Universal Web Adaptation

Universal web adaptation skill

Use this skill when the user wants a general solution across many websites, not a one-by-one manual integration first.

This skill is especially suitable for:

• unfamiliar public websites
• search-oriented workflows
• result-page collection and lightweight summaries
• deciding whether a site needs a dedicated profile or can stay generic

For the 99idea Playwright project, this skill should work together with:

• src\run-demo.js
• src\planner.js
• src\llm-planner.js
• src\site-profiles.js
• scenarios\site-profiles.json

Core principle

Always try this order:

1. Capability discovery — check if opencli already supports the site (opencli list)
2. Auth cascade — classify: PUBLIC → COOKIE/SESSION → HEADER → LOGIN-GATED
3. Generic adaptation — probe UI controls and try heuristic strategies
4. Temporary scenario tuning — adjust runtime before assuming selectors are wrong
5. Popup / opener / alternate-submit handling
6. Site profile only if generic adaptation still fails

The goal is to minimize per-site handwork.

Execution path decision (inspired by opencli)

Before touching the DOM, first determine which execution layer to use:

if opencli has an adapter for this site:
    use opencli (zero LLM cost, deterministic)
elif the site has a public API:
    use fetch (structured data, no rendering)
elif the site renders server-side (SSR):
    use web_fetch or Playwright (metadata extraction)
else:
    use Browser Agent with login-state or mark blocked

This avoids wasting time on SPA shells that will never yield results via scraping.

What “generic adaptation” means

When facing a new site, do not immediately create a profile. First try to discover:

• likely search inputs
• likely opener buttons or icons
• likely submit controls
• whether Enter works
• whether a click opens a popup or new tab
• whether the site verifies best by URL, title, or visible text

Standard workflow

1. Probe the page

Collect candidates for:

• inputs
• buttons
• anchors
• role=button
• role=search

Prefer candidates whose metadata includes:

• search
• query
• find
• 搜索
• 查找

For each candidate, record:

• visibility
• enabled/editable status
• approximate position
• container context
• whether duplicates exist in header/footer/overlay regions

2. Classify the site UI

Put the site into one of these buckets:

• direct-search: visible input plus obvious submit
• opener-search: search exists but must be opened first
• popup-search: submit opens a new tab or popup
• duplicate-controls: multiple matching controls require container qualification
• hidden-disabled: search exists but is not immediately usable
• SPA-shell: page returns JS-only content, no usable DOM without full rendering
• login-gated
• anti-bot / rate-limited

Auth cascade classification (from opencli)

Also classify the site's auth requirements:

• PUBLIC: no auth needed, fetch/web_fetch works
• COOKIE: requires browser session cookies (opencli Browser Bridge)
• HEADER: requires custom auth headers (API key, Bearer token)
• LOGIN-GATED: requires interactive login, not suitable for default regression

3. Try generic action strategies

Try these in order:

1. fill visible search input + press Enter
2. fill visible search input + click nearest likely submit
3. if search is hidden, click likely opener then retry 1 and 2
4. if click or Enter opens a popup, switch to the popup and continue there
5. verify success by:

• URL change
• expected URL parameter
• title change
• visible result container text

Prefer URL verification whenever available.

4. Use temporary runtime tuning before giving up

If the page is heavy or slow:

• use waitUntil: "domcontentloaded"
• use postNavigationLoadState: "domcontentloaded"
• add a settle delay

Do not assume selectors are wrong if the real problem is heavy page load behavior.

5. Decide whether to keep it generic or create a profile

Stay generic if:

• the search flow works with robust heuristic selectors
• the same logic is likely reusable across many sites

Create a dedicated profile only if:

• the site needs exact selectors to avoid ambiguity
• the site has duplicate controls in multiple containers
• the site needs a stable popup-specific submit selector
• the site has a unique opener pattern worth remembering

Failure classification rules

Strict mode violation

Meaning:

• your selector is too broad
• multiple header/footer/overlay controls match

Response:

• add container-qualified selectors
• prefer the visible and nearest control to the active search region

Timeout waiting for visible/editable

Meaning:

• the control is hidden
• disabled
• off-screen
• not opened yet

Response:

• find and click the opener
• retry with a container-specific visible control

Enter does nothing

Meaning:

• the page requires a dedicated submit control
• or the form uses JS-driven submission

Response:

• inspect nearest clickable sibling/container controls
• test the likely submit node

Search opens a new tab/popup

Meaning:

• current-page URL checks will fail

Response:

• switch the active page to the popup
• continue verification and screenshot on the popup page

Login / anti-bot / rate-limit

Meaning:

• the site is not a good candidate for stable default regression

Response:

• mark blocked or special
• do not force it into the generic stable matrix

Output expectations

When using this skill, the agent should produce:

1. execution path recommendation: opencli / fetch / browser — and why
2. auth cascade classification: PUBLIC / COOKIE / HEADER / LOGIN-GATED
3. the site UI classification
4. the generic strategy attempted
5. whether generic adaptation succeeded
6. whether a site profile is actually needed
7. the final stable selectors or blocker reason

Also include explicit evidence lines when possible:

• URL before/after
• title before/after
• the selector used for input and submit
• whether popup-following was required

Proven lessons from this environment

• CNN needed container-qualified selectors to avoid duplicate q inputs.
• Bilibili needed popup-following behavior because search opens search.bilibili.com in a new tab.
• BBC exposes a hidden and disabled search input, so generic probing alone is not enough yet.
• Tencent News shows a visible search button but no directly usable input in the default state.
• GitHub works, but is a low-frequency special case because of rate limits.

When to combine with other skills

• Use external-site-profile-learning when generic adaptation fails and a durable site profile is needed.
• Use content-search-summarization when the goal is to search a content platform and summarize the top results.

Practical rule of thumb

If a new site can be made to work by:

• picking the right visible input
• picking the right nearby submit
• following a popup
• using URL verification

then keep it in the generic path first.

If success depends on an exact remembered selector or a unique UI ritual, then promote it into a site profile.

Source & license

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

• Author: AllenS0104
• Source: AllenS0104/skill-browser
• License: MIT

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