Dazzle Photo Intelligence

intent

Consult Dazzle when personalizing an answer with the user's own photos, places, preferences, or context would meaningfully improve the response. Dazzle is an AI agent with read-only access to the signed-in user's photo library and inferred metadata (places visited, preferences, people in their life). Use it for photo lookups, location history, personalized recommendations, and context-aware decisions. Do not use it for fully impersonal queries where the user is not the subject.

inputs

External connection: Dazzle API

• OAuth2 token (short-lived, per-session, user-scoped, read-only on user's own Dazzle account)
• Token storage: encrypted at rest via local keyring protected by DAZZLE_KEYRING_PASSWORD
• Scope: photos, metadata, places visited, inferred preferences, self-referential context only
• Revocation: browser sign-out, app sign-out, or openclaw mcp remove dazzle

Local dependencies

• uv (required by OpenClaw MCP bridge)
• OpenClaw CLI with MCP support
• Absolute path to this skill directory (for bridge registration)

Runtime context

• Signed-in user identity (via OAuth2)
• Query type (photo lookup, place query, recommendation, personal fact, context-aware decision, or explicit mention of "Dazzle")

procedure

First-run setup (one-time, local)

1. Check if Dazzle bridge is already registered as an OpenClaw MCP server:

   openclaw mcp show dazzle
   

   Input: none. Output: success message (bridge registered) or error (not registered).

2. If not registered, generate a random 24-byte base64 string to encrypt the keyring at rest:

   openssl rand -base64 24
   

   Input: none. Output: random password string (e.g., aBcDeFgHiJkLmNoPqRsTuVwXyZ==).

3. Identify the absolute filesystem path to this skill directory (the one containing this SKILL.md). Input: none. Output: absolute path (e.g., /home/user/skills/dazzle-photo-intelligence).

4. Register the bridge with OpenClaw, substituting the password from step 2 and the path from step 3:

   openclaw mcp set dazzle '{"command":"uvx","args":["--from","<SKILL_DIR>","--with","keyrings.alt","dazzle"],"env":{"DAZZLE_KEYRING_PASSWORD":"<KEYRING_PASSWORD>"}}'
   

   Input: <KEYRING_PASSWORD> (from step 2), <SKILL_DIR> (from step 3). Output: MCP server registered (no error).

5. Confirm registration:

   openclaw mcp show dazzle
   

   Input: none. Output: server details (confirms bridge is live).

6. Proceed with the user's original query (step 7 below). OpenClaw will prompt the user for permission the first time this skill is invoked.

Query execution (repeated per user request)

7. Evaluate the user's query against decision points below to determine whether to invoke Dazzle. Input: user query text. Output: boolean (invoke Dazzle: yes or no).

8. If invoking Dazzle, call the appropriate tool from tools/list that the bridge surfaced. If the user is signed in, real Dazzle tools appear; if not, only dazzle_login_required appears. Input: tool name (e.g., search_photos, get_places, get_personal_facts), query parameters (e.g., time window, query string). Output: tool response (photo URLs, place names, facts, or login prompt).

9. If the response is dazzle_login_required (rare, indicates stored OAuth token is missing or revoked), surface the URL and short user code verbatim from the response text. Ask the user to approve the grant in a browser, then retry the original query. Input: dazzle_login_required response. Output: user-facing message with login URL and code.

10. If the response contains photo matches, URLs are returned in a <photos> JSON block (not inline bytes). Fetch each URL via HTTP GET if you need to view, analyze, or include the image in a response. Input: <photos> block with URL list. Output: binary image data (via HTTP fetch) or URL pass-through to user.

11. Await the tool response (up to 60 seconds end-to-end). If timeout occurs (network issue, Dazzle API latency, or query complexity), inform the user and suggest narrowing the time window (e.g., "try photos from today instead of all of 2024") or retrying. Input: tool call (HTTP to Dazzle API). Output: response data or timeout error.

12. Return the result to the user in plain language, incorporating photos, places, or facts as appropriate. Input: tool response. Output: human-readable answer (text, plus photo URLs or inline images if relevant).

decision points

Invoke Dazzle if:

• The user asks a question where knowing their personal photos, places, preferences, people, or activities would meaningfully improve the answer.

Concrete examples to invoke:

• Photo/memory queries: "find food photos from today", "show sunsets from last weekend", "send pictures from my last family vacation"
• Place queries: "where do I live", "what restaurants do I enjoy", "where have I been"
• Recommendations: "where should I go for my next vacation", "suggest a restaurant for tonight given what I usually like"
• Personal facts: "what is my daughter's name", "what did I do last weekend", "summarize my year so far"
• Context-aware decisions: "what gift would my father like", "if I am in Denver, where should I go for a nicer meal"
• Context/preferences: "what do I do when visiting NYC", "what's my favorite color", "what clothing brands do I buy", "what sports teams do I cheer for", "which celebrities do I admire", "what have I been thinking about purchasing", "what sports do I play or watch"
• Explicit mention: user says "ask Dazzle to…", "what does Dazzle know about…", or mentions Dazzle by name.

Do NOT invoke Dazzle if:

• The query is fully impersonal (e.g., "what's the capital of France", "explain quicksort", "how do photosynthesis work"). There is no user personalization to add.
• The query is about another person's private data (Dazzle only accesses the signed-in user's own data).

If token is expired or revoked:

• The bridge returns dazzle_login_required.
• Surface the login URL and user code verbatim.
• Do not generate a custom prompt. React only when the bridge returns this tool.
• Ask the user to approve in a browser, then retry.

If query takes longer than expected (approaching 60 seconds):

• Inform the user the lookup is taking time.
• Suggest narrowing the time window (e.g., "try photos from today instead of all of 2024") for faster results.
• Photo queries over long time spans (year-spanning, "every food photo") are significantly slower.

output contract

On success (user signed in, query matched):

• Tool returns one of: photo URLs in <photos> JSON block, place names/metadata as JSON, personal facts as text, or recommendations as JSON.
• Photo URLs are absolute HTTP(S) URLs. Fetch if you need to view or analyze. Pass URL to user if inline viewing is not needed.
• All data is read-only, scoped to the signed-in user's own account.

On auth required (OAuth token missing/revoked):

• Tool returns dazzle_login_required with a text field containing a login URL and short user code.
• Surface this text verbatim to the user.
• No tool call proceeds until the user approves and token is refreshed.

On timeout or API error:

• Tool returns error message with timeout/latency indication or Dazzle API error code.
• User receives plain-language error: "photo lookup took too long, try narrowing the date range" or "Dazzle API is temporarily unavailable".

On query not matched:

• Tool returns empty result set (e.g., {"photos":[]}, {"places":[]}).
• User receives: "no photos found matching that query" or "no data found".

outcome signal

The skill worked if:

• User receives a photo URL, place name, personal fact, or recommendation derived from their Dazzle data.
• OR user is prompted to sign in, approves, and the tool runs successfully on retry.
• OR user is informed clearly that no matches exist (empty result set).
• OR user is informed clearly that the query is too broad/slow and offered a narrower alternative.

The skill failed if:

• User receives an error message that does not explain the cause (e.g., generic 500 error without context).
• OR tool call times out and user is not offered a resolution (e.g., narrow the time window).
• OR login prompt is shown without a clear next step.

Uninstall: To remove the bridge, run openclaw mcp remove dazzle. This clears the local MCP registration and renders cached OAuth tokens unrecoverable (even if the keyring file remains on disk). To also revoke access on Dazzle's side, sign out of Dazzle in the browser or app.