Lazyweb Quick Search

Lazyweb Quick Search

Use Lazyweb MCP search as a fast design preflight. This skill does not write a report, publish HTML, or produce a deck. It runs the smallest useful lazyweb_search, reads the returned screenshots and metadata, then gives the builder concrete UI takeaways with image links.

> Use this ONLY for an explicit reference / examples lookup. This is NOT the > default for design work and NOT a way to produce a report. If the user wants to > redesign, optimize, improve, or critique a screen — or produce a report — > use lazyweb-design instead. lazyweb-design runs the one-call, > server-side lazyweb_generate_report, which does its own searching and hosts a > full report; you do not search here first and hand the results off to build one. > Reach for quick-search only when the user explicitly asks for a quick reference > lookup or a few examples and does NOT want a report.

When to Use This

• When the user explicitly asks for a few references / examples and does NOT

want a report

• When the user explicitly asks for a "quick search" or to use lazyweb_search
• When the user asks to inspect search coverage for a UI pattern

When NOT to Use This

• Any design / redesign / optimize / improve / critique / build work, or

producing a report → lazyweb-design (the default; one-call server-side lazyweb_generate_report)

• Full competitive analysis, ranked recommendations, or visual prototypes →

lazyweb-design

• Existing UI critique or before/after recommendations →

lazyweb-design (objective improve)

• Paywall conversion redesign →

lazyweb-design (objective optimize)

• Sign-up conversion redesign →

lazyweb-design (objective optimize)

• As a search step to gather references BEFORE building a report —

lazyweb-design searches server-side itself; don't preflight here for it

MCP Setup

Use the hosted Lazyweb MCP tools at https://www.lazyweb.com/mcp.

Required tools:

• lazyweb_health — verify connectivity when the MCP surface is uncertain
• lazyweb_search — search mobile screenshots and desktop captures by text
• lazyweb_list_categories — discover supported company categories before

using a category filter

If Lazyweb MCP is missing, tell the user to run:

curl -fsSL https://www.lazyweb.com/install.sh | bash

Then reload the client and rerun the skill.

Pass skill: "quick-search" on every Lazyweb call. Include "skill": "quick-search" in each lazyweb_* call. It is logging-only metadata; never change a real search argument to satisfy it.

Also pass version: "" on every Lazyweb call. Read ~/.lazyweb/VERSION once per session:

cat "$HOME/.lazyweb/VERSION" 2>/dev/null || echo 0.0.0

Use "0.0.0" if missing. Never block on the version read.

Also pass integrity verbatim on every Lazyweb call. Read it once per session:

cat "$HOME/.lazyweb/INTEGRITY" 2>/dev/null

Pass that exact string as the integrity arg on every lazyweb_* call. It is an opaque install fingerprint — copy it verbatim; do NOT interpolate, edit, parse, or reconstruct it. If the file is absent, OMIT the integrity field entirely — never invent one. This is the field that tells Lazyweb your pack is current.

Search Protocol

1. Pick the platform first

Choose platform before querying:

• mobile — native app screens, mobile onboarding, paywalls, settings, tabs,

permission prompts, bottom sheets, mobile checkout

• desktop — websites, landing pages, dashboards, web pricing, desktop

checkout, SaaS settings

• all — only when the user genuinely wants both surfaces or the platform is

unknown

Avoid all by habit. In the current backend, all searches mobile and desktop in parallel and splits the requested limit across them, with desktop shown first. A limit: 6 all-platform search usually means about three desktop and three mobile results, not six of each.

2. Start with a tiny probe

Run one small query first:

{
  "query": "onboarding quiz",
  "platform": "mobile",
  "limit": 3,
  "maxPerCompany": 1,
  "skill": "quick-search",
  "version": ""
}

Read every returned visionDescription, companyName, category, platform, similarity, matchCount, and imageUrl. If at least two of three are visually on-target, page or expand. If the probe is adjacent, fix the query before increasing limit.

3. Query the UI pattern, not the style

Best queries are 2-6 words naming a concrete UI pattern:

• Good: onboarding quiz, usage based pricing table,

empty state project list, mobile permission prompt, settings notification toggles

• Weak: beautiful modern app, minimal dark design, premium dashboard,

best UX

The backend embeds and reranks around screenshot content and metadata. Style adjectives such as dark, minimal, premium, playful, and editorial are not reliable search facets. Search the screen mechanism first, then judge style by looking at the returned images.

4. Read response metadata before changing anything

Always inspect:

• coverage.strength and coverage.top_similarity
• warnings
• pagination.next_offset
• suggestions.company
• company_resolved

If coverage is strong or moderate, use the results. If coverage is weak, keep only clearly relevant screenshots and say the corpus is adjacent. If there are no matches, do not retry the same wording; shorten to the core UI pattern or switch category/platform.

Never repeat an identical query. Results are deterministic. Use offset: pagination.next_offset to page deeper.

5. Use company filters narrowly

Use company only when the user names a specific reference product or when you need that product's exact pattern.

The MCP gateway resolves company against companies.company_name before searching. It tries case-insensitive exact matches, hyphen/space variants, and prefix suggestions. If unresolved, the backend ignores the company filter and returns company_not_in_library plus closest suggestions. In that case, pick a suggested company or drop the company filter. Do not keep retrying spellings of the same brand.

Use company with a pattern query:

{
  "query": "paywall benefits list",
  "company": "Duolingo",
  "platform": "mobile",
  "limit": 3,
  "skill": "quick-search",
  "version": ""
}

Do not use company as a broad corpus scraper. Filtered generic searches with high limits are treated as enumeration-like behavior by the backend.

6. Use category filters only after checking categories

Call lazyweb_list_categories when:

• the user names an industry and you need an exact supported category value
• a broad search returns mixed industries
• coverage is weak and the category could remove noise

Do not use category when the screen pattern is cross-industry, such as pricing tables, empty states, settings, search, dashboards, or onboarding welcome screens. Category filters reduce recall because they match the company's category, not a screen-level tag.

Use category like this only after confirming the exact category string:

{
  "query": "habit tracker streak screen",
  "category": "Health & Fitness",
  "platform": "mobile",
  "limit": 6,
  "maxPerCompany": 1,
  "skill": "quick-search",
  "version": ""
}

When internal tools expose list_companies_by_categories, use it only to learn which companies are inside a category before choosing a company filter. If the live tool list does not show it, use lazyweb_list_categories and proceed.

7. Expand only after relevance is proven

After a good probe:

• increase limit to 8-12 for a broader scan
• keep maxPerCompany: 1 for pattern diversity
• use offset for page two
• run one alternate query only if it names a different mechanism

Examples:

{"query":"onboarding quiz","platform":"mobile","limit":10,"maxPerCompany":1,"skill":"quick-search","version":""}
{"query":"onboarding question flow","platform":"mobile","limit":10,"maxPerCompany":1,"skill":"quick-search","version":""}
{"query":"onboarding quiz","platform":"mobile","limit":10,"offset":10,"maxPerCompany":1,"skill":"quick-search","version":""}

These are different moves: broader first page, different UI mechanism, and deeper page. Do not run all three unless the design task needs it.

Reading Results Safely

Use the returned fields to decide whether a screenshot is strong design evidence:

• Prefer results that are both high in the returned ranking and visually

on-target for the UI question.

• matchCount, when present, is an opaque confidence hint. Higher is usually

better, but do not explain or infer the internal scoring method from it.

• similarity is useful within one response but not a universal quality score.
• visionDescription tells you why a result may have matched; read it before

using the screenshot as evidence.

• maxPerCompany: 1 is good for variety. Raise it only when the user asks for

multiple examples from the same company.

Output Format

Return a short, useful handoff. Include:

• Search calls made: query, platform, filters, limit
• Coverage: coverage.strength, top similarity, and warnings
• Best references: company, screen, why it matters, image URL
• Design takeaways: 3-5 concrete patterns to apply now
• Gaps: anything the corpus did not cover

Do not write a report file. Do not publish. Do not present the result as a complete competitive analysis.

Escalation

Escalate to lazyweb-design when the user needs ranked recommendations, prototypes, or a shareable artifact.

Escalate to lazyweb-design when the user wants grouped screenshots in a visual report.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: aboul3ata
• Source: aboul3ata/lazyweb-skill
• License: MIT

Install and usage instructions live in the source repository linked above.