handout-typos

---

name: handout-typos description: Use this skill to proofread a handout .tex file for serious objective errors only — misspellings, missing/wrong diacritics, broken agreement, missing words, repeated words, clear punctuation mistakes. Applies confident fixes directly and flags the rare genuinely ungrammatical sentence for user review. Trigger words: "proofread", "typo pass", "find typos", "spell check", "grammar check" on a handout. Languages: CS, SK, EN. Do NOT use for rephrasing, style polish, awkward-phrasing rewrites, or translation — use the handout-editor skill for that.

Handout Typo & Grammar Fixer

You proofread one olympiad math handout .tex file (PlainTeX + AMS-TeX + OPmac) for serious objective errors only — the kind a native speaker would instantly recognize as a mistake. Apply confident fixes directly with Edit. Everything else stays untouched.

Hard rule: never change the author's style or identity. Word choice, sentence rhythm, idiom, level of formality, signature turns of phrase — all off-limits. You are fixing mistakes, not editing writing. When you can hear the author in a sentence, leave it alone even if you would phrase it differently.

Higher-order hard rule: math writing is allowed to sound a bit stiff, a bit clunky, or a bit unpolished. That is not a mistake. A sentence that is slightly awkward, a word choice that isn't the most natural, a construction that a copy editor would smooth out — all of that is fine and must be left alone. Only fix what a competent native speaker would read and say "that's wrong," not "I'd phrase it differently."

---

Scope

• Operate on one .tex file at a time in data/handouts/. If the user names several, process them one by one.
• Edit prose only. Math, macro names, commands, labels, URLs, and verbatim content are off-limits.
• Language is set by \setlanguage{CS|SK|EN} in the file (and mirrored in the filename suffix .cs.tex / .sk.tex / .en.tex). Apply CS/SK/EN rules accordingly.
• Two input shapes. Default is one whole file. If the user asks to check only recent changes (a diff), identify the changed passages via git diff or recent in-session edits — and operate on those passages plus their surrounding sentence or paragraph, since agreement, case, and punctuation errors need context to detect. Passes 1–2 run over that scope, not the whole file.

---

Workflow

1. Identify the target file. If the user hasn't named one: prefer the handout created or edited in the current session (e.g. the output of a just-finished translate/edit pass) and say so. Otherwise check git status — when exactly one handout .tex is modified or added, take it as the target and say so. Otherwise ask.

2. Read the file and note the language.

3. Pass 1 — collect candidates across prose only. Sweep the High-frequency pitfalls list below explicitly — those errors parse fine and slip past a prose read. Include only serious, objective errors:

   • Misspellings and missing/wrong diacritics.
   • Broken agreement (gender/number/case), clearly wrong verb forms, a missing or doubled word that breaks the sentence.
   • Unequivocal punctuation mistakes that change parsing or are clearly wrong (not stylistic comma choices).
   • Casing errors: sentence-initial uppercase; proper names capitalized. Don't over-capitalize — in CS/SK, days, months, and language names stay lowercase unless sentence-initial.
   • Repeated words (case-insensitive): immediate duplicates separated only by whitespace, ~, or trivial punctuation (token token → token).
   • Genuinely ungrammatical sentences — word salad, missing essential particle, mangled construction that no native speaker would produce. Merely awkward, stiff, clunky, or non-idiomatic phrasing does not qualify. If a native speaker could read the sentence and say "that's an unusual way to put it, but it's fine," it stays.

4. Pass 2 — classify each candidate into one of three buckets:

   • Fix directly: objective mistake (typo, diacritic, agreement error, obvious duplicate word, sentence-initial case) where the correction is unambiguous and cosmetic.
   • Uncertain: proper noun you can't verify, a clearly ungrammatical sentence where the minimal fix is non-obvious, or a spelling/case-ending choice that hinges on meaning you can't infer. Be stingy with this list — if you're about to flag something because it "could be smoother," drop it.
   • Leave alone: author's voice / idiom / rhythm / clunkiness — not a mistake, just not how you'd write it. This is the default; when in doubt, here.

5. Apply the "fix directly" items with Edit. One fix per Edit call; include 2–3 words of surrounding context so old_string is unique. Use replace_all only when the replacement is safe in every occurrence. Batch independent Edit calls in a single message.

6. Report and resolve Uncertain. Start with a one-sentence summary of directly-applied fixes (e.g. "Fixed 3 diacritics, 1 repeated word, 1 agreement."). If nothing is Uncertain, say so in a few words and stop.

   Otherwise, for each Uncertain item print a short paragraph with: a line-number anchor (e.g. L296), the problematic clause with surrounding context, a minimal proposed fix, and a few-word reason (e.g. "missing verb", "wrong case ending", "doubled preposition"). Deduplicate — each distinct item once.

   Then call AskUserQuestion (batched up to 4 per call) with options per item: Apply suggested fix / Leave alone. Apply approved fixes via Edit and emit a final one-liner with what was applied. If you genuinely have no idea what the author meant for an item, leave it out of AskUserQuestion and ask in plain text instead.

7. No fan-out — proofread in the main session, even for long handouts (the longest are ~850 lines, well within one context). Whole-file context is part of the detection toolkit (consistency evidence, referent tracking), and slice-based subagent proofreading has shown near-zero recall on real inflection errors while producing exactly the style false positives this skill forbids. On long multi-problem handouts, keep attention sharp by running pass 1 as sequential chunks — one \Problem block (or ~150-line window) at a time — instead of one continuous skim.

Do not compile. Typo fixes in prose cannot break TeX. If the user wants a compile check afterwards, they will ask.

---

Ignore zones (never inspect or edit)

• Comments: from an unescaped % to end of line; lines starting with %; \iffalse … \fi; \begin{comment} … \end{comment}.
• Math: $…$, \(…\), $$…$$, \[…\], and math envs (equation, align, eqalign, cases, gather, matrix, pmatrix, …).
• Verbatim / code: verbatim, lstlisting, minted, alltt, \verb|…|.
• Control / IDs: any token starting with \; contents of \label{}, \ref{}, \cite{}, \url{}, \href{}, \path{}, \MathcompsLink{}, \setlanguage{}, \input, filename-like arguments.

Allowed prose wrappers (check inside the braces only)

• In math: \text{…}, \hbox{…}, \mbox{…}.
• In text: \textit{…}, \textbf{…}, \emph{…}, \textrm{…}, \textsf{…}.

Never alter the macro name or its options — only the prose inside.

---

TeX spaces

~ is a non-breaking space. Treat it as a normal space for tokenization and repeated-word detection (e.g. k~nájdenému). Never remove, insert, or flag ~ itself.

---

High-frequency pitfalls (sweep explicitly)

These errors leave the sentence parsing fine, so a prose read glides over them — check occurrences deliberately:

• SK singular inštrumentál needs í: svojím ťahom, ním, tvojím; the short forms svojim/nim are plural datives.
• Inanimate math nouns decline inanimately: SK v čitateli, v menovateli — čitateľovi is the animate form (a reader, not a numerator). CS: v čitateli, ve jmenovateli.
• CS classics: mně/mě, ji/jí, by jsme → bychom, aby jste → abyste.
• Contaminated or incomplete constructions: crosses like vyzerá byť, že (mix of vyzerá byť X and vyzerá, že); clefts missing their verb (Posledné, čo zostáva rozmyslieť, je, ako…).
• Referent slips: the same role-noun twice in one clause where the second should be the other actor (každý ťah hráča vedie do vyhrávajúcej pozície hráča → súpera). When a player/opponent noun repeats, check who is actually meant. Meaning-dependent → Uncertain, not a direct fix.
• Consistency evidence: before fixing an inflection, search the file for other occurrences — a correctly inflected form elsewhere in the same handout confirms both the slip and the fix.

---

Confidence bar for direct fixes

Apply directly only when the correction is unambiguous, local, and cosmetic. Move to Uncertain (or leave alone entirely) when any of the following apply:

• The original could be a valid but unusual form.
• It is a proper noun whose correct form you cannot verify.
• The choice depends on meaning (e.g. CS/SK i vs í, s vs z, SK mi vs my) and context is thin.
• A punctuation change could alter meaning or emphasis.
• The token is domain jargon you don't confidently know (mathematical terms in CS/SK: check twice before "fixing").
• The fix would rewrite a phrase rather than correct a token. Don't. Not directly, not in Uncertain — leave it alone.
• The phrasing is merely clunky, stiff, or unpolished, not ungrammatical. Leave it alone. Math writing is allowed to be plain, even slightly awkward. A sentence that parses and means what it should is not a mistake.

When in doubt: leave unchanged, and do not list under Uncertain.

What counts as "style" (never touch, never flag)

• Choice of connective (teda vs preto, tedy vs proto, thus vs so).
• Sentence length, comma-splice-ish rhythm the author uses consistently.
• Favorite hedges, interjections, or signature phrasings ("Všimnime si", "Zrejme", "Note that", "Clearly").
• Level of formality, first-person plural vs passive, active vs passive voice.
• Word order that parses, even if you'd reorder it.
• Non-idiomatic but grammatical word choices — e.g. a slightly unusual adjective, a translated-feeling turn of phrase that still parses.
• Stiffness, stuffiness, repetition, or a sentence that could be "smoother."

If in doubt whether something is a mistake or a voice choice, it's voice. Leave it. The bar for flagging anything at all is: a native speaker would read this and say it's wrong, not a native speaker could improve it.