What to give the agent
Snapdown is read by the agent, not controlled by it. There is no MCP server or local API: an agent cannot start a capture or query your library. You hand the evidence over from the Library:
- Copy Markdown: choose it on the bundle's row and paste the result into the agent chat. The image links are absolute paths into your vault, so they include your Windows user name, and only an agent running on the same computer can open the images.
- Copy image: copy one finding's image, with or without the annotations burned in, and paste it into a chat that accepts images.
Copy one bundle for one coherent review pass. The Markdown carries one section per finding, its image link, its notes, and one numbered line per marker with the marker's position.
The bundle should answer four questions without a follow-up: what is visible, where is it visible, what should be true instead, and how can the agent reproduce it? Numbered markers make “where” unambiguous.
A repeatable review loop
Reproduce before you capture
Start from a known route, viewport, scaling level, account state, test data, or feature flag. Put those conditions in the finding, not only in your memory.
Capture the relevant region and blur what is sensitive
Use the pixel loupe for exact boundaries. Blur credentials, API keys, personal information, customer content, and anything else that does not belong in an agent conversation.
Write a checkable finding
Use a numbered marker and write the current behavior, the expected behavior, and the reproduction facts. Prefer visible statements over guesses about the implementation.
Hand over the bundle and ask for a bounded action
Paste the copied Markdown and ask the agent to open the images it links to, locate the responsible UI, propose or implement a minimal fix, and report how it verified the result.
Recapture after the change
Review the new state with a fresh finding. A before and after pair explains a later regression better than “looks fixed”.
Prompt templates
Claude Code
Paste the copied Markdown after a short request with a verification boundary:
The Markdown below is a Snapdown bundle. Open the images it links to. Reproduce the conditions listed in the bundle. For each finding: 1. identify the responsible component and stylesheet, 2. explain the smallest safe fix, 3. implement it, 4. verify the result at the stated viewport and display scale. Do not change unrelated layout. Tell me which files changed and what remains unverified. (paste the copied Markdown here)
Cursor and Windsurf
Paste the bundle into the workspace chat and ask the agent to ground each conclusion in a finding and marker number:
Use the Snapdown bundle below as the visual acceptance record. Open its images before editing. Resolve markers 1 and 2 of Finding 1 only. For each fix, cite the finding and marker number, name the source file, and run the most relevant local check. If the image does not prove a proposed change, ask a specific question instead of guessing.
For a larger bundle, limit the delivery so the task does not turn into a broad redesign:
Treat the numbered marker notes in the Snapdown bundle below as acceptance criteria. Implement the minimum change needed to satisfy the markers of Finding 1 and Finding 2, preserve existing behavior outside those states, and summarize: reproduction used, files changed, validation performed, and any assumption that needs human confirmation.
Token economics: region first, budget second
Image-token cost differs between models and agent products, and their accounting can change, so this page gives no percentage. The comparison that holds is structural: a full-screen 4K capture contains large areas an agent does not need; a region capture plus a short, anchored finding makes the relevant detail explicit.
| Hand-off | What the agent must infer | Snapdown alternative |
|---|---|---|
| One full-screen capture with “fix this” | Which panel, state, and pixel relationship matter. | Capture the relevant region, add marker 1, and state the expected result. |
| Several unlabeled images | Which image belongs to which complaint and whether they are related. | Copy one bundle whose Markdown numbers each finding and links each image. |
| Raw high-resolution image for a small UI defect | How to ignore most of the frame. | Start with a tight region; let the image size preset keep only useful detail. |
| A text bug report without a screenshot | The actual visual state, hierarchy, and spacing. | Attach the annotated evidence and describe the reproduction condition in text. |
Use a full-screen or full-page capture when global context is the defect: page hierarchy, a cross-panel relationship, an unexpected modal, or overflow across the whole viewport. Even then, anchor the local issue with markers and avoid sending other monitors, private notifications, or credentials.
A reproducible visual issue record
This bundle format is deliberately explicit. It works for a human teammate and for an agent because it separates evidence from instruction.
# Checkout review 2026-09-24 10:15 ## Finding 1  ### Notes Checkout at 1440 × 900, Windows 11, browser zoom 100%. Route: /checkout. State: cart contains one physical item; signed-in test account; feature flag checkout_v2 on. Build: local branch feature/checkout-summary at commit <commit-id>. ### Marker Notes 1. [loc:62%,78%] The primary button overlaps the order-total row. Expected: 16 px of visible separation and no overlap at this viewport. 2. [loc:30%,54%] The discount label is truncated after "promotional". Expected: label wraps or expands without hiding its value.
Put the request in your prompt, not in the finding: locate the responsible layout rules, apply the smallest fix for markers 1 and 2, verify this exact route, state, viewport, zoom, and display scale, and report the files changed with a fresh post-fix capture.
Do not put API keys, real customer details, session cookies, or production-only URLs in the reproduction block. Use a test account, placeholders, and blurred images. Snapdown makes sensitive values in the image unreadable; the note needs the same care from you.