2026-09-22 (session 2) — Build run: thirteen conjunctions and modal verbs
This was the second automated session of the day on the TKG English Learner's Dictionary, an original English-English dictionary for language learners built as one JSON file per word ("entry") and published as a website. Each session is unattended and picks its own task from a set of modes; this one was assigned "build" — drafting new entries from the word queue.
What the session did
Pre-flight was clean: no leftover pull request, no abandoned work branches, nothing waiting in the owner's inbox, and the local copy was already up to date with the previous session's work. The automatic word-claiming tool, asked for the next thirteen words by the queue's own ordering (most common words first, grouped by part of speech), happened to land exactly on the last two closed categories of "function words" — small grammatical words rather than content words — that the dictionary had not yet finished: the five remaining conjunctions (when, where, whether, while, yet — words that join clauses) and all eight modal verbs (could, may, might, ought, shall, should, will, would — the helping verbs that express ability, permission, possibility, and obligation). This was a good batch to take as a whole, since finishing an entire grammatical category in one pass avoids leaving loose ends for a later session.
Each entry went through the required process: writing definitions, examples, and grammar notes from my own knowledge (never copied from a published dictionary); confirming that these words take no inflected forms, which is correct for this part of speech; checking pronunciation against a panel of other AI models plus a public pronunciation reference (all 26 transcriptions across the 13 words came back fully agreed, no disputes); and a second review by two more AI models reading every field of every entry against a checklist. That review raised 72 issues, 40 serious enough to block publication before I read and decided each one myself against the entry and the house style guide. I agreed with and fixed 64; I rejected 5 where the reviewer was simply wrong — one flagged the word "polite" as an invalid label when it is in fact one of the dictionary's approved options, another wanted a contraction like shan't hand-marked in an example when the website already marks such forms automatically.
The interesting finding
A clear pattern ran through many of the genuine fixes: modal verbs like will and would are full of small, real exceptions to their own general rules (for example, will can appear in a clause where it normally shouldn't, specifically to show someone is being stubbornly willing to do something), and my first drafts stated the general rule too absolutely, without naming the exception. I softened each of those claims rather than deleting the useful general rule. Separately, twice I had written a short historical aside — that a word was "originally the past tense of" another — inside a field meant for present-day usage advice; both reviewers correctly flagged this, since word-history claims belong only in a dedicated field that requires checking two outside sources first, which I had not done. I removed both asides rather than trying to verify them under time pressure. Finally, twelve definitions used a word that sounds ordinary but turns out to fall outside the roughly 2,400-word list the dictionary limits its own definitions to (words like "permission," "contrast," and "invitation"); I reworded each definition with an already-approved word instead of creating a new entry mid-session.
What this means and what's next
The dictionary grew from 273 to 286 finished, published-quality entries, completing two entire grammatical categories. I also noticed and wrote up, without fixing, a data problem: eleven common conjunctions (although, as, but, if, nor, once, or, since, so, though, unless) are stuck in the word queue marked as "already being worked on" by no session that actually exists, which silently hides them from the automatic word-picking tool. This needs a small tool fix in a future session; I left a note explaining exactly what happened so that fix doesn't have to be rediscovered from scratch. The scheduler's own next pick is an "originality" check — a periodic search of the web to confirm the dictionary's wording isn't accidentally matching a published dictionary's phrasing.
What needs the owner
Nothing new to decide immediately. The three previously open items are unchanged: turning on the public website (GitHub Pages), and two small style questions already answered with a working assumption in the meantime. The ten pronunciation and word-classification questions already waiting for the owner are unchanged. Total spending today: $0.47, well under the $5.00 daily budget, all of it in this session (the previous session made no paid calls).