| name | vocab-deck |
| description | Build an Anki flashcard deck teaching a domain's vocabulary so a learner names and describes things precisely. Use when asked for flashcards or an Anki deck to learn terminology or visual vocabulary from a reference. |
Vocab Deck
Turn a rich reference source into a spaced-repetition deck that builds precise
recognition-and-naming vocabulary — the exact words a learner needs to describe
things, including when prompting an AI.
Route
Read only the reference for the phase you are in.
| Need | Read |
|---|
| Design the cards — types, schema, anti-leak, pruning, tagging | references/card-types.md |
| Source content and capture clean visuals from the reference | references/sourcing.md |
Orchestrate generation and package the .apkg | references/pipeline.md |
Workflow
- Scope with the user before building: the source, breadth (how many items),
concept-vs-code emphasis, and rough size. Batch choices that materially change
the deck into one interview; infer the rest.
- Source — extract one structured record per item (term, aliases,
plain-language "describe-the-thing" phrases, a tight definition) and capture
one clean visual each. →
sourcing.md
- Author a diverse card set that balances recall and understanding, with
discrimination cards for confusable siblings. Delegate the generation to a
different model family than the orchestrator. →
card-types.md
- Package to a single
.apkg with genanki in an isolated venv:
deterministic GUIDs, subdecks, embedded media, loud validation. → pipeline.md
- QA and deliver — render real cards to confirm no answer leaks and every
image resolves, prune redundant cards, then deliver with import steps and a
short "what's in / what was cut and why" note.
Hard rules
- A recall card tests one cue → the term. Never expose the answer in the prompt —
including a term label baked into a captured image.
- Cover recall AND understanding; always include discrimination cards for
confusable sets. Cut redundant and near-duplicate cards — quality over count.
- Generate card content with a different model family than the orchestrator, and
package with a separate agent; verify output against the source, not the brief.
(→
crew, agent-orchestration)
- Prefer uniform catalog/grid thumbnails over annotated detail captures. If the
source renders a concept too small or too interactive to capture cleanly,
author a crisp SVG icon rather than shipping a blurry crop.
- Package as one
.apkg via genanki in a project-local venv; derive GUIDs
deterministically so re-import updates in place; embed all media; fail loudly
on any missing file. Never hardcode time- or random-based deck/model IDs.
- Get the user's OK before installing packages or driving their browser; keep
capture visible through the authorized bridge (→
local-http-browser-use).
Stops
- Scope that changes deck composition is ambiguous → ask (batched interview),
don't guess.
- Clean visuals are unobtainable for some items → stop and offer options (author
an SVG, skip the visual, or keep the term verbal-only). Never ship blurry or
answer-leaking images.
Output
Report the deck path, import steps (Anki desktop + AnkiWeb sync + AnkiMobile),
per-deck/per-type counts, and what was included, cut, and why. Keep the source
records, media, and build script reproducible in the work directory.
