DocsOverview
Direction
What shoots is trying to be, what is being worked on, and what it will not become. This page is intentionally short on dates: it records intent, and intent that has not shipped is not a promise.
Current release: 0.7.*. The tool is pre-1.0 — the commands below are in daily use, but the shape of a young one can still change between minor versions. What each release asks of you is in the migration notes.
Direction change. Until launch,
shootstargets Lightroom Classic only, through a plugin and a local app it opens. The interactive shell and every other editor are paused. The plugin reads and writes the Lightroom catalog directly; the CLI on its own keeps writing XMP sidecars, so it stays usable without the plugin.
Design goals
These are the constraints every feature is measured against.
| Goal | What it means in practice |
|---|---|
| Nothing is destroyed | Originals are never deleted or overwritten. Edits go through the Lightroom catalog, rejects are moved and not removed, every mutating command accepts --dry-run. |
| Everything runs locally | No cloud inference. The models ship to ~/.shoots and run on your machine, on your photographs. Nothing leaves it without an explicit opt-in, and even then only derived measurements — never an image. |
| The CLI stays free | Every command documented here stays free to run. The commercial product is the Lightroom Classic plugin built on top of it. |
| Scriptable before interactive | Every command speaks --json and returns meaningful exit codes. The plugin and its local app are fronts for the CLI, never the only way to do something. |
| An orchestration layer, not an editor | shoots sits before and after Lightroom and automates the tedium around it. The plugin is the interface. |
| Lightroom first | One host, done properly, before any other. The engine keeps its adapter interface, so other editors stay additive later. |
| Honest about what it knows | A model that cannot support a prediction says so, and reaches only as far as its evidence carries it. |
Where it stands
| Area | State |
|---|---|
import, rename, exif |
Stable. The template engine and the metadata layer are settled. |
cull, rate, triage |
Stable. Focus-aware culling and CLIP rating, with review. First candidates for the plugin. |
Preference learning — embeddings, match |
Working end to end: duels in, a personal rating profile out. |
Develop predictor — develop |
Usable. Every parameter states whether it is predicted, leveled or constant, and none ships with negative held-out skill. More progress now waits on data, not code. |
Pipelines, schedule |
Recent. The surface may still gain steps — there are no conditionals yet. |
setup, doctor, update, release-notes |
Stable. Provisioning, self-update and the migration notes that apply to your machine. |
cache |
cull, rate, embeddings and develop export share what they derive, so a re-run costs nothing. |
| Shell | Being removed. The plugin replaces it; every command it fronted stays on the CLI. |
| Test suite | Every package carries unit tests (packages/*/test), run together by npm test. |
Next
In order:
- Lightroom Classic plugin. The fast operations first — culling, rating, picks and labels written straight into the catalog, no sidecars.
- A local app the plugin opens. Training, calibration and review in the
browser, served on your machine: opening it from Lightroom starts the server,
closing it stops it. Built on
@shoots/ui, a shared design system (tokens, dark/light themes, Preact components) meant for every shoots web surface. Cull verdicts are reviewed there, and every one can be reversed: this replaces the shell'scull --review. - Send to Lightroom, and back. Results and decisions made in the app go straight into the catalog through the plugin, as one undo step: no sidecars, nothing to re-read. Flags, stars and labels changed in Lightroom show up in the app within seconds; the latest change wins.
- Train from the catalog. The plugin exports your develop settings as JSON, so training through the plugin does not depend on XMP sidecars being up to date. Training and calibration start from the app and run on their own: they finish even if Lightroom closes. The standalone CLI keeps training from sidecars.
- Develop from Lightroom, and learn back. The plugin applies the profile's predictions to the selection as one undo step; once you have finished the photos, the app compares what you kept with what was predicted and improves the profile from the difference.
- Develop with intents. Before applying, you tick what the shoot needs — preserve highlights, open shadows, save the sky, expose correctly — and pick a detail profile (B&W, warm, …): measured as how a subset of your edits departs from the general profile and applied on top of it, never on its own. Intents are explicit targets a tone solver meets frame by frame; region intents find the sky and the subject with a local vision model (Florence-2, MIT). Semantic features did not beat the histogram as predictor inputs on one photographer's catalog; they are re-measured on beta data first. Keywords come from the same optional vision pass; the CLI side is built, the app screens and the Lightroom field test of intent strength are next.
- A closed beta that shares data. Testers who opt in share derived measurements, never images, so the develop predictor can learn from many photographers instead of one. After the beta, sharing becomes an explicit opt-in in the plugin, off by default.
Editors
| Editor | State |
|---|---|
| Lightroom Classic | The focus — via the plugin, reading and writing the catalog. The standalone CLI keeps its XMP adapter. |
| RapidRAW | Paused — the adapter works end to end, but is not maintained until after launch. |
| darktable, RawTherapee, ON1 Photo RAW, Capture One | Paused until after launch. |
The engine speaks to an editor through an adapter, and that interface stays: develop settings do not transfer between editors, and the adapter is the only place that has to know. Adding an editor later is additive work, not a rewrite.
Later
Candidates, not commitments. None of these has been started.
- More rating profiles for genres the five built-ins (
street,generic,portrait,wildlife,wedding) do not cover. - Richer pipeline steps — conditionals, and reuse of a pipeline as a step.
- Selection help beyond focus: near-duplicate grouping, eyes-closed and expression checks on the frames culling leaves behind.
- Other editors, in the order demand sets.
Non-goals
Things that would make shoots a different tool, and are therefore out of scope
whatever the demand:
- Becoming an editor. No local masks, no generative fill, no retouching. The deliverable is the best starting point to refine, not a finished image.
- Becoming a DAM. No catalog database of its own, no asset manager. The local app runs jobs; Lightroom owns the library.
- Cloud inference. Sending photographs to a third party for a better model is not a trade this tool will make. An account exists for usage analytics and opt-in data sharing — never as a gate in front of a prediction.
- Editing in place. No feature will ever earn the right to overwrite an original.
Disagree with any of it, or need something that is not here? Open an issue.