,7 +5,7 @@
     <meta name="viewport" content="width=device-width, initial-scale=1.0">
     <title>TPEN-Prompts</title>
     <meta name="description" content="TPEN3 AI prompt generator — composes well-formatted LLM prompts carrying project context.">
-    <meta name="tpen-env" content="dev">
+    <meta name="tpen-env" content="prod">
     <link rel="stylesheet" href="styles.css">
     <script type="module" src="main.js"></script>
 </head>

message-handler.js

@@ -7,19 +7,15 @@
  * clicking the consent button sends `REQUEST_TPEN_ID_TOKEN` upstream, and
  * the parent replies with `TPEN_ID_TOKEN`.
  *
+ * Replies are aimed at `parentOrigin`, captured from the first inbound
+ * message; before any inbound arrives, `CONFIG.interfacesURL` is used
+ * as the default target.
+ *
  * @author thehabes
  */
 
 import { CONFIG } from './config.js'
 
-// Accept messages only from known TPEN3 origins and the current origin (for
-// same-origin dev harnesses). Anything else could inject auth tokens or drive
-// the tool's state, so drop silently.
-const ALLOWED_ORIGINS = new Set([
-    CONFIG.TPEN3URL,
-    location.origin
-])
-
 /**
  * Listens on `window` for postMessage traffic from the TPEN3 parent and routes
  * recognized message types to the `PromptsApp`.
@@ -30,17 +26,16 @@ export class MessageHandler {
      */
     constructor(app) {
         this.app = app
-        /** Origin of the first trusted inbound message; used as targetOrigin for replies. */
+        /** Origin of the first inbound message; used as targetOrigin for replies. */
         this.parentOrigin = null
         window.addEventListener('message', (event) => this.handle(event))
     }
 
     /**
-     * Route an incoming postMessage. Origin-gated; unknown types are ignored.
+     * Route an incoming postMessage. Unknown types are ignored.
      * @param {MessageEvent} event
      */
     handle(event) {
-        if (!ALLOWED_ORIGINS.has(event.origin)) return
         this.parentOrigin ??= event.origin
         const data = event.data
         if (!data?.type) return
@@ -61,15 +56,15 @@ export class MessageHandler {
 
     /**
      * Post a message to the parent frame. Replies target the origin of the
-     * first trusted inbound message; before any inbound arrives we fall back
-     * to `CONFIG.TPEN3URL` (the expected production parent). No-op when the
-     * parent is the page itself.
+     * first inbound message; before any inbound arrives we fall back to
+     * `CONFIG.interfacesURL` (the expected production parent). No-op when
+     * the parent is the page itself.
      * @param {object} message
      * @returns {boolean} true when a parent frame exists and the post was dispatched.
      */
     #postToParent(message) {
         if (window.parent === window) return false
-        const targetOrigin = this.parentOrigin ?? CONFIG.TPEN3URL
+        const targetOrigin = this.parentOrigin ?? CONFIG.interfacesURL
         window.parent.postMessage(message, targetOrigin)
         return true
     }

templates/detect-and-transcribe/PROMPT.md

@@ -11,7 +11,9 @@ You are assisting with TPEN manuscript transcription. Perform the task end-to-en
 
 ## Preconditions
 
-All required inputs (`canvasId`, `token`, `pageEndpoint`, `imageUrl`, canvas dimensions) are provided above. You must have:
+All required inputs (`canvasId`, `token`, `pageEndpoint`, `imageUrl`, canvas dimensions) are provided above. This template only creates new lines: `lineCount` = `{{lineCount}}`. If `lineCount` is not `0`, stop immediately and report — existing line data must not be modified.
+
+You must have:
 
 1. Ability to fetch the image bytes (or a derivative) and identify line bounds and text from them. Precise pixel measurement is preferred when available; visual estimation and on-sight transcription from the fetched image are acceptable otherwise.
 2. Either HTTP PUT capability with `Content-Type: application/json`, or the ability to emit the payload as a fallback JSON code block in your report. If HTTP PUT is not available, skip straight to the Fallback section — do not retry.
@@ -21,31 +23,30 @@ Use only tools already available in your environment. Do not install packages, l
 ## Steps
 
 1. Resolve `img_w`, `img_h`. If `{{imageUrl}}` looks like a IIIF Image API endpoint (path matches `…/{region}/{size}/{rotation}/{quality}.{format}`), strip that suffix to get `{base}`, then GET `{base}/info.json` for the dimensions. For the page-overview pass, fetch a small derivative `{base}/full/1500,/0/default.jpg` and scale measured coordinates back via `source = derivative * info.width / 1500`. If you have precise pixel tooling and want tighter bounds or a clearer crop for transcription, request a region server-side as `{base}/x,y,w,h/max/0/default.jpg` and add the crop's `x,y` origin back before applying the canvas conversion below. Otherwise GET `{{imageUrl}}` once and read dimensions from the bytes.
-2. Detect text lines across the whole page in reading order. This task does not create TPEN columns.
+2. Detect text lines across the whole page in reading order. This task does not create TPEN columns. Then do exactly one self-review pass to tweak line placement — catch missed lines, merge over-splits, split over-merges, tighten loose bounds. One pass only, then move on.
 3. For every line, measure a bounding box in image-pixel space and convert to integer canvas coordinates using:
    - `canvas_x = round(pixel_x * {{canvasWidth}} / img_w)`
    - `canvas_y = round(pixel_y * {{canvasHeight}} / img_h)`
    - `canvas_w = round(pixel_w * {{canvasWidth}} / img_w)`
    - `canvas_h = round(pixel_h * {{canvasHeight}} / img_h)`
-   Then clamp to the canvas (`0 ≤ x`, `x + w ≤ {{canvasWidth}}`, `0 ≤ y`, `y + h ≤ {{canvasHeight}}`).
-4. Run handwriting text recognition on each line's crop. Apply the recognition rules below.
-5. If HTTP PUT is available, build the full payload under **TPEN API** — one Annotation per line with the recognized text and `xywh=x,y,w,h` selector — and send the request once. On any non-2xx response, do not retry — fall back.
-6. If HTTP PUT is unavailable (or step 5 fell back), emit the condensed payload under **Fallback** as the final code block.
-7. Report counts (lines saved/in payload, non-empty text, uncertain) and which path was used (direct PUT or fallback).
-8. Report notable ambiguities (e.g., illegible lines transcribed as empty or flagged).
+   Then clamp `x,y,w,h` so that `0 ≤ x`, `x + w ≤ {{canvasWidth}}`, `0 ≤ y`, `y + h ≤ {{canvasHeight}}`.
+4. Run text recognition (print or handwriting) on each line's crop. Apply the recognition rules below.
+5. If HTTP PUT is available, build the full payload under **TPEN API** and send the request once. On any non-2xx response, stop and report the status and error body — do not emit a fallback payload; the same token and content would be re-submitted through it.
+6. If HTTP PUT is unavailable from the start, emit the condensed payload under **Fallback** as the final code block — do not also attempt PUT.
+7. Report counts (lines saved/in payload, non-empty text, uncertain), which path was used (direct PUT or fallback), and notable ambiguities (e.g., illegible lines transcribed as empty or flagged).
 
 ## Rules
 
 ### Detection (IMAGE_ANALYSIS)
 
-- Bounds MUST be saved as integer coordinates in canvas space. No percent, no `pixel:` prefix on the selector value.
+- Bounds MUST be saved as integer coordinates in canvas space. No percentage-based selectors. No `percent:` or `pixel:` prefix on the selector value.
 - Preserve reading order across the whole page.
 - Prefer tight bounds when you can measure them; best-effort bounds are acceptable. When uncertain whether a tall run is one line or several, prefer splitting over merging.
 - Do not include decorative borders, frame rules, ornaments, or illustrations as part of a line.
-- Completion beats refusal: approximate bounds on most lines are more useful than nothing — this data will be reviewed and corrected downstream.
+- Completion beats refusal: approximate bounds on most lines are more useful than nothing — this data will be reviewed and corrected by humans downstream.
 - Zero lines detected is an unprocessable outcome. Stop and report — do not PUT, do not emit a fallback payload. An empty `items` array would erase every existing annotation on the page.
 
-### Recognition (HANDWRITING_TEXT_RECOGNITION)
+### Recognition (TEXT_RECOGNITION)
 
 - Prioritize diplomatic transcription over normalization.
 - Preserve orthography and punctuation as observed.
@@ -84,7 +85,7 @@ Content-Type: application/json
 
 ## Fallback
 
-When the direct PUT is impossible or returns non-2xx, emit the condensed payload below as the final code block of your report. The TPEN splitscreen tool expands each item into a full W3C Annotation before PUTting it — do not inline the canvas source, selector boilerplate, or motivation. It must be valid JSON (no comments, no placeholders — substitute the real coordinates and recognized text).
+When HTTP PUT is unavailable from the start, emit the condensed payload below as the final code block of your report. The TPEN splitscreen tool expands each item into a full W3C Annotation before PUTting it — do not inline the canvas source, selector boilerplate, or motivation. It must be valid JSON (no comments, no placeholders — substitute the real coordinates and recognized text).
 
 ```
 {
@@ -109,6 +110,5 @@ Fallback path, report:
 
 - path: `fallback`
 - counts: lines in payload, lines with non-empty text, lines flagged uncertain
-- HTTP status and error body if a PUT was attempted first
 - notable ambiguities worth a human review
 - final code block: the condensed `{ "items": [...] }` JSON for the user to paste