.tag.shock

No sign-up required to use the Free local tier. Open .tag.shock in your browser, enter your Anthropic API key in Settings, and you're ready to drop in a folder. No software to install, no account to create. If you want cloud processing across devices, create a free account for the Starter plan. Paid plans add monthly cloud-run allocations so you don't need your own API key.

.tag.shock uses Anthropic's prompt caching. The system instructions (the keyword-generation prompt) are cached after the first image in a session — you only pay tokens for the image itself on every subsequent call. In practice: the first image costs roughly $0.006–$0.010 at current Claude Haiku pricing; each subsequent cached image costs roughly $0.002–$0.004. A 200-image batch typically costs well under $1 in API fees. The Free local face-matching modes (CLIP + landmark models) have zero API cost.

A typical Generate call — title, description, 50 keywords, 9-axis score, categories and release status — returns in 10–30 seconds depending on image complexity and current API latency. The first call in a session may be slightly longer as the prompt cache is being populated; subsequent calls in the same session are typically under 15 seconds. Generate all processes images sequentially, so a 200-image folder takes approximately 30–50 minutes end to end.

.tag.shock reads JPEG, TIFF, PNG and all major RAW formats natively: NEF (Nikon), ARW / SRF / SR2 (Sony), CR3 / CR2 (Canon), DNG (Adobe universal RAW), RAF (Fujifilm), ORF (Olympus), RW2 (Panasonic) and others. No pre-conversion to JPEG is needed. Embedded XMP is written back into JPEG and TIFF; sidecar XMP files are generated for RAW and PNG formats.

One Claude API call returns: a commercial title (up to 150 characters), a platform description (up to 200 characters), up to 50 ranked keywords in 4 priority tiers, up to 3 primary and 3 secondary submission categories, a content-type flag (Photo / Illustration / Vector / Video), the 9-axis commercial sales score with a written reasoning paragraph, and a model release status flag. All in a single structured response — no second API call, no extra latency.

Claude ranks keywords into four tiers by commercial weight. Primary — the most specific, high-intent terms a buyer searches (e.g. "orange grove harvest"). Secondary — broader subject terms that widen reach (e.g. "citrus fruit", "agricultural landscape"). Supporting — descriptive context (e.g. "golden hour", "copy space", "Mediterranean"). Context — technical or supplementary tags (e.g. "horizontal", "colour", "outdoors"). The export places primary keywords at position 1 so agency ranking algorithms see the strongest signals first.

The standard export covers Shutterstock, Adobe Stock, Getty / iStock, Alamy, and Pond5. Each agency has different keyword count limits, category naming and column headers — the export engine shapes the metadata correctly for each platform so you import directly without manual editing. Embedded XMP is written to JPEG and TIFF source files; sidecar XMP is generated for all other formats including RAW.

The 9 axes are: searchability, commercial relevance, emotional resonance, technical quality, authenticity, trend alignment, market saturation, usability and licensing risk. The overall score is a submission priority rank — it tells you which frames to upload first, not whether an image will pass or fail agency technical review. Re-grade any image after metadata edits; the score refreshes in one call without re-running the full keyword pipeline.

When you load a folder, face detection runs in parallel across all images — no separate trigger. In Free mode, on-device landmark models and CLIP cosine similarity compare each detected face to the release photos in .release.shock — zero API cost. In AI mode, a multi-stage pipeline runs: CLIP embedding match first, then Claude Vision verification for ambiguous or rear-view cases. Matched release names attach to the metadata and export in the submission CSV. Unmatched faces are flagged in the gallery before export.

Yes — in AI mode. When a subject is turned away, shot from behind, or has their face occluded, the detection pipeline flags the image as "person visible, face not detected" and routes it to Claude Vision with the body-based matching prompt. Claude analyses body shape, distinctive hairstyle, clothing, posture and contextual cues against the reference photos in your .release.shock library, returns a match confidence score and a written rationale. This is especially important for lifestyle, fashion and action photography where subjects are rarely front-facing. Matches below your configured confidence threshold land in a manual review queue — nothing ships unchecked. Free mode (on-device only) requires a visible face for geometric matching; AI mode is recommended for any shoot with rear-facing or occluded subjects.

.tag.shock uses a structured three-layer prompt architecture built on microstock submission expertise. The metadata layer encodes keyword tier definitions, per-platform character limits, deduplication rules, singular/plural conventions and a 26-category taxonomy table — so the model produces agency-shaped output, not raw description. The scoring layer encodes the 9-axis rubric with weighted definitions and numeric anchors (what an 85 vs a 60 looks like on each axis), plus instructions to return a written reasoning paragraph alongside each score. The release matching layer provides a structured visual checklist including body-based identification cues for rear-facing subjects. All three layers are prompt-cached after the first image, so every subsequent image in a batch only pays tokens for the image itself — not for re-sending the instruction set.

Yes. .tag.shock works as a fully standalone stock photo keywording tool. Point it at finished JPEGs or TIFFs and run the full metadata generation, scoring and export workflow without any other Suite module installed. Release matching is optional — if you don't have .release.shock set up, .tag.shock still generates all metadata and scores, and flags images with detected but unmatched faces so you can attach releases manually.

v1 runs in your browser using your Anthropic API key — all processing stays local, no files leave except the API call itself. v2 is a hosted cloud build currently in Early Access: files are processed on dotshock.ai infrastructure, results sync across devices, and there is no API key to manage. Both versions produce the same output. v2 adds team collaboration, shared project libraries and scheduled batch processing.

The main differentiators are: (1) the 9-axis commercial sales score with a prioritised submission queue so you upload in order of likely return, (2) AI model release matching that confirms consent before export in the same workflow, and (3) batch cost efficiency — prompt caching means a 200-image folder costs under $1 in API fees. Xpiks and Microstock+ focus on metadata entry and platform sync; .tag.shock adds commercial assessment and consent management to the same single workflow. Built in the spirit of both tools; not affiliated with either.

The app shell and your previously loaded images work offline — you can browse your library, review existing results and edit metadata without a connection. Generate calls require internet access because they send the image to the Anthropic Claude API. The ArcFace local face-matching mode runs fully on-device and works offline; the Claude Vision verification step for release matching needs a connection. All data — your images, generated metadata and sales scores — stays in your browser's IndexedDB and is never uploaded to any dotshock server.

Not directly — .tag.shock generates metadata from still images. For video, extract a representative still frame (most video editors export a poster frame in one click), run it through .tag.shock to generate title, keywords and category, then copy the metadata to your video clip's submission form or CSV. The export columns for Pond5 and Shutterstock video share the same field names as the image exports, so the workflow is identical — you're just sourcing the frame manually. Native video support is on the v2 roadmap.

The batch pauses and the failed image is flagged with an error badge. Any images already generated keep their metadata — progress is saved to IndexedDB after every successful call so nothing is lost. You can top up your Anthropic credits and resume from the first failed image; the batch picks up exactly where it left off. Prompt caching means re-running a flagged image after a credit top-up doesn't re-process images that already completed.

Yes — every field is editable before export. Click any keyword chip to remove it; type in the keyword field to add your own. The title and description fields are plain text inputs. You can also drag chips to reprioritise tier order, or use the bulk override to promote a keyword to primary tier across all images in a batch. Changes are saved automatically to IndexedDB; the export reflects whatever is in the editor at the time you click Export — not the original AI output.