A working history of how one temple — and one translator — learned to render sacred teachings with machines. 2020–2026.
Since at least 2020, a small circle at Tong De temple has rendered its sacred teachings — 訓文, holy words received through 三才 and channeled instruction — from Classical Chinese into English by hand. Over the years that followed, as one of that circle took to AI tools early and kept experimenting, year after year, the process changed almost past recognition: from a shared dictionary and six-hour sessions, through personal ChatGPT experiments and jury-rigged spreadsheets, past two generations of Zapier automation, into a purpose-built multi-agent Claude pipeline that treats every teaching as a living, multi-register document.
What follows is that record, reconstructed from receipts, files, code, and memory. It is as much a portrait of one member’s AI-adoption curve as it is of the tools themselves — and since that member has, for years, run the temple’s technology, it is quietly a portrait of the temple’s too.
Where the translation itself lived — the source of truth migrating from paper to git:
Before any AI tool touched a single line, translation was entirely a human act. A Google Translate tab for spot-checking single words, a thesaurus, a dictionary — that was the whole toolkit. The rest was people, arguing carefully over register and meaning until a line felt right.
It was slow by design as much as by necessity: more than six hours to carry four to six lines of a holy teaching into English, each phrase weighed for classical allusion and theological precision by a small standing team — several 前賢 (senior temple practitioners) — with one to three others joining occasionally.
| 師尊 / 祖師 | Patriarch | 10/30/24 meeting |
| 與時伸屈 以義變應 | Adapt to each situation, always respond with uprightness | English class prep group |
| 義 | uprightness (less preferred: righteousness) | Mohr temple |
| 人心惟危, 道心惟微 | the human heart is precarious; the Dao heart is subtle | 前賢 |
| 有教無類 | Teach without discrimination | 前賢 |
The account was created the evening of 2023-03-26, timed almost exactly to GPT-4’s public launch two weeks earlier — three independent records (a Gmail verify email, a Plus receipt, the first Monarch charge) agree to the day. Chat by chat, teaching by teaching, translations “got better and better, but were still not quite up to snuff.”
On the very same day, the translator ran his first real scripture-translation prompt on both platforms he’d just gained access to. ChatGPT, then on GPT-4:
The following is a poem written in old poetic Chinese (文言文). Please translate it into plain English: 心無挂碎 如日普照虚空 / 無入而不自得 天人感通 / 一切法不取扑 正見般若行 / 化導見性 心源若水常流通
And that same afternoon or evening, on Bard — Gemini’s predecessor:
Please translate the below from old Classical Chinese to plain English: 修辰人生有苦也有甘 禁得起…
Neither platform had any special claim yet. He was simply trying both.
GPT-3.5, the ChatGPT default, handled casual lookups from April through September 2023 — but was never the tool for real scripture work. That belonged to GPT-4, from the very first request above, through August 2024. A stable, reusable glossary template emerged by late 2023 and held for nine months:
Please translate the below text delimited with triple quotes. If it is in English, please translate to traditional Chinese. If it is in Chinese, please translate it to English. Please preserve as much of the original meaning and contextual tones and clues as possible. Of note, please consider the following as equivalent in your translation … hx = 後學 / qx = 前贤 / temple = 佛垫
Much of the real craft was single-term wordsmithing — hunting for the exact right English word across many turns:
He went deep enough into the mechanics to ask about API parameters directly — and in doing so, confirmed exactly how the Sheets-formula era worked:
Not a hand-rolled script, but a known Sheets add-on — run alongside a separate, hand-rolled Apps Script (“Airtable Integration,” Dec 2023) trying the same bridge a different way. The card ledger shows exactly when that habit turned into real spend:
The smoking gun: a folder named “Translations in Progress” was created that same day, 2023-05-24. Its seed sheet ran the same lines two ways at once — GPT-3.5 Line-By-Line beside GPT-4 Together. The API balance emptied years ago, so the live cells now read only [GPT ERROR] — but the frozen columns kept the actual 2023 output. Here is the opening line, exactly as each model left it:
That is the translator A/B-testing models and granularity inside the spreadsheet itself, three years before the reader’s own register comparison. From there the model arc ran on: gpt-4o (everyday lookups, May 2024–mid-2025) → a single o1 conversation (Feb 2025) → o3 (May–Jul 2025, the whole-teaching workhorse — see the next exhibit) → the gpt-5 family, tapering through early 2026. Last chat of any kind on the account: 23 Feb 2026.
Genuinely “sprinkled in,” exactly as recalled — a minor thread inside a mostly-personal Gemini history, quiet for most of three years, until January 2026: a dense cluster of line-by-line pinyin requests for 訓文/善歌 lyrics, the direct manual ancestor of the reader’s current pinyin layer.
The same week, an instinct shows up before it was ever codified into a rule:
One week later, an actual workflow would begin automating exactly this.
The Sheets experiments hardened into something more permanent: pull the Chinese straight out of Airtable, translate it, push the English straight back in. It was, in the translator’s words, “a major but clunky undertaking” — and the record shows exactly how clunky. In the same week of December 2023, two competing approaches were tried at once: a Zapier automation (“Create Translation Google Sheets from Airtable Teachings Entry,” 11 Dec) and a hand-rolled Google Apps Script literally named “Airtable Integration” (17 Dec).
Zapier won out. By April 2024 every path had consolidated into a single automation — a checkbox on each Airtable record, ticked by hand, that fired GPT-4o once per section (鎮壇詩, 吾乃, 本訓, 訓中訓, and the sung lyrics with their pinyin) and wrote the English straight back into the cell.
You are a world-renowned Classical Chinese scholar that can translate Classical Chinese to modern English fully capturing the flavor and tone of the original. You respond with the translation only, without explanation or text before or after. You translate the provided text line by line, staying loyal to where lines break, but you translate each line also keeping in mind the context of the line in the frame of the entire work. You understand that the provided piece is written by someone who embodies tremendous Dao, compassion, humility, benevolence, harmony, simplicity, and prajna/wisdom. Despite being humble, they are also enlightened and awakened. Of note, you translate "Φ" as "Heavenly Mother" or "God", whichever seems more appropriate. You provide the translation of the provided text only, without explanation. When there is ambiguity in whether something is written in first, second, or third person, you translate equally ambiguously in poetic fashion. For example, instead of "I walk through the fields", you might just say "walking through the fields."
Already present here, in the very first automated prompt: the Φ → Heavenly Mother rule, tone-marked pinyin, line-break fidelity — every one of these survives, more rigorously enforced, in the current pipeline three years later.
What it couldn’t do: see a teaching as a whole. Each section was translated alone, with no research, no cross-line memory, and no way to know what the line before or after was doing.
The turn came from a simple, hard-won observation: a line translates better when the model can see the whole teaching, not just its neighbors. Early Sheets formulas had tried feeding in the line before and after as context — it helped, but it wasn’t the same as the model actually understanding the piece.
The fix was to stop translating in fragments. A Zapier automation would pull an entire teaching — 鎮壇詩, 吾乃, 本訓 together — into a Google Doc, pre-loaded with a full, ready-to-run prompt for ChatGPT’s reasoning model. A person ran it, pasted the result back in. It worked, for a while.
We need a very good translation of the below holy spiritual teaching. Of note, "Φ" (read: Mǔ) is translated as "Heavenly Mother" or "God", whichever seems more appropriate. We can translate '語寄' as 'Message'. For each line in the provided source text, please return Chinese, pinyin using diacritics (without hyphenation and without combining compound words, and don't capitalize pinyin unless a proper noun), and English, using the indicated format. When providing additional explanation, you can addend that in indented format below the English translation line before moving on to the next line. Please understand the general gist of the teaching first so that the line-by-line translation makes coherent sense. If you encounter parenthetical things containing "取" such as "（A取B）", this is beyond the scope of this translation, and the full parenthetical segment can be fully ignored in your output (no Chinese, no pinyin, no English). For advanced terminology such as 五乘, 十聞, 八正道, etc., or with words that are certain references to old classics or sutras, you will help me greatly if you do the appropriate research and provide the historical original Chinese context, explaining accordingly in English. If you provide citations with links, please have the visible text of the link simply be the domain, e.g. [fo.sina.cn](https://url). You can set the title of the conversation to be based on 誘寄, 語寄, or if not available, based on the name of the spiritual teacher that introduced themselves. Here's the preferred output for each line that will subsequently be pasted to a Google doc — four-level heading for Chinese, italic pinyin, plain English, and any explanation as a blockquote beneath — given as a raw Markdown code block, sections in H2, one horizontal rule between sections and none within.
Nearly everything the current pipeline still does was first asked for here: understand the whole piece before translating any of it; research 五乘, 十聖, 八正道 and cite sources; produce Chinese, pinyin, and English together, line by line. What was missing was only the machinery to do it without a human pasting text between two apps.
The idea of an actual workflow for translation — not a chat, not a spreadsheet formula, a genuine multi-step pipeline — had been forming for months. It needed the right tool to land in. Claude Code turned out to be it. A couple of weeks into using Claude in earnest, the first version appeared: /forge.
| Step | Model | Job |
|---|---|---|
| 1 · Research | Gemini 2.5 Pro | trace phrases to classical sources, live web search |
| Gleaned Insights | Claude Sonnet | research → 3–6 translation-critical calls |
| 2 · Literal | Claude Sonnet | word-for-word, research-informed |
| 3 · Smooth | Claude Sonnet | natural modern English |
| 4 · Review | Claude Sonnet | fidelity check vs. the original |
| 5 · Commentary | Claude Sonnet | reflective prose reading |
forge proved the two-model split — Gemini for grounded research, Claude for translation craft — and a living norms file that never silently deletes a retired term, only supersedes it with its reasoning attached:
皇Φ — "Imperial Mother" → replaced by "Sovereign Mother" Reason: "Imperial" carries unavoidable dynastic connotations from Chinese history — it evokes the emperor's court, not the cosmic feminine. In spiritual teachings, the 皇 register is about supreme, originating authority, not historical empire. "Sovereign Mother" preserves the authority and elevation while shedding the political-historical freight.
But every step shared one context window — an early step could bias a later one — and Gemini’s research sometimes carried phrases over from a previous teaching’s session. A successor was already named before forge’s own files were even finished: an internal note dated 27 February 2026 earmarks “/distill — new translation workflow.” forge’s spec was finalized 7 March 2026. distill’s first real run landed that same day.
distill fixed forge’s central flaw by refusing to share context at all: every step runs as its own agent, seeing only what it needs, so an early guess can never quietly bias a later judgment. What began as three outputs — a modern-Chinese explanation, a faithful translation, an English explanation — has grown into a full pipeline with its own institutional memory, over thirty teachings deep.
| Step | Model | Job |
|---|---|---|
| 1 · 研究員 | Gemini 2.5 Pro | allusion & classical-source research |
| 2 · 解經者 | Opus | 白話解說, modern-Chinese explanation |
| 3 · 譯者 | Opus | literal translation, pinyin, elevation map |
| 4 · 潤筆 | Opus | poetic smoothing, in isolation |
| 5 · 校讀 | Opus | fidelity review against the source |
| 6 · 解說者 | Opus | English explanation |
| 7 · 省察 | Sonnet | write the learning-log entry |
| 8 · 總編 | Opus | assemble scholar + plain registers |
Pinyin used to be its own guess, made cold from bare characters. Now the translator commits to a word’s grammatical role while translating it — and that commitment is the pronunciation. One recent pass folded that insight back through the whole pipeline: pinyin, and a per-teaching map of which sacred names appear, are now byproducts of steps that already had to work them out, not separate AI calls re-deriving the same answer. The Overview synopsis is skipped entirely when it’s already been authored. Nothing runs twice.
That same week, the reader itself went live — a page where anyone can toggle Chinese, pinyin, English, and annotation on or off independently, switch between scholar and plain registers, and rate a line directly. The translation is no longer a document that gets produced once and sits still; it’s a living surface, built once, expensively, and then reused richly.
A quiet surprise lives in that register switch. scholar is the faithful, footnoted rendering — the audit layer, technical vocabulary kept whole. plain was only ever meant to be the accessible one: the same meaning, jargon stripped, no dictionary required. Yet again and again it is plain that lands the line more truly. Where scholar renders 正者止一 as the terse “stop at one,” plain writes “come to rest in the One and see your own true nature” — and the second isn’t merely easier, it is closer. Both fall out of the same final pass, from one vetted meaning; the difference is only that plain trusts the reader to receive the depth without the scaffolding around it. The register built to be simplest keeps turning out to be the best translation in the room.
From distill’s first day, the Chinese-output steps — 解經者 (Step 2) and its review, 白話校閱 (Step 2b) — are prompted entirely in Traditional Chinese, by design, not habit. Their own header says why:
Five weeks later, a rigorous study proved out why. Benchmarking five local models against Opus across every distill step (2026-04-26), the single most important finding wasn’t about cost — it was about language pull. Step 6 (English Explanation) carries an English system prompt, but reads Step 1’s Chinese-language research notes as input:
Two of three local models tested (mistral-large:123b, deepseek-r1:70b) produced Chinese output anyway -- overriding an explicit English instruction, because the Chinese-language CONTEXT pulled harder than the English INSTRUCTION. Only Opus, and qwq:32b (itself natively Chinese-trained), held the instructed output language. "The Step 6 prompt is fragile to a language-anchor flip -- when a substantial amount of input is in another language, weaker models drift to that language."
Put together: the March design choice wasn’t stylistic preference, it was working with a real, measurable pull rather than against it. The study’s own recommended guard for the reverse case — explicit “your output language is English” anchors on Step 6 — was proposed but never actually applied to the production prompt. Not urgent today, since only Opus runs there in practice; a real open item the day a cheaper model gets swapped in.
Every exhibit above describes a method. This one runs them — the same source text pushed, for real, through each era’s actual prompt, side by side, so the difference is something you watch happen rather than take on faith.
The piece is the opening verse (鎮壇詩) of a 濟公活佛 teaching on 正精進 — Right Effort, the sixth limb of the Noble Eightfold Path — received at 同興壇, Taiwan, in October 2024. Eight lines, and inside them at least six traps that reward exactly the depth each era did or didn’t have: two archaic false-friend characters, a character that defines itself by coming apart, a double citation from the Book of Changes, a Yogācāra technical term, and a line that is really a buried quotation from the Book of Documents. It is a piece built to punish a shallow reading.
Two of the traps make the cleanest demonstration. For each, here is the same line as every method rendered it — earliest at the top, today at the bottom. Watch the meaning surface rung by rung.
The line defines its own key character. 正 (“upright”) is written as 止 (“stop”) + 一 (“one”): to be upright is to stop at the One. A reader who can’t see the character being dismantled renders a vague platitude and moves on.
The closing line abbreviates the Book of Documents (尚書·說命): 「若藥弗瞑眩，厥疾弗瘳」 — “if the medicine does not bring on vertigo, the illness will not be cured.” The dizziness isn’t a side-effect to tolerate; it is the proof the cure is working. Miss the citation and you invert the logic.
The pattern is the same in both, and it isn’t a clean staircase — it’s a real one. GPT-3.5 produces confident mush. gpt-4o is smoother but just as blind to the depth, and sometimes lucks into a right word while getting the logic backwards. o3, given the whole piece and told to research, genuinely finds some allusions (it names Yogācāra and the Great Learning unprompted) — yet still walks straight past the two hardest traps. forge, with a dedicated research pass, is the first to get both. And only distill also hands you the reason:
This is the Book of Documents's line "if the medicine does not bring on vertigo, the sickness will not be cured": bitter advice and strong medicine must disturb in order to work; treatment that feels comfortable is not reaching the root. So where your practice stalls -- where you cannot move forward -- that is exactly where the illness is showing itself. Do not detour around it; probe the principle. The vertigo is not failure; it is proof the cure has begun.
And then the honest twist — the one the museum owes you after five boxes of forward motion. Go back to that acrostic: 正 · 精 · 進 · 行, the teaching’s own theme written down the left margin. Not one of these methods put it on the page — distill’s published reader included. forge’s rerun happened to notice it in its private working notes and let it lie there; every shipped rendering, current pipeline and all, walks past it.
That is the real lesson of the time machine, and it is not “the last box wins.” It is how much a machine can now carry — a dismantled character, an inverted proverb, a school of Buddhist psychology, all surfaced and explained in one automatic pass — and exactly where the carrying still stops. The gap didn’t close. It moved, to somewhere higher up the page.
It’s easy to look at a run finishing in minutes and conclude the machine did it all. This record says otherwise: every era got faster to run only because someone spent real ingenuity, resourcefulness, and iteration getting it there — and that setup labor didn’t shrink as the pipeline matured. It grew. This arc took years, and it moved in lockstep with whatever the cutting edge actually was at the time — nobody skipped from a committee dictionary to an eight-agent pipeline; every awkward, gerry-rigged intermediate step was a real rung someone had to climb.
Two clocks matter here, and blending them hides the story. Claude-hours and human-hours are worth keeping apart — one shows what the machine carried, the other what the person did — and so is a second split: time spent setting up a method vs. time spent running it. The arc across every era: setup grew, running shrank.
| Era | Setup time | Running time |
|---|---|---|
| Stone Age | Maintaining the shared glossary, convening the group. | The session itself is the work: 6+ hrs for 4–6 lines. |
| ChatGPT / Gemini | Landing on a reusable prompt template — nine months stable. | One chat per term — six turns just to land 緣影. |
| Airtable / Era 1 | Building the Zap, wiring Airtable, a parallel Apps-Script attempt. | Seconds, one checkbox — but shallow, single-shot output. |
| Reasoning models | Refining the whole-piece prompt into its mature form. | Still hand-run, read in full, carried into a Doc. |
| forge / distill | An 8-agent architecture, 30+ retrospectives, a 5-model benchmarking study, an efficiency pass. | Minutes to kick off. Senior review now happens after publication. |
Every subscription era left a real trail — not just ChatGPT’s:
Read plainly: the current era costs 9–20× more per month than any prior one — because it does categorically more. Capability went up. Cost went up with it. Neither the money nor the hours went to zero; they moved to where the new capability actually lives.