--- name: memorytechniques-generate-test description: Generate an AI practice test for the Memory Techniques app (VŠE FPH přijímačky — matematika nebo angličtina) locally, bez volání Anthropic API. Vloží test přímo do D1 (`saved_tests` tabulka) přes `npx wrangler d1 execute`, takže se okamžitě objeví v Practice → Saved tests. Triggers on `/memorytechniques-generate-test`, `/memorytechniques-test`, "vygeneruj test", "memory techniques test", "practice test math", "practice test english", "VŠE přijímačky test". --- # Memory Techniques — Generate AI Test (local, no API spend) Patrik runs this when he wants a fresh VŠE FPH entrance exam practice test but doesn't want to burn Anthropic API credits through the deployed `/api/generate-test` endpoint. Claude generates the test JSON itself, following the exact same format + LaTeX rules the deployed prompt enforces, and inserts it straight into the D1 `saved_tests` table so the Practice UI picks it up. ## Inputs - `` — `math` or `english` (required; accept `matika`, `mat`, `en`, `aj` as aliases) - `--weak=` — optional: focus ≥60% of questions on these topics (same topic IDs as `MATH_TOPICS` / `ENGLISH_TOPICS` in `src/data/vse`) - `--include=` — optional: Patrik's free-text request for specific problem types to include. Examples: `--include="soustava x^2+y^2=5 a xy=2"`, `--include="logaritmická rovnice s podmínkami"`, `--include="parametrická úloha — dva různé kladné kořeny"`. Parse the freeform list and guarantee at least one question per named type. If the include item doesn't map cleanly to a standard topic, emit it as `soustava-rovnic`/`parametricke-ulohy`/the closest match. - `--title=` — optional override; default `"AI Test — Matematika"` / `"AI Test — English"` - `--dry-run` — write only the JSON file, skip the D1 insert. Examples: - `/memorytechniques-generate-test math` - `/memorytechniques-generate-test english --weak=phrasal-verbs,conditionals` - `/memorytechniques-generate-test matika --weak=logaritmy,kombinatorika` - `/memorytechniques-generate-test math --include="soustava kvadratická + nelineární, xy+y=6 styl"` ## Hard constraints - **NO Anthropic API call.** Claude composes the test inline as part of this skill run. The deployed `/api/generate-test` endpoint is bypassed entirely. - **Math in LaTeX inside `$...$` delimiters.** The frontend renders through `src/components/MathText.jsx` (KaTeX). See [LaTeX rules](#latex-rules-math) below. Never use Unicode math (`√`, `π`, `²`, `≤` etc.) — always wrap in `$...$`. - **D1 write via wrangler, not curl.** The prod `pages.dev` API is fine, but wrangler is the canonical path for memory-techniques data writes (matches `feedback_koudy-finance-d1-import` pattern; prod API is public here but wrangler is still deterministic). - **Repo root is `~/WebstormProjects/memory-techniques`.** All wrangler commands run from there. - **JSON only between delimiters.** When composing, Claude keeps the questions array as valid JSON — backslashes in LaTeX must be escaped as `\\` in JSON strings (so `$\frac{1}{2}$` in the rendered output is `"$\\frac{1}{2}$"` in the JSON). For SQL safety we pass the INSERT through a file, not inline. - **Difficulty floor: VŠE FPH přijímačky, NE maturita.** Patrik feedback 2026-04-24: první iterace byla moc jednoduchá ("Vypočtěte $\log_2(8) + \log_3(1/9)$" = trivial). Every math problem MUST require multiple steps or a non-obvious technique. Apply these rules: - **Q1–Q7 (5 pts):** Exponenciální/logaritmické rovnice vyžadují substituci (`$t=2^x$`) NEBO kontrolu definičního oboru (odhazování cizích kořenů). Goniometrie: počítání kořenů v intervalu, ne "spočti $\sin(\pi/6)$". Komplexní čísla: převod do polárního tvaru, Moivre, ne "spočti $|3-4i|$". Mocniny/odmocniny: tři různé exponenty se smíšenými zlomky. - **Q8–Q11 (13 pts, PŘESNĚ 4 slovní úlohy):** Vícekrokové realistické úlohy. Kombinatorika typicky "aspoň N" přes komplement; pohyb s posunutým startem (jeden vyráží dřív); směsi jako soustava dvou rovnic; procenta jako kombinace tří multiplikativních změn. Ne "dosaď $s=vt$". - **Q12–Q15 (doplňkové, součet 13 pts):** Konceptuální — lomená nerovnice s testem znamének, soustava kvadratická + nelineární, rovnice kružnice, parametrická úloha s Viètovými podmínkami, binomický koeficient pro konkrétní člen. - **Soustava rovnic (kvadratická + nelineární) je "dobrá úloha".** Patrik explicitly wants this type (např. `$x^2+y^2=5$` + `$xy=2$`, nebo `$x^2+y=7$` + `$xy+y=6$`) — zahrň ji minimálně jednou v každém matematickém testu pokud Patrik nespecifikuje jinak. Topic: `soustava-rovnic`. - **`--include=` je hard guarantee.** Když Patrik vyžádá konkrétní typ přes `--include`, MUSÍ být alespoň jedna otázka přesně toho typu v testu (i kdyby to znamenalo vytlačit jiný standard topic). ## Procedure ### Step 1 — Parse args Resolve ``: - `math`, `matika`, `mat`, `matematika` → `math` - `english`, `en`, `aj`, `anglictina` → `english` - anything else → error: `❌ Neznámý subject "". Použij math nebo english.` Resolve `--weak`: - Comma-split, trim, dedupe. Validate against the topic whitelist: - **math:** `logaritmy, exponencialni-rovnice, goniometrie, komplexni-cisla, mocniny-odmocniny, funkce-definicni-obor, posloupnosti, nerovnice, kvadraticke-rovnice, parametricke-ulohy, kombinatorika, spolecna-prace, smesi, pohyb, prumery, procenta, pomery, binomicka-veta, analyticka-geometrie` - **english:** `tenses, conditionals, passive, reported-speech, modals, gerund-infinitive, relative-clauses, articles-prepositions, inversions, subjunctive-causative, phrasal-verbs, synonyms-antonyms, collocations, business-vocab, confused-words, reading-comprehension` - Invalid topics → warn and drop, proceed with the rest. ### Step 2 — Compose the test Claude composes the test **in this conversation**. Format matches the deployed `/api/generate-test` endpoint exactly. **Re-read the canonical prompts** from `~/WebstormProjects/memory-techniques/functions/api/generate-test.js` (`buildMathPrompt` / `buildEnglishPrompt`) before composing — if Patrik has changed them, follow the newer version. The generator here MUST stay in lockstep with what `functions/api/generate-test.js` asks for. #### Math test shape ```json { "id": "ai-math-", "title": "AI Test — Matematika", "time": 3600, "questions": [ { "id": 1, "points": 5, "topic": "logaritmy", "text": "...LaTeX math in $...$...", "options": ["...", "...", "...", "...", "žádná z předchozích odpovědí není správná"], "correct": 0, "explanation": "..." } // 15 questions total: ids 1-7 (5 pts each), 8-11 (13 pts each), 12-15 (splits the remaining 13 points to hit 100 total) ] } ``` **Scoring distribution (must hit 100 pts):** 1–7 × 5 = 35, 8–11 × 13 = 52, 12–15 must sum to 13. Typical split: 12=3, 13=3, 14=3, 15=4 (or any split summing to 13). #### English test shape — REAL VŠE FPH formát (2026-04-25 update) **Canonical reference (always re-read before composing):** `~/.claude/skills/memorytechniques-generate-test/reference/vse-fph-english-real-exam.pdf` — scan reálných příjímaček (verze A0, A1, A2, A4, A6) s Patrikovými rukopisnými poznámkami. Use Read tool with `pages: "1-10"` na začátku každé english generation. **Reálný formát (NE to, co měl prompt dřív):** - **50 otázek total**, vše v JEDNÉ posloupnosti — žádné `sections` array, žádné Part 1/Part 2/Part 3. - **Každá otázka = jedna anglická věta s prázdným místem (`.....`)**, fill-in-the-blank style. - **4 možnosti** (A/B/C/D), **NE 5** — žádná "none of the above" pseudo-option. - **Žádné reading comprehension passages.** Žádný `passage` field. - **British English textbook standard** — explicitně řečeno v hlavičce každé verze: "Only standard written textbook British English will be accepted". - Témata smíchaná v jedné posloupnosti (NE seskupené po blocích): tenses, conditionals, modals, phrasal verbs, prepositions, articles, quantifiers, reported speech, inversions, question tags, causative (have/get sth done), wish + past, gerund/infinitive, relative clauses, comparatives, idioms, phrasal verbs, vocabulary nuance, confused words. - Úroveň: B1/B2 (upper-intermediate), často "trick" otázky kde rozdíl mezi A a B je jemný (např. `for` vs `since`, `to take` vs `taking`, `would have` vs `had`). **Příklady reálných otázek z PDF (use as style anchor, NOT verbatim copy):** - "Look ..... ! There's a car coming!" → A/for B/through C/out D/after (correct: C) - "I'll try to get in touch with him but he's ..... ever at home when I phone." → A/hardly B/occasionally C/rarely D/almost (correct: A) - "By the time you arrive ....." → A/he left B/he'll have left C/he leaves D/he'll leave (correct: B — future perfect after time clause) - "While the builders were repairing the roof they ..... the bathroom window." → A/broke B/was broken C/break D/have broken (correct: A — past simple in past continuous context) - "It will cost a lot of money to have ....." → A/made that work B/that work done C/done that work D/that work made (correct: B — causative) - "I am increasingly optimistic that we can ..... it changed." → A/get B/let C/be allowed D/had (correct: A — get sth done) **JSON shape:** ```json { "id": "ai-en-", "title": "AI Test — English (VŠE FPH)", "time": 2700, "questions": [ { "id": 1, "topic": "phrasal-verbs", "text": "Look ..... ! There's a car coming!", "options": ["for", "through", "out", "after"], "correct": 2, "explanation": "Phrasal verb 'look out' = pozor, dej si bacha. Varovné zvolání před nebezpečím. (Britská angličtina, učebnicový tvar.)" }, { "id": 2, "topic": "tenses", "text": "By the time you arrive .....", "options": ["he left", "he'll have left", "he leaves", "he'll come"], "correct": 1, "explanation": "Po časové větě 'by the time' v budoucnosti se používá future perfect, protože akce odejde dřív, než se ten druhý stihne dostavit." } // ... 50 questions, all same shape, NO sections, NO passages ] } ``` **Topic IDs (English):** tenses, conditionals, passive, reported-speech, modals, gerund-infinitive, relative-clauses, articles-prepositions, inversions, question-tags, causative, phrasal-verbs, prepositions, quantifiers, comparatives, synonyms-antonyms, collocations, confused-words, idioms, vocabulary-nuance. **English questions don't need LaTeX.** Plain text fill-in-the-blank. Use `.....` (5 dots) for the gap to match the real exam visual. **Distribuce témat (target ratios for 50 Q):** - Tenses: 8–10 - Modals + causative + wish: 5–7 - Conditionals + inversions + reported speech: 4–6 - Phrasal verbs + prepositions: 8–10 - Articles + quantifiers + comparatives: 4–6 - Gerund/infinitive + relative clauses + question tags: 5–7 - Vocabulary (collocations, confused words, idioms, nuance): 8–10 **Vysvětlení v češtině** zůstává — krátké, ukáže pravidlo (např. "Po 'deny' následuje gerundium." nebo "'Hardly ever' = skoro nikdy, jediná možnost která dává smysl v negativním kontextu."). ### Step 3 — LaTeX rules (math) Every math expression MUST be inside `$...$`. Obyčejný český text mimo `$...$`. | Concept | LaTeX (inside `$...$`) | JSON escape | |---|---|---| | Zlomek | `\frac{a}{b}` | `"$\\frac{a}{b}$"` | | Odmocnina | `\sqrt{x}`, `\sqrt[3]{x}` | `"$\\sqrt{x}$"` | | Mocnina | `x^{2}`, `2^{n+1}` | `"$x^{2}$"` | | Index | `a_{n}`, `\log_{2}(x)` | `"$a_{n}$"` | | Goniometrie | `\sin(x)`, `\cos(\alpha)`, `\tan(\frac{\pi}{4})` | `"$\\sin(x)$"` | | Konstanty | `\pi`, `\infty`, `e` | `"$\\pi$"` | | Nerovnosti | `\leq`, `\geq`, `\neq`, `\approx` | `"$\\leq$"` | | Sety / intervaly | `\langle 0; 1 \rangle`, `x \in \mathbb{R}` | `"$x \\in \\mathbb{R}$"` | | Komplexní | `\|z\| = \sqrt{a^2 + b^2}`, `z = r(\cos\varphi + i\sin\varphi)` | `"$\|z\| = \\sqrt{a^2+b^2}$"` | | Suma / limita | `\sum_{k=1}^{n} k`, `\lim_{x \to 0}` | `"$\\sum_{k=1}^{n} k$"` | | Kombinační číslo | `\binom{n}{k}` | `"$\\binom{n}{k}$"` | **JSON escape reminder:** Each `\` in LaTeX becomes `\\` in the JSON string. When you compose the test and then `JSON.stringify` mentally / explicitly, verify that backslashes are doubled. LaTeX also belongs in `options` and `explanation`, not just `text`. ### Step 4 — Write JSON + SQL to tmp files 1. Generate `test_id` = `ai--` (matches the convention used by `generateTest()` in `src/components/practice/PracticeDashboard.jsx`). 2. Write the composed test JSON to `/tmp/mt-test-.json` (pretty-printed, for Patrik to eyeball if needed). 3. Write SQL to `/tmp/mt-test-.sql`: ```sql INSERT OR IGNORE INTO saved_tests (test_id, subject, title, questions, source) VALUES ( '', '', '', '<JSON.stringify(questions) with single-quotes escaped as ''>', 'ai' ); ``` **Escape rule:** SQLite strings escape `'` as `''`. Use `python3 -c` or Bun to build the SQL — don't hand-escape large JSON. Suggested one-liner: ```bash bun -e 'const t=JSON.parse(await Bun.file("/tmp/mt-test-<id>.json").text()); const sql=`INSERT OR IGNORE INTO saved_tests (test_id, subject, title, questions, source) VALUES (${[t.id, t.subject||"<subject>", t.title, JSON.stringify(t.questions), "ai"].map(v=>`'${String(v).replace(/'"'"'/g,"'"'"''"'"'")}'`).join(", ")});`; await Bun.write("/tmp/mt-test-<id>.sql", sql);' ``` Simpler fallback (Python, always available on macOS): ```bash python3 - <<'PY' import json, sys test = json.load(open('/tmp/mt-test-<test_id>.json')) q = json.dumps(test['questions'], ensure_ascii=False).replace("'", "''") title = test['title'].replace("'", "''") sql = f"INSERT OR IGNORE INTO saved_tests (test_id, subject, title, questions, source) VALUES ('{test['id']}', '<subject>', '{title}', '{q}', 'ai');" open('/tmp/mt-test-<test_id>.sql', 'w').write(sql) PY ``` ### Step 5 — Insert into D1 (remote) From `~/WebstormProjects/memory-techniques`: ```bash npx wrangler d1 execute memory-techniques-words --remote --file=/tmp/mt-test-<test_id>.sql ``` On failure: - **"D1_ERROR: UNIQUE constraint failed"** → `test_id` already exists; regenerate with a new timestamp and retry. - **"no such table: saved_tests"** → migration `0010_saved_tests.sql` not applied in remote. Prompt Patrik: `⚠ Migration 0010 not applied in remote D1. Run: npx wrangler d1 migrations apply memory-techniques-words --remote` - **Auth error** → prompt: `⚠ Wrangler auth expired. Run: npx wrangler login` ### Step 6 — Output receipt Print in this format (no Discord DM — this is a local action, not a team one): ``` ✅ AI test vygenerovaný a vložený do D1 (bez volání API) Subject: <subject> Test ID: <test_id> Title: <title> Questions: <count> Weak topics: <comma list or "—"> File: /tmp/mt-test-<test_id>.json Otevři Practice: https://memory-techniques.pages.dev → Procvičování → <Matematika | English> → Uložené testy ``` Do NOT paste the full test JSON into chat — it's ~12k tokens. If Patrik wants to see it, offer to cat the file. ## Edge cases | Case | Behavior | |---|---| | Patrik passes no subject | Ask: `Math nebo english? (math / english)`. Wait for answer, then proceed. Don't guess. | | `--weak` has topics from the wrong subject | Drop silently with a one-line warning; proceed with valid topics. | | Claude generates fewer questions than expected (math<15 or english<50) | Stop. Don't write. Print `⚠ Test generation incomplete (<N>/<target> questions). Retry.` | | JSON has invalid LaTeX (KaTeX would throw) | MathText.jsx catches with `throwOnError: false`, so the raw `$...$` would render as plain text. Before insert, optional: validate by rendering through a node KaTeX check — for now, rely on Patrik's eye. | | Patrik passes an existing `test_id` via `--id=` flag (not standard) | Not supported; always generate a fresh id. | | Wrangler not found | `⚠ wrangler not installed. Run: bun install (in memory-techniques repo) — it's a devDependency.` | | Remote D1 write hits Cloudflare rate limit | One retry after 5 s, then surface the error. | | Patrik wants a test *without* saving (dry run) | Support `--dry-run`: skip Step 5, only write JSON file, print path. | ## What this skill does NOT do - Does not POST to `/api/generate-test` (the whole point is to skip it). - Does not hit any LLM API — Claude IS the generator. - Does not touch static tests in `src/data/vse/` — those stay authored. - Does not deploy anything. D1 write is the only side effect. - Does not send a Discord DM — local dev action, not team-visible.