# AI usage LiturgyGen was built with AI assistance. This file is the record of it, kept as the work happened rather than written up at the end. **Assistant:** Claude (Anthropic), used through Claude Code in the terminal and through the Antigravity IDE for the earlier server work. **Roughly how much it touched:** most of the first draft of the services layer, the `.docx` builder and the React components. It did not touch the liturgical rules, the parser heuristics, the Prayers of the Faithful cascade, or the PostgreSQL migration — those are listed under *Who wrote what* below. Commit links point at `https://github.com/Damudachi/LiturgyGen`. --- ## 1. How I used AI ### 2026-09-13 — Windows desktop packaging - **Tool:** Claude (Claude Code) - **What I asked for:** a way to ship the app to an office PC with no Node.js, no browser tab and no administrator rights. - **What it gave back:** a C# launcher using WebView2 (the web view already in Windows 10 and 11), an Inno Setup script for a per-user install, and a build script that bundles a Node runtime, the server and the built client. - **What I kept, what I changed, and why:** kept the WebView2 approach — it was the right call and I would not have found it myself. Changed the process lifecycle: its first version left the Node process running after the window closed, so the port stayed held and the next launch failed. I added the parent-PID watchdog in `server.js` that exits when the launcher goes away. - **Commit:** `c34f132` — https://github.com/Damudachi/LiturgyGen/commit/c34f132 ### 2026-09-13 — Splash screen and window lifecycle - **Tool:** Claude (Claude Code) - **What I asked for:** something on screen while the bundled server boots, because the window was blank for three or four seconds. - **What it gave back:** a splash window in the launcher, shown until the server answers its health check, then handed off to the main window. - **What I kept, what I changed, and why:** kept it. I changed the seal image to be clipped to a circle, and fixed a bug where minimising the window made it restore itself about two seconds later. - **Commit:** `06c40ae` — https://github.com/Damudachi/LiturgyGen/commit/06c40ae ### 2026-09-14 — Shadow and depth system - **Tool:** Claude (Claude Code) - **What I asked for:** the calendar looked flat; I wanted depth without changing the layout, the colours or the type. - **What it gave back:** three shadow tokens (`rest`, `raise`, `float`) tinted with the palette's navy rather than grey, and a `.lift` class pairing the raise shadow with a small upward translate. - **What I kept, what I changed, and why:** kept all of it, strengthened the values because the first pass was too subtle to see, and added the `prefers-reduced-motion` branch that drops the translate. - **Commit:** `6ade5fe` — https://github.com/Damudachi/LiturgyGen/commit/6ade5fe ### 2026-09-14 — Dual installer builds - **Tool:** Claude (Claude Code) - **What I asked for:** one installer for the public that contains no copyrighted prayers, and one for the office that does. - **What it gave back:** a `--public` flag on `desktop/build.mjs` that skips `server/data/orillo` and writes to a different output folder, plus the `.gitignore` rules that let the public one be committed and keep the office one out. - **What I kept, what I changed, and why:** kept it. This is the piece that made the repository publishable at all. - **Commit:** `c34f132` — https://github.com/Damudachi/LiturgyGen/commit/c34f132 ### 2026-09-20 — Planning documents - **Tool:** Claude (Claude Code) - **What I asked for:** wireframes and a design system document generated from the code that already existed, rather than drawn by hand and left to drift. - **What it gave back:** Python scripts that render the screen map, the wireframes at two widths, the component tree and the design system sheets as images, reading the real token values out of `client/src/index.css`. - **What I kept, what I changed, and why:** kept the approach. It found a real bug while doing it: `--color-gold-edge` at 2.42:1 against the page fails the contrast requirement for a meaningful icon. - **Commit:** `cfba87c` — https://github.com/Damudachi/LiturgyGen/commit/cfba87c ### 2026-09-27 — Merging the class template - **Tool:** Claude (Claude Code) - **What I asked for:** fold the class template's required shape into this repository without breaking the 92 tests or the desktop build: `server.js` at the server root, `db/` with a pool and a schema, the `src/api/` demo-mode split, `/healthz` and `/readyz`, the Pages workflow, `.env.example` files. - **What it gave back:** all of it, plus `server/tools/build-demo-seed.mjs`, which generates the demo data for GitHub Pages from the real services. - **What I kept, what I changed, and why:** kept it. The one thing I insisted on was that the demo seed strip the scripture text and the office's prayers before writing a file that goes into a public repository — see case 3 below. - **Commit:** `d6baf70` — https://github.com/Damudachi/LiturgyGen/commit/d6baf70 ### 2026-09-27 — Security pass before going public - **Tool:** Claude (Claude Code) - **What I asked for:** audit the repository against the security announcement — secrets in history, workflow leaks, personal data, open write routes — and scaffold a Basic Auth gate **without writing the credential check**, because the brief says to write that myself. - **What it gave back:** the audit (clean history, no workflow secrets, but 19 unauthenticated write routes and my email on all 18 commits), the middleware wiring with `/healthz` and `/readyz` deliberately outside the gate, nine tests describing the contract, and a rewrite of the one SQL `WHERE` clause that was still built by string concatenation. - **What I kept, what I changed, and why:** kept the wiring and the tests. The detail I would not have thought of is that the stub has to **fail closed** — an unimplemented check throws, and the middleware catches it and returns 401 rather than calling `next()`. A gate that falls open when it breaks is worse than no gate, because you think you have one. - **Commit:** `9182b6d` — https://github.com/Damudachi/LiturgyGen/commit/9182b6d (the audit's own write-up landed alongside it in `468c5a7`) ### 2026-10-04 — The signed-out screens, a dedicated mark, and three auth regressions - **Tool:** Claude (Claude Code) - **What I asked for:** the landing page and the sign-in screen looked bland and too narrow on a laptop. Then, as they came up: the prayers from my local folder were not in the new accounts, the progress bar and the USCCB countdown had stopped moving, the email confirmation link errored, and I wanted a dedicated logo instead of the chapel's seal. - **What it gave back:** - `client/src/components/auth/EntryLayout.jsx`, the signed-out spread, and the same layout rewritten into `server/src/views/pages.js` for the page at `/`. - `client/src/components/Logo.jsx` and `client/public/favicon.svg`. - **Three bugs the accounts work had left behind**, all found by reading rather than by a failing test: `routes/settings.js` took `_req` and then called `req.orgId`, so `GET /api/settings` threw `ReferenceError: req is not defined` on every call and the client reported "LiturgyGen's server is not running"; progress used `new EventSource`, which takes no options and so cannot send an `Authorization` header, meaning the stream 401'd and both the progress bar and the server's own one-second USCCB countdown went dead; and `npm run seed` called `seedPotfTemplates()` with no `orgId`, hit the early-return guard and printed "0 inserted" as though it had worked. - `/auth-forward.js`, because Supabase sends email links to the Site **origin** and the app lives at `/app` — a fragment is never sent to the server, so a valid confirmation link was dropping people on a page that could not read it. - `server/test/boot.test.js` and `client/test/stream.test.js`, which cover the two regressions that were invisible. - The prayer-book importer: `server/src/services/importService.js`, `routes/potfImport.js`, `client/src/components/prayers/ImportBook.jsx`. - **What I kept, what I changed, and why:** kept nearly all of it. Two things I pushed back on. It first built the ribbon in the seam as **six** ribbons, one per liturgical colour; it read as a striped comb and the white one vanished against the paper, so it is one gold ribbon now — the same thing `Book.jsx` already uses. And it had restyled the sign-in screen while telling me there was no landing page in the repository, having searched only `client/`; the real one is server-rendered in `views/pages.js`, which is why my first two redeploys changed nothing. - **What this costs me:** a fair amount. The entry screens, the mark and the importer are AI-written, and section 3 says so. - **Commits:** `b989f40` — https://github.com/Damudachi/LiturgyGen/commit/b989f40 (the signed-out entry spread); `c6bc7ab` — https://github.com/Damudachi/LiturgyGen/commit/c6bc7ab (the mark, the favicon and `/auth-forward.js`); `ffcbf93` — https://github.com/Damudachi/LiturgyGen/commit/ffcbf93 (the `req`/`_req` boot regression and `boot.test.js`); `8a5a2f4` — https://github.com/Damudachi/LiturgyGen/commit/8a5a2f4 (the importer, the authenticated stream and `stream.test.js`); `7adcbfe` — https://github.com/Damudachi/LiturgyGen/commit/7adcbfe (OCR and the multi-file import flow) --- ## 2. Where the AI got it wrong ### Case 1 — It rebuilt the whole visual system, and it was worse - **What it gave me:** I asked for the design to look less generic. It replaced the fonts, the palette and the type hierarchy in one pass — a whole new identity, delivered confidently. - **What was wrong with it:** it looked worse than what it replaced. The cream/navy/gold palette came from the chapel seal and meant something; the replacement was a generic Ordo-inspired theme that meant nothing. The AI had no way to know that, because the reason was outside the code. - **What I did instead:** reverted the entire commit. Then I asked for depth and shadow only, keeping the layout, colours and type exactly as they were, which is what the screens actually needed and took an hour instead of an evening. - **Commit:** reverted in `8f979f1`, replaced by `6ade5fe` — https://github.com/Damudachi/LiturgyGen/commit/6ade5fe ### Case 2 — It suggested solving the USCCB bot challenge - **What it gave me:** when I said USCCB was blocking the scraper, one of the options offered was to drive a real browser at the page so the proof-of-work challenge would be solved by the browser itself. It would have worked. - **What was wrong with it:** the challenge is an access control that USCCB put there deliberately. Defeating it is not a technical problem to be solved, it is a decision about someone else's site, and the AI presented it as a straightforward option among others. - **What I did instead:** did not do it. The app keeps the grace cookie it is given, spends it fast (eight to eleven dates in ten to fifteen seconds), then leaves the site completely alone for three minutes. Everything it fetches is cached permanently, it falls back to Evangelizo, and there is a paste box for the date nothing else reaches. - **Commit:** `11d94f1` — https://github.com/Damudachi/LiturgyGen/commit/11d94f1 ### Case 3 — It would have committed copyrighted text to a public repository - **What it gave me:** the first version of the demo-mode seed generator serialised whatever `buildDay()` returned straight into `client/src/api/seed.json`, which is committed. - **What was wrong with it:** that output contains the full New American Bible reading text, copyright USCCB and the Confraternity of Christian Doctrine, and on this machine it would also have pulled in the office's transcriptions of two published ST PAULS volumes. Both would have been redistributed in a public repository, permanently, with my name on it. - **What I did instead:** the generator now replaces every body of text with one stand-in line and ships citations only, and the templates in the seed are the placeholder prayers written for this tool rather than anything from the books. `grep` for a distinctive phrase from the readings is part of the check. - **Commit:** `server/tools/build-demo-seed.mjs`, added in `d6baf70` — https://github.com/Damudachi/LiturgyGen/commit/d6baf70 ### Case 4 — It assumed the parser could be written from the page structure - **What it gave me:** a first USCCB parser that read each heading and the block under it as one reading. - **What was wrong with it:** days with optional readings do not work that way. On 8 September the First Reading is a choice between Micah and Romans, and the Gospel has a long and a short form; the "or" blocks belong to the section above them rather than standing on their own. The parser silently produced a first reading made of two different books stuck together. - **What I did instead:** saved the real HTML from the edge-case dates as test fixtures (`server/test/fixtures/usccb-2026-09-08-optional.html`) and rewrote the block-walking logic against them. This is why the parser tests exist. - **Commit:** `9e59339` — https://github.com/Damudachi/LiturgyGen/commit/9e59339 ### Case 5 — It restyled the wrong page, and told me there wasn't one I asked it to improve the landing page and the log-in. It searched `client/` for the landing page, found only the React sign-in screen, and told me plainly that there was no landing page in the codebase — then asked me a question built on that, and restyled the sign-in screen. The landing page is real. It is server-rendered at `/` by `server/src/views/pages.js`, with the React app mounted at `/app`. So I pushed and redeployed twice and saw nothing change, because the page I was looking at was the one file it had not looked in. **What I take from it:** the confident negative is the dangerous output. "There is no X in this codebase" is a claim about everything it did *not* read, and it had read one directory. When it told me later that `GET /api/settings` had been throwing on every call since the accounts commit, I checked that one against the database before believing it — and it was right, all 19 prayers in both parishes were placeholders. The lesson is not to distrust it; it is that an absence is a much weaker finding than a presence, and I should ask which files it actually opened. **Commit:** the spread finally reached the real landing page in `c6bc7ab` — https://github.com/Damudachi/LiturgyGen/commit/c6bc7ab, after two redeploys that changed only `client/`. ### Case 6 — It hid its own error message behind a 500 The prayer-book importer refused an image upload with "Something went wrong on the server." I sent it looking for a crash. There wasn't one. The code was raising a deliberately helpful error — "this looks like a scan, which needs optical character recognition, install it with..." — and giving it HTTP status **501**. The error handler hides the message of any 5xx in production, because a stack trace or a driver error describes your file layout to a stranger. 501 is a 5xx. So the one sentence that explained the problem was the exact sentence being suppressed, by a rule it had written itself a week earlier. **What I take from it:** it got both halves right and the interaction between them wrong. That is the shape of most of its mistakes — not a wrong line, but two correct lines that have never met. The fix was to let an error opt in to being shown (`error.expose`) and to use it only where the message was composed for the reader. **Commit:** `7adcbfe` — https://github.com/Damudachi/LiturgyGen/commit/7adcbfe. ### Case 7 — It gave me commands it had never run It told me to run `npm run seed -- --org ` to get the office's prayers onto the host. I ran it and got "DATABASE_URL is not set", while looking at a filled-in `.env`. Nothing in this project had ever loaded `.env`. Not the seed script, not `npm run dev`, not `npm run db:schema` — the project relied on Render putting variables in the environment, and `db/run.js`'s own header says to pass `--env-file=.env` by hand. The command it handed me could not work on any machine, and it had written that command, in a README section, twice. **What I take from it:** it will write instructions with the same confidence it writes code, and the instructions are the part nothing checks. Code at least has to import. From then on I asked it to run anything it told me to run. **Commit:** `cb90955` — https://github.com/Damudachi/LiturgyGen/commit/cb90955 — the commit that made `.env` load at all. ### Case 8 — It shipped the importer twice without running it I said it was fine not to test the prayer-book importer yet, and it took that further than I meant. Two separate faults were sitting in it: - `tesseract.js` v7 exposes `recognize` on its default export only, so reading it off the module namespace gave `undefined`. - Passing a third options argument — one whose single key was `undefined` — killed the OCR worker from inside a `MessagePort`, which escapes a `try/catch` and takes the whole process down instead of rejecting. Both surfaced in the first thirty seconds of actually running it on a PNG. **What I take from it:** "not tested" and "probably works" are not the same claim, and it does not reliably distinguish them. When it says a path is untested, that is a statement about its confidence, not about the odds. **Commit:** `7adcbfe` — https://github.com/Damudachi/LiturgyGen/commit/7adcbfe — both faults were fixed in the commit that added OCR, because running it was what found them. ### Case 9 — It estimated its own output instead of measuring it It wrote the presentation script with a word budget per slide and a stated total of four minutes fifty. Then it measured what it had actually written: **six minutes twelve**. Half the beats were over their own budgets, and three of the headings claimed word counts that did not match the words underneath them. It only caught this because I had asked for the timings in the first place; it would otherwise have handed me a script that overran a five-minute cap, and I would have found out mid-take. **What I take from it:** the arithmetic it does about its own work is a guess wearing a number. It is good at writing the script and good at counting the words, and it will not do the second one unless told to. **Commit:** none in this repository — `docs/08-presentation-script.md` is git-ignored, because it says my name in its first line. It is in the private coursework folder with the rest of the identifying material. ### Case 10 — It ignored a note the codebase had already written for it Asked to improve the sign-in screen, it put six ribbons in the seam, one per liturgical colour. Two things were wrong: at seven pixels they read as a striped comb rather than bookmarks, and the **white** one was invisible against the cream half of the page. That second fault is the interesting one. `client/src/index.css` has a comment directly above the liturgical colours that says, in so many words, that white is an ivory stripe and needs an edge so it shows — written months earlier for the calendar, for exactly this reason. The file was open. It had read it well enough to copy the palette out of it. **What I take from it:** it reads a file for what it is looking for, not for what the file is telling it. The comments that exist to stop somebody repeating a mistake are the ones it is most likely to skim, which is an argument for the comments being there and not an argument against them. **Commit:** `b989f40` — https://github.com/Damudachi/LiturgyGen/commit/b989f40 — the entry spread as it landed, with one gold ribbon rather than six. --- ## 3. Who wrote what ### Written by me - **File:** `server/src/services/philippineOrdo.js` - **Commit:** `60cbd17` — https://github.com/Damudachi/LiturgyGen/commit/60cbd17 - **What it does and why it is built this way:** romcal ships a Philippine particular calendar, but it is missing local celebrations that the printed Philippine ORDO carries, and it ranks a few of them differently. This file is the overlay that fixes that. It holds the local feasts romcal does not have, and a `precedenceRank()` function that decides which celebration wins when two land on the same date — the liturgical table of precedence, in code, where a Solemnity beats a Feast beats a Memorial beats a ferial day. I wrote it because the rules come out of the ORDO, not out of anything a model has read, and because getting them wrong means the wrong readings are printed for a school Mass. `nthWeekdayOfMonth()` is in here too, for the feasts fixed to a weekday rather than a date. - **File:** `server/src/services/potfService.js` — the `resolveForDay` cascade - **Commit:** `28b679d` — https://github.com/Damudachi/LiturgyGen/commit/28b679d - **What it does and why it is built this way:** picks which Prayers of the Faithful page belongs to a date. Six match rules, most specific first: celebration id, then the calendar date (MM-DD, because the book keys 17–24 December and 2–5 January to dates rather than to liturgical weeks), then season + week + weekday, then season + week, season + weekday, and season alone. Those six are run as six passes — the day's own season, its ferial season, then the Ordinary Time book for the same week and weekday, with every specific prayer tried before any whole-season catch-all. It stops at the first hit. (This paragraph said "seven tiers" until 2026-10-11, as did three other documents; counted against the code it is six and six.) The order is the order the office uses when they flip through the two physical volumes, and I worked it out by watching them do it. The fallback to Ordinary Time is the part a naive matcher gets wrong: a day in a proper season with no page of its own does not get *nothing*, it gets the Ordinary Time page for that weekday. - **File:** `server/src/lib/orilloParser.js` - **Commit:** `9f8f3c8` — https://github.com/Damudachi/LiturgyGen/commit/9f8f3c8 - **What it does and why it is built this way:** turns a page typed or scanned out of the General Intercessions books into a structured template. The printed page has no markup — it is a heading, a paragraph of invitation, a response in capitals, numbered intentions, and a concluding prayer, separated by nothing but layout. The parser works on those conventions: capitals mean a response, a leading numeral means an intention, the last paragraph is the conclusion. It also repairs the OCR mistakes that kept recurring in the scans. This is the file I rewrote most often, because every new section of the book broke an assumption in it. - **File:** `server/src/middleware/retired/basicAuth.js` — the `checkCredentials` function. **Not mine. Written by Claude on 4 October 2026** and disclosed in section 1 under that date. I had planned to write it myself; I ran out of week and asked for it instead. The surrounding middleware, the wiring in `app.js` and the eleven tests that covered it were also scaffolded with assistance. **This gate is retired.** Supabase Auth replaced it later the same week — a single shared password cannot name a parish — so the file sits under `retired/` and its tests are parked as `server/test/retired-basicAuth.test.js.txt`. It is left in the repository, and described here, because the disclosure stands whether or not the code is still wired in: `server/src/middleware/requireAuth.js` is what guards `/api` today. - **What it does and why it is built this way:** decides whether an `Authorization: Basic ` header carries the configured username and password. Four things in it are not obvious until you read RFC 7617: the username cannot contain a colon but the **password can**, so it has to split on the first colon only rather than on every colon; the scheme token is case-insensitive, so `basic` has to be accepted as well as `Basic`; a malformed or missing header has to be a rejection rather than a crash; and the comparison has to be constant-time via `crypto.timingSafeEqual`, because `===` on a string returns the moment two bytes differ and that timing difference leaks how much of the password is right. Hashing both sides first is the usual way round `timingSafeEqual` throwing on buffers of unequal length, which would otherwise leak the password's length. - **The PostgreSQL migration is NOT mine.** This bullet used to say it would be "mine end to end" and that "every one of the twenty query rewrites is my own". I wrote that as a plan, in advance, and then it is not how the week went: I ran out of time and Claude did the migration - the schema, the twenty query rewrites, the async conversion through every call site, the `readings_cache` table, and the tests. `docs/07-postgres-migration-map.md` records the work; section 1 of this file records who did it. Leaving the original sentence in would have been the one dishonest line in this document, so it is replaced rather than quietly deleted. - **File:** `server/src/routes/` — the whole HTTP surface - **Commit:** `7e30ee4` onward — https://github.com/Damudachi/LiturgyGen/commit/7e30ee4 - **What it does and why it is built this way:** the REST API: `calendar`, `readings`, `generate`, `batch`, `potf`, `settings` and `account`. About nine hundred lines, and until now filed in this document under the AI, which was a filing mistake rather than a decision: the AI did not write them. Two things in here are worth naming. The first is what each route refuses: a malformed date is a 400 with a message written for the office rather than a 500 with a driver error in it, because the person reading it is a parish secretary and not me. The second is which routes are **streams** rather than requests — a batch over two hundred days cannot answer in one response, so `POST /api/batch` returns a job and `/api/batch/:id/events` reports on it as it goes, cancellable half way. That shape is the reason the panel can say "making day 4 of 22" at all. - **Files:** `server/src/services/calendarService.js` and `server/src/lib/dates.js` - **Commit:** `7e30ee4` onward — https://github.com/Damudachi/LiturgyGen/commit/7e30ee4 - **What they do and why they are built this way:** turning a date into a liturgical day, and the month grid the calendar screen draws. These sit directly on top of `philippineOrdo.js`, which is also mine: romcal answers "what does the general calendar say", the ORDO overlay corrects it, and this is the layer that asks both in the right order and hands one answer to the rest of the app. `lib/dates.js` is small and deliberately boring, and it is where the timezone handling finally settled. Every date in this application is a calendar day in the Philippines, never a moment in time, and these functions are the only place allowed to know that. ### Written by the AI, named here - **Files:** `client/src/components/auth/EntryLayout.jsx`, `client/src/components/Logo.jsx`, `client/public/favicon.svg`, the spread in `server/src/views/pages.js`, `server/src/services/importService.js`, `server/src/routes/potfImport.js`, `client/src/components/prayers/ImportBook.jsx`, `server/test/boot.test.js`, `client/test/stream.test.js`, and the `/auth-forward.js` route. - **Commits:** `b989f40` — https://github.com/Damudachi/LiturgyGen/commit/b989f40, `c6bc7ab` — https://github.com/Damudachi/LiturgyGen/commit/c6bc7ab, `ffcbf93` — https://github.com/Damudachi/LiturgyGen/commit/ffcbf93, `8a5a2f4` — https://github.com/Damudachi/LiturgyGen/commit/8a5a2f4 and `7adcbfe` — https://github.com/Damudachi/LiturgyGen/commit/7adcbfe - **What they do and what I decided:** the signed-out spread and the app's own mark, which replaced the chapel's seal because that seal belongs to one parish and the app now serves more than one; the prayer-book importer; and the two tests that cover the regressions nothing else caught. My decisions inside them: one gold ribbon rather than six liturgical ones, and — the one I would defend hardest — the importer does **not** guess a page's season, week or weekday. It reads a title and stops. A page number is not liturgical data, and the whole argument of `philippineOrdo.js` is that this application does not invent that mapping. So every imported draft arrives unticked and needs a day set by hand before it can be saved. That makes importing a hundred pages slower, on purpose. ### Mine, but not a file in this repository - **Not in this repository:** the Supabase project - **Evidence:** migration `20261004110845_multi_parish_schema_with_rls`; ten policies across seven tables, listable with `select * from pg_policies` - **What I did and why it is not a file here:** I set up and configured Supabase — the project, the Auth settings, the Row Level Security migration, and the environment wiring on Render. None of it is a line of code in this repository and all of it is load-bearing, so it is named here rather than left out because it does not show up in a diff. The part worth defending is the RLS policies. The scoping the application relies on is in the services, which name `org_id` on every query; these policies are the second lock, for anything that reaches the database without going through the app at all. They hang off one function, `user_org_ids()`, which is `SECURITY DEFINER` for a specific reason: a policy on `memberships` that reads `memberships` to decide who you are recurses into itself. That function is also the one standing Supabase advisory, written up in `Docu/project/SECURITY-CHECKLIST.md` rather than quietly fixed. `readings_cache` is the one table whose policy is read-only and not scoped to a parish, which is the same deliberate exception the schema makes. ### The AI-written part I understand best - **File:** `server/src/services/docxService.js` - **Commit:** `44a3298` — https://github.com/Damudachi/LiturgyGen/commit/44a3298 - **What it does and why we kept it:** builds the Word file. I had not used the `docx` library before and this would have taken me a week on my own; the first draft of the paragraph builders came from the AI. I understand it because I spent a day fixing it against the office's sample document, and the fixing is where the learning was. The office's missalette is not "a Word document with the readings in it". It is a specific geometry: Letter paper, margins of 1.0625 / 0.8125 / 0.5625 / 1 inches, Book Antiqua, and a right-aligned tab stop at 9630 twips so the scripture citation sits flush with the right margin on the same line as the heading. A twip is a twentieth of a point, which is the unit Word actually stores positions in, and that single number is why the citations line up. Settings feed in as style overrides: whether each reading starts a new page, whether the first psalm response is printed in capitals, whether "Let us pray to the Lord." is appended to each intention. The structure I kept; the numbers I measured myself out of the sample.