---

name: logo-get description: Fetch verified company logo images for one or more companies in parallel. Searches the company's own website, then DuckDuckGo image search, then Wikipedia infobox; downloads candidates, visually verifies each with the host LLM (correct identity, current variant, accurate brand colors), and saves the winners to disk. Use when the user wants to download, fetch, grab, pull, or get logos for one or more companies. Invoked as /logo-get <company1>, <company2>, ... [--out <path>]. argument-hint: ", , ... [--out ]" user-invocable: true

Logo Get

You are about to fetch verified company logos for: $ARGUMENTS.

You are a foreman dispatching parallel sub-agents — one per company. Each sub-agent independently searches, downloads, verifies, and saves a single logo. You coordinate, batch user prompts, and produce a final summary table.

Argument Parsing

Parse $ARGUMENTS as follows:

1. Detect optional --out <path> flag (anywhere in the string). Strip it; resolve <path> to an absolute path. Default: ./logos/ in the current working directory.
2. Treat the remainder as a comma-separated list of company names. Trim whitespace from each entry. Empty entries are skipped.
3. If zero company names parse out, ask the user for input via AskUserQuestion and stop.

Create the output directory if it doesn't exist (mkdir -p).

Step 1 — Disambiguation Pass

Before dispatching any sub-agents, scan the company list with your own judgment:

• For each name, decide: is there a clearly-dominant commercial entity? (e.g. "apple" → Apple Inc, "google" → Google LLC, "stripe" → Stripe Inc — yes. "acme" → no. "delta" → ambiguous: Delta Air Lines, Delta Faucet, Delta Dental.)
• For genuinely-ambiguous names, batch them into ONE AskUserQuestion call with one question per ambiguous name. For each, present 3–5 plausible options + an "other (specify)" choice.
• Use the user's qualified answers as the canonical company name for the rest of the run.

Skip this step entirely if no names are ambiguous.

Step 2 — Dispatch Sub-Agents (parallel waves of 8)

Dispatch one Agent per company in parallel waves of 8 max (single message with up to 8 Agent tool calls). Wait for the wave to complete before dispatching the next.

Each Agent call uses subagent_type: general-purpose and gets a self-contained prompt — see Sub-Agent Prompt Template below. Interpolate {{COMPANY}}, {{OUT_DIR}}, {{MODE}} per call.

Track which sub-agent corresponds to which company so you can attribute results in the final summary.

Step 3 — Aggregate Results

Each sub-agent returns a structured report:

status: ok | needs_soft_fallback | failed
path:   <absolute path to saved file, or null>
note:   <≤120 chars: tier hit, dimensions, format, variant — OR — reason for failure>

Categorize by status. If any are needs_soft_fallback, ask the user once via AskUserQuestion:

Strict verification failed for: <company list>. Show weaker matches for these (current-ish logos with approximately correct colors, possibly icon-only)? Pick which to retry.

For each name the user approves, dispatch a fresh sub-agent in soft mode (parallel wave, cap 8). Merge results.

Step 4 — Final Summary

Print to the user:

Saved to: <absolute output directory>

apple      ✓  apple.svg              (Tier 3 Wikipedia, SVG)
google     ✓  google.png             (Tier 1 website, 800×272 PNG, wordmark)
acme-corp  ⚠  needs soft fallback    (only outdated logo found)  → user declined
zorp       ✗  no qualifying logo found

---

Sub-Agent Prompt Template

Pass the following self-contained prompt to each sub-agent. Interpolate {{COMPANY}} (qualified company name), {{OUT_DIR}} (absolute output dir), {{SLUG}} (lowercased, spaces and & → -, other punctuation stripped, but preserve . for .ai/.io domains, e.g. JP Morgan → jp-morgan, Adopt.ai → adopt-ai), {{MODE}} (strict or soft):

Your job: fetch ONE verified logo for the company **{{COMPANY}}**.

Output directory (absolute): {{OUT_DIR}}
Create it if it doesn't exist.

Mode: {{MODE}}

WORKFLOW (try tiers in order, stop at first verified success):

**Tier 1 — Company website header / press kit**
1. Determine the official website URL using your own knowledge (e.g. "Apple" → https://www.apple.com, "JP Morgan" → https://www.jpmorganchase.com).
2. WebFetch the homepage. Look for the primary logo in the HTML — usually inside <header>, <nav>, or an element with class/id containing "logo". Resolve relative URLs to absolute.
3. Many companies link to a press kit, brand assets, or media page (look for /press, /brand, /media, /about, /trust, footer links). Follow them — they often have the highest-quality wordmark.
4. Download with: curl -L -A "Mozilla/5.0" -o /tmp/{{SLUG}}_candidate_N.<ext> "<url>"  (use the right extension based on URL or content-type)
5. Check size: sips -g pixelWidth -g pixelHeight <file>. REJECT if either dimension < 100. Skip the size check for SVG (vector, infinitely scalable).
6. If the homepage WebFetch fails or no logo is found, do ONE retry: WebSearch for "{{COMPANY}} official site", grab the top result's domain, retry from step 2.

**Tier 2 — DuckDuckGo image search**
1. WebFetch https://duckduckgo.com/?q={{URL-encoded COMPANY}}+logo&iax=images&ia=images
2. If the first query returns weak results, try variants: "{{COMPANY}} logo png wordmark", "{{COMPANY}} brand assets", "{{COMPANY}} press kit logo".
3. Extract up to 5 candidate image URLs. Download, size-check, verify each.

**Tier 3 — Wikipedia infobox**
1. WebFetch https://en.wikipedia.org/wiki/{{COMPANY with spaces as underscores}}
2. If the page is a disambiguation page or 404, try variants: `{{Company}}_(company)`, `{{Company}}_Inc`.
3. Locate the infobox image (typically the first image inside `.infobox`). Download, size-check.

**VERIFICATION** — after each successful download, use the `Read` tool to view the image (you can see images directly).

**Strict mode** — REJECT the candidate if any of these fail:
- **Identity**: unmistakably {{COMPANY}}'s logo (not a competitor, fan art, meme, or a different company with a similar name).
- **Currency**: the current logo per your knowledge cutoff (not a clearly-outdated version).
- **Color fidelity**: brand colors match what you know.
- **Variant**: prefer the FULL WORDMARK or LOCKUP (icon + text) over icon-only. If only icon-only is findable in strict mode, return `needs_soft_fallback` rather than `ok`.

**Soft mode** — REJECT only if:
- Identity is wrong (different company, fan art, meme).
- Tolerate outdated-but-recognizable variants, minor color shifts (JPEG compression, lighting), and icon-only when no wordmark exists.

If verified: rename `/tmp/{{SLUG}}_candidate_N.<ext>` to `{{OUT_DIR}}/{{SLUG}}.<ext>`. If a file with that name already exists, increment: `{{SLUG}}-2.<ext>`, `{{SLUG}}-3.<ext>`, etc.

REPORT — return ONE message in exactly this format (no extra prose):

status: ok | needs_soft_fallback | failed
path:   <absolute path to saved file, or null>
note:   <≤120 chars: tier hit, dimensions, format, variant — OR — reason for failure>

Rules:
- Don't retry transient failures (timeouts, 5xx) — fall through to the next tier.
- Never save an unverified file.
- "No API" = no auth-required / paid APIs (e.g. Clearbit's logo API, Brandfetch's commercial API). Incidental hits on logo CDNs and aggregators (logo.dev, seeklogo.com, worldvectorlogo.com, etc.) discovered via image search are fine — they're just hosted images, not API calls.
- needs_soft_fallback = candidates exist but failed strict (icon-only when wordmark expected, outdated, color-shifted).
- failed = no candidates at all across all tiers.

---

Notes on Behavior

• Concurrency cap is real: do not exceed 8 sub-agents in flight at once.
• Don't over-fetch: if Tier 1 succeeds, the sub-agent stops there and does not try Tiers 2 or 3.
• Soft fallback is opt-in: never run soft mode without explicit user approval.
• Don't fabricate: if no logo can be verified, return failed honestly with a useful note.