A new mode where the user explains a topic in their own words and the AI
plays a curious classmate: short reactions, exactly one follow-up question
at a time, never "correct"/"wrong", never lecturing, and everything it knows
comes from the user's own decks and notes.
- Entry points: a sidebar row next to Exam mode and a "Teach it back" action
in every deck's menu (which prefills the topic and scope).
- Start screen: deck or typed topic, friendly/skeptical student, the same
context and model chips as Chat (/teach?deck=<id> pre-selects the deck).
- Session: the student prompt is a mode of the existing /api/chat route, so
model resolution, allowlists, rate limits and the streaming client are the
same code; the personality lives in lib/prompts.ts (TEACH_*). The deck's
cards travel as a numbered checklist and the student is told to ask about
the ideas that were skipped; it must say so rather than guess when the
notes cover nothing. One constant (TEACH_MAX_EXCHANGES = 10) ends the
session, or the Finish button does.
- Recap: a teach_summary call (JSON, one repair retry) yields well / gaps
with card references / next steps; each gap links to the card it means.
"Add missed cards to review" schedules those activities due now; "Turn
gaps into cards" generates one flashcard per gap from the deck's notes and
shows an editable review list before saving.
- Sessions are chat conversations with a mode:"teach" marker + session
record: same IndexedDB history, backup and sync, filtered out of the Chat
screens; the recent list reopens a recap or continues a session.
- Reused pieces: MessageRow, Composer (now optional attach/scope chips and a
mic), Markdown/KaTeX, AiDisclaimer, ContextPicker/ModelPicker, plus a
shared useChatModels hook and useKeyboardInset extracted from Chat.
- Dictation: Web Speech API recognition (lib/speech-input.ts); the mic only
renders where the browser supports it, and dictated phrases land in the
draft.
- All strings in en + cs. teach-test: 44 checks (API prompts and guards,
both entry points, the session, recap actions, history, exchange limit,
360/390px layouts, mic on/off). Version 0.8.1, SW v103.
The third segment of the note-powered wizards (Study deck / Homework /
Podcast) takes the same intake - photos, documents, pasted text, the share
target - and turns it into a spoken episode.
- prompts: PODCAST_SYSTEM_PROMPT writes for the ear (plain prose, no markdown,
no visual references), chaptered into 3-8 sections around a length target.
- /api/podcast mirrors /api/generate's plumbing (rate limit, model gate,
provider keys, heartbeats, the 2x deadline, the optional vision pass).
- storage: frezmo.podcasts.v1, zod-validated, included in the backup and the
encrypted sync payload; a new library tab lists episodes with listen /
rename / delete-with-undo, and the wizard switch deep-links in.
- player: the transcript is word-wrapped spans and the word under the voice
rides a lyric-style highlight that keeps itself centred. Two clocks: exact
onboundary events where the browser reports them, a per-word length timer
where it does not (iOS Safari); the first boundary retires the timer.
Pause = cancel + re-speak from the word, which makes resume, chapter jumps
and tap-any-word-to-seek the same operation; the position saves on pause,
page-hide and unmount, so reopening resumes. Speed rebuilds the engine and
keeps playing; the Aero soundtrack pauses itself when the voice starts; a
browser with no voice gets an honest note and a readable transcript.
- the script sits on a glass pane (plain paragraphs over Aero's grass failed
contrast; the scan now covers the player too).
tests/podcast-test.mjs (17th suite) stubs the model and the speech engine and
drives the highlight by hand: 26 checks. All suites green, contrast clean.
Version 0.7.3, SW v95.
From your notes: the generator now stores a verbatim source quote with every
activity (schema + prompt + a server-side clamp so a chatty model cannot drop a
card), and the feedback panel shows the sentence that carries it with the quote
highlighted green. Decks made before the change, and quotes the model trimmed,
fall back to a local IDF-weighted match over the saved notes text
(lib/notes-source.ts): nothing is shown unless at least two distinct terms
land. The passage hides in exam mode with the other reveals.
Read aloud: a Web Speech API speaker control on the text a learner reads
(flashcard faces, questions, statements, sentences, instructions) via
lib/speech.ts and components/SpeakButton.tsx — one utterance app-wide, a voice
matched to the deck's language, and no button at all where the browser has no
speech synthesis. It used to sit inside the flashcard's flip button (nested
buttons, tapping it flipped the card); it now lives next to the hint row.
Timed runs: the feed chrome carries a timer chip (off by default) that gives
each card 30 seconds; a card that runs out is recorded as a miss through the
same path as a wrong answer (history + SRS agree), shows a toast, and the feed
moves straight on with a fresh budget. handleAnswered ignores a second result
per card, so a timeout plus a manual answer cannot record twice.
The generator's type spec now names what each type suits (language decks lean
on flashcards/matching/fill blank, ordering needs a real sequence, short answer
reasoning), so the deck-aware guidance covers all seven types.
New activities-test suite (24 checks: quote path, local path, the green mark,
the speech stub, the timed run on Playwright's fake clock), registered in
run-all; phase2 pins that the notes passage stays hidden in exam mode; en+cs
keys for every new string; no new dependencies. Version 0.6.1, SW v83.
UI bug from a phone screenshot (short answer, Czech):
- The textarea was flex-shrunk inside the slide's column (58px instead of
~82px), clipping the typed answer mid-line. Every direct child of a
study/review/guide slide is now shrink-0 and the slide scrolls instead.
- Two AI disclaimers were visible at once: the activity's own after the
grading box plus the feed's row. The activity-level copy is gone; the
feed/guide row (which also carries the report control) is the single one.
The phase-2 check now asserts the note sits below the graded feedback.
- The mobile suite's backdrop-tap check is deterministic too (the spring
animation made the measured box mid-flight).
Ordering activities: the deck prompt now allows them only when the notes
hold a real sequence (chronological events, process stages, procedure steps,
a lifecycle, a ranked scale) and points at matching / multiple choice /
true-false when they do not - "sort these" was landing on decks with no
order at all.
Version 0.2.4, SW -> freezmo-v57.
UI batch from feedback:
- The study/review/exam/homework chrome is now a single 58px pill: round
close button, title + position over the progress bar, chips (or the exam
timer) at the far end — everything on the centre line, so nothing crowds
the rounded corners any more. Homework guide gets the same recipe.
- The immersive feeds carry z-[45] so they paint over the app chrome (on
/exam the header used to cover the top of the chrome), and the page chrome
behind an open overlay (.app-header, .segmented, .tabbar) is hidden while
the transparent overlay is up, so it can no longer ghost through the glass.
The guide's disclaimer moved inside the overlay with it.
- Chat: the context and model chips toggle their own popover (pressed state,
aria-expanded), opening one closes the other, a press outside closes it —
no more being forced to pick something to get out.
- Homework: the model's plain-text step explanations/hints/verdicts keep
their line breaks (whitespace-pre-wrap) — a step used to collapse into one
run-on paragraph with the whole solution inside it; the guide prompt now
says the explanation is coaching, 2-4 sentences.
- The four glacier blobs are at 25% of their old opacity (0.045 light,
0.1 dark).
Tests: phase-2 chrome geometry + exam layering checks (62), chat desktop
picker checks (101), the standing suites all green. README: decisions 54-56,
decision 23 updated. SW -> freezmo-v44.
A generated Czech literature deck reached a phone with three identical
matching definitions — the card could not be solved. The pipeline only
deduped matching terms (and ordering items), so repeated definitions and
options sailed through.
- parseDeck now rejects any card whose options, items, terms or definitions
repeat (case-insensitive), through the same drop-or-repair path as the
ordering giveaway guard: the card is dropped, or a deck left with too few
cards gets one repair round with a clear reason.
- The matching dedupe covers definitions too, and both it and the ordering
one only trim when the card still meets the schema minimum afterwards —
a shorter card would fail validation on the next read and take the whole
deck down with it (latent data-loss bug).
- The generation prompt states the uniqueness rule.
- phase-1 suite: a stub Ollama returns a deck with a repeated-definition
matching card (the reported shape), a repeated-option card and a
salvageable one; asserts the first two are dropped, the third is trimmed
to three pairs and the good cards ship (43 checks).
- README: decision 51.
Server-side only (lib/ai.ts, lib/prompts.ts) — no client bundle change, no
SW bump.
Entry points (no sixth tab): Ask in the Library header (desktop button,
phone icon), Ask about this deck in the deck menu, Ask a question on
study/homework results (compact result travels as context), and a Chat
row in the desktop sidebar. Phones get a full-screen route with back +
New chat; the tab bar steps aside.
- Client-side retrieval: chunk scoped decks (~800 tokens), BM25 against
the question + recent turns, top chunks under ~7k tokens, each
labelled with deck + section. Notes come from the deck's stored
sourceText (already kept for Regenerate); decks without it fall back
to generated content and the UI says so.
- Stateless server: /api/chat validates against the CHAT_MODELS
allowlist, trims the history to ~6k tokens, caps context at ~9k and
streams token/thinking/done/error NDJSON. /api/chat/extract reduces
attachments to text through the existing vision step.
- UI: user bubbles + full-width assistant text, streaming with stop,
auto-scroll that pauses with a jump-to-bottom button, safe markdown
(React nodes, no HTML injection; KaTeX dynamically imported only when
a message has math), Thinking row for reasoning, copy/regenerate/
edit-and-resend, model + context pickers (last model remembered),
composer with visualViewport keyboard handling, safe-area, 16px
textarea, Enter-to-send on desktop.
- Usage-limit failures keep the draft and show the shared message with
Settings/Ollama actions; stopping maps the reader abort to a clean
stopped state; offline disables sending with a reason.
- Settings: Keep notes text for chat toggle; deck menu: Delete notes
text. RATE_LIMIT_CHAT + CHAT_MODELS in .env.example and README.
- Tests: chat suite (63 checks, stub Ollama) wired into run-all.
Full set: 12/12 suites, 487 checks. SW -> freezmo-v32.
Also: tests/model-compare.mjs (diagnostic) benchmarks two models on the
same notes (latency, repair rounds, activity types) so the gpt-oss vs
gemma default can be decided on numbers; defaults unchanged.
Check my work (third homework mode)
- One model call with its own system prompt marks the student's own
attempt: verdict, what's right, each issue with why + fix, and the
corrected version that keeps everything the student got right.
Unasked-for extras are never errors. The attempt is typed/pasted
(photos of handwriting are documented as unsupported) and stored in
the record so Redo re-checks the same work.
- Server: mode "check" + solution on POST /api/homework (zod, 400
server.needSolution without one) and the same NDJSON stream; new
checkHomework() pipeline with repair-once validation.
Follow-ups
- New POST /api/homework/refine (plain JSON, RATE_LIMIT_REFINE, zod,
auth before body): shorter / more detail / explain, plain text back,
language preserved.
- Shorter + More detail + Original rewrite the finished answer and the
corrected version in place (FollowUpBar); every guide step has
Explain step N differently; failed calls say so without losing text.
Turn into a study deck
- Any result hands its content (answer, corrected version or assembled
guide) to New deck through a one-shot sessionStorage seed with the
options step pre-filled.
UI
- Phone step 2 gains the third card: tapping it reveals the solution
field + submit; desktop has three cards and the solution field shows
with the mode. Homework history labels and share/card previews cover
the new mode.
Verified: new phase-3 suite 20/20 against a stub Ollama (check payload,
missing-solution 400, refine validation, phone + desktop check flows,
shorter follow-up with restore, guide explain, deck hand-off) plus
regression 74/74, mobile 52/52, declutter 116/116, wizard 80/80,
animation 8/8, native 35/35, phase1 38/38, phase2 43/43, i18n 25/25.
Key parity 734/734. Lint + build clean. SW freezmo-v27.
A short-answer case: "Why was Kryštof Harant executed?" with the student
answering "because he took part in the anti-Habsburg uprising" was failed
for not naming the year 1621 — a detail only present in the reference
answer, never asked by the question. The generator had put the date into
keyPoints, and the grader treated every key point as mandatory.
- Grading prompt: the QUESTION defines what is required. A key point
counts only when the question asks for it or the answer is wrong or
incomplete without it; extra detail in the reference answer (dates,
places, names, numbers) is a bonus — never listed as missing, never a
reason to fail. An accepted answer must return an empty "missing"
array.
- Generation spec: short_answer keyPoints must contain only the ideas
the question asks for (a "why" question requires the reason, not the
year).
- gradeShortAnswer drops any stray "missing" entries when correct is
true, so the UI can never show a gap on an accepted answer.
Verified: mock-Ollama integration test 9/9 through the real /api/grade
route (new rules sent upstream, incorrect answers keep their missing
point, accepted answers normalise to missing: []), regression 74/74,
mobile 51/51, i18n 27/27. Lint + build clean. SW freezmo-v18.
Homework (/homework — entry points in the drawer and on the create screen)
- Assignment intake reuses the study pipeline: pasted text, photos
(camera/picker, compressed in the browser) and documents (PDF/DOCX/PPTX/
ODT/text) through lib/image + lib/documents, the same vision step and the
same language guard. Raw files never leave the device — only text and JPEG
data URLs reach the server.
- Two large choice cards: “Just make it” returns the finished homework
(answer only, plain text, copy + make it again); “Explain how to make this
homework” builds an interactive guide.
- The guide is one step per screen in the study-feed snap layout:
explanation, three progressive hints (nudge -> hint -> the answer) and an
optional checkpoint rendered by the existing activity components (graded
through /api/grade). The last screen assembles the solution from the step
answers, with a copy button. Checkpoints carry xp 0 so the guide stays free
of XP/streak chrome.
- Results live in freezmo.homework.v1 with a small history (open/delete);
“Make it again” re-runs from the stored assignment text (the vision
transcription when photos were used) — no original files are stored.
- The answer and the guide are written in the assignment's language, never
translated into the UI language.
Server
- POST /api/homework: NDJSON stage events like /api/generate, the same
http-guard, auth checked before parsing, and its own RATE_LIMIT_HOMEWORK
(documented in .env.example + the README table).
- New prompts in lib/prompts.ts and zod schemas in lib/schemas.ts with the
same single repair retry; TEXT_MODEL for content, VISION_MODEL only for
photos. New shared client pieces: useNotesIntake, lib/homework store,
lib/clipboard, and a shared NDJSON reader in lib/client-api.
Also finishing the mobile polish pass from the previous request
- Phone ice scene: the top-left orb moved behind the glass header (it sat
behind the page heading/subtitle). The settings sheet's scroll cue is now a
bare chevron inside a content fade (mask) instead of a chip over the status
banner, and the API key card's note/disclosure/actions are tightened.
- theme-color now follows the in-app theme (SSR metas still fall back to the
OS preference); apple-mobile-web-app-status-bar-style stays “default”
because iOS fixes the status-bar text colour at launch.
Verified: 78/78 homework checks (text/photo/PDF, both modes, EN+CS, dark+light,
360/390px, copy, history persistence, storage privacy, API language, axe), plus
the existing suites re-run (mobile 47/47, pixel contrast 0 below AA, i18n 27/27,
desktop regression 72/72). Lint + build clean; service worker freezmo-v6.
Ordering cards
- Generation prompt now forbids dates, years, times, step numbers and
ordinals in ordering items (spec + hard rules), and the server-side
validation drops any ordering card whose items still contain them
instead of shipping a giveaway.
True/False
- The 'false ✕' swipe label was permanently visible on the card: its
opacity curve was inverted, so the clamp at rest was 1. Flipped the
range (and tightened both ranges for the elastic drag) — each label now
only fades in while swiping in that direction.
Regenerate deck
- Results screen gains a Regenerate deck button: rebuilds the activities
from the deck's stored source notes, replaces the deck in storage
(attempts/best score reset) and starts the fresh run from slide 1.
- StudyFeed tracks the current deck so the header, export and messages use
the regenerated title/activities.
- Fixed a spurious 0% attempt recorded when a regenerated deck swapped in:
scroll anchoring held the old results position, so the feed briefly
reported 5/5 with zero answers. The scroll now resets before the new
slides render, overflow-anchor is disabled on the feed, and a run with
no answers is never recorded.
Tests: new deck-fixes suite 15/15 (TF overlay at rest/drag, regenerate
with loading state + storage reset, ordering guard drop/keep); retry
14/14, deck chrome 15/15, pwa 16/16, final-ux 12/12, security 8/8,
contrast 20/20.
Next.js 15 app that turns photos, PDFs, DOCX/PPTX and text into a swipeable
feed of AI-generated activities (flashcards, multiple choice, true/false,
fill-in-the-blank, matching, ordering, short answer) using the user's own
Ollama models.
Highlights:
- All AI calls run in server route handlers; the browser never talks to Ollama.
- Bring-your-own-key mode (default): the key lives only in the browser and is
sent per-request; server, local-daemon and shared-key modes also supported.
- Client-side document extraction (pdf.js, mammoth, jszip) with scanned-PDF
page rendering for the vision model; notes stay in their original language.
- ChatGPT-style shell: collapsible sidebar / mobile drawer, neutral token-based
light + dark themes, folder organisation, PWA (manifest, SW, offline page).
- Strict JSON validation with zod + auto-repair, rate limiting, and request
guards (cross-origin, body-size, API-key sanitising).
- setup.sh bootstraps deps, .env.local and a production build.
Gitignore excludes .env*, the generated pdf.js worker and local scratch files.
