# Hebrish project context

Context memory for AI agents (and humans) working on Hebrish. It is public: served from the website
repository's `project-context/` folder, starting at <https://hebrish.app/project-context/README.md>.
Started 2026-10-22 from Daniel's voice note of that day; from now on, humans and agents edit it as
the project moves.

Read this folder before planning work on Hebrish. Read the file for your task, not all of them.

| File | What it holds |
| --- | --- |
| [`project.md`](project.md) | What Hebrish is, why it exists, who makes it, the principles that decide trade-offs |
| [`models-and-data.md`](models-and-data.md) | Model lines and naming, how the training data is composed, the TTS line, data collection |
| [`product.md`](product.md) | How the apps should behave and why: usage patterns, shortcuts, output modes, long form, retention |
| [`tasks.md`](tasks.md) | Work outside the apps (recording app, training, website), with status |
| [`decisions.md`](decisions.md) | Project-level decisions (models, data, licensing), dated, with their source |
| [`open-questions.md`](open-questions.md) | Questions Daniel has raised and not settled |

## Where everything else lives

| What | Where |
| --- | --- |
| **Sources** (voice notes: audio and transcript, unedited apart from marked spelling fixes) | [`Hebrish-ASR/Hebrish-App` `planning/voice-notes/`](https://github.com/Hebrish-ASR/Hebrish-App/tree/main/planning/voice-notes) |
| App requests, the per-platform sprints, ideas | [`planning/requests.md`](https://github.com/Hebrish-ASR/Hebrish-App/blob/main/planning/requests.md), [`planning/backlog.md`](https://github.com/Hebrish-ASR/Hebrish-App/blob/main/planning/backlog.md) in the same repository |
| App decisions (`D-<date>-<nn>`) | [`planning/decisions.md`](https://github.com/Hebrish-ASR/Hebrish-App/blob/main/planning/decisions.md) |
| Model and version naming, authoritative | Website repository `README.md`, "Model and version names"; <https://hebrish.app/models> |
| Voice notes for listening, with an edited transcript | <https://hebrish.app/updates> |

## Rules for editing

1. **Sources are the truth; everything here derives from them.** Cite them as
   `<source> <timestamp>`, e.g. `voice note 2026-10-22 41:04`. A source transcript is never
   rewritten; corrections go in its "Corrections after recording" section, dated.
2. **One home per fact.** App work goes in the app repository's `planning/` (agents: the
   `queue-sprint` and `log-decision` skills); this folder links to it and holds what has no other
   home: what the project is, models and data, product principles, work outside the apps.
3. **Dates are absolute** (YYYY-MM-DD). Say who decided ("Daniel, 2026-10-22").
4. **Update in place.** A task changes status (`todo`, `doing`, `done YYYY-MM-DD`,
   `dropped YYYY-MM-DD: reason`); a reversed decision is marked, not deleted; a settled question
   moves to `decisions.md` or the app decision log.
5. **Public folder.** No credential, infrastructure identifier (bucket names, account ids, file ids,
   internal hosts), personal-account model or anything Daniel keeps off hebrish.app. Those go in
   the website repository's `internal/` folder, which is not deployed.
6. **Site rules apply** (website `CLAUDE.md`): canonical romanization, model names as in
   `models-and-data.md`, no invented results, "Daniel", never "we".

Skills: `project-context` in the public `hebrish-maintainers` plugin reads and updates this folder;
the maintainer's private `voice-update` skill (plugin `asr-dev`) takes a new voice note from the
project Drive through to the site, the planning folder and this folder.