match-loop（匹配循环）

Match Loop

Run a generator ↔ visual analyst loop.

This skill is for vibe-coding where one-shot generation is not enough. The core pattern is:

• one sub-agent codes/builds the thing
• one sub-agent analyzes the code and previews the live UI like a human
• the analyst gives concrete feedback
• the generator revises
• the loop continues until the analyst accepts the result, progress stalls, a blocker appears, or the user says stop

The defining feature: the analyst must visually inspect the frontend whenever a frontend exists. It should not only read code or logs. It should open the app in a browser, take screenshots or otherwise inspect the rendered UI, interact with it, and verify that it looks and behaves correctly.

Roles

1. Generator agent

Responsible for:

• implementing the requested app, feature, page, or artifact
• fixing issues found by the analyst
• preserving working parts while revising broken/ugly/confusing parts
• running basic local checks when useful
• producing concrete revisions, not endless plans

The generator should optimize for fast, working iterations.

2. Analyst agent

Responsible for:

• inspecting the code for correctness, maintainability, obvious bugs, and missing pieces
• running or previewing the app locally when possible
• visually inspecting the frontend in a browser like a human user
• taking screenshots or using browser/computer-use tools to examine the UI
• testing basic flows and interactions
• checking whether API calls, forms, buttons, navigation, and visible state actually work
• giving prioritized, concrete feedback to the generator
• deciding whether the result is good enough for the task

The analyst is not a replacement generator. It should critique toward revision, not rewrite the whole project unless the implementation is structurally broken.

Non-negotiable visual QA rule

For frontend work, the analyst must not rely on code review alone.

It should:

1. start the dev server or use the provided preview URL
2. open the app in a browser
3. inspect the rendered page visually
4. capture screenshots when useful
5. test key interactions like a human would
6. report visual and functional defects clearly

Look for things that code review often misses:

• text cut off, overlapping, too small, too large, low contrast, or misaligned
• mobile/desktop layout problems
• broken spacing, hierarchy, or visual balance
• confusing CTAs or navigation
• forms that appear fine in code but fail in the UI
• loading/error states that look ugly or broken
• API calls that silently fail
• console errors
• buttons that do nothing

If browser-native tools fail, escalate using the local browser/computer-use stack rather than pretending visual inspection happened.

Preferred order:

1. Playwright MCP / Playwright CLI / direct Playwright or CDP
2. app/browser screenshots and DOM inspection
3. macbot / Hammerspoon / AppleScript / Peekaboo for desktop-level inspection

Do not use Safari unless the user explicitly asks. Chrome is the default browser target.

Best use cases

Use this for:

• vibe-coded web apps
• landing pages
• dashboards
• onboarding flows
• mobile-responsive pages
• React/Next/Vite/Expo web previews
• UI-heavy prototypes
• API-backed frontend apps
• code where visual behavior matters as much as implementation

Loop workflow

1. Define the target

Before spawning the loop, define:

• what is being built
• what “perfect” or “good enough” means for this task
• target user / use case
• must-have features
• visual/style expectations
• functional/API expectations
• test commands or preview commands if known

If the target is fuzzy, write the best current target and let the analyst sharpen acceptance criteria during the first review.

2. Spawn the two roles

Create:

• one generator worker
• one analyst worker

Keep the roles explicit. The parent orchestrates handoffs and convergence.

3. Generator builds v1

The generator should create a real working attempt:

• implement files
• install needed dependencies when appropriate
• run basic checks if cheap
• provide how to run/preview/test it

4. Analyst reviews v1

The analyst should perform three layers of review:

Code review

• architecture / file structure
• implementation completeness
• obvious bugs
• missing edge cases
• maintainability

Visual browser review

• launch or open the app
• inspect the rendered UI
• take screenshots when useful
• check layout, text, hierarchy, spacing, contrast, responsiveness, and polish

Functional smoke test

• click key buttons/links
• fill forms if relevant
• verify navigation and visible state changes
• check console/network/API failures when available
• run tests or small API checks if appropriate

5. Analyst produces feedback packet

The analyst feedback should include:

• what works
• what is broken or ugly
• exact visual defects observed
• exact functional/API defects observed
• screenshots or screenshot paths when useful
• prioritized changes for the next generator pass
• acceptance status: accepted, revise, or blocked

6. Generator revises

The parent sends the analyst packet to the generator.

The generator should:

• fix the highest-impact issues first
• preserve accepted parts
• avoid unnecessary rewrites
• state what changed
• provide the next preview/test instructions

7. Repeat until convergence

Continue generator → analyst → generator until:

• analyst accepts the output
• remaining issues are trivial and not worth another round
• the loop is thrashing
• the task is blocked
• the user says stop

Convergence rule

The analyst can call the result “perfect” only after:

• code review passes for the task’s scope
• the frontend was visually inspected, if applicable
• key interactions were tested
• major visual defects are resolved
• major functional/API defects are resolved

Do not chase fake perfection. If only tiny polish remains and the output is fit for purpose, stop and report caveats.

Feedback quality bar

Good analyst feedback is:

• visual when visual defects exist
• specific
• prioritized
• actionable
• tied to acceptance criteria

Good:

• "The hero text wraps badly at mobile width and overlaps the CTA. Reduce headline size on <480px, add vertical spacing, and make the CTA full-width. Screenshot: /tmp/match-loop/mobile-v2.png."

Bad:

• "Looks bad."
• "Improve UI."
• "Probably fine" without previewing.

Stop / handoff conditions

Pause or stop when:

• the user says stop
• the app cannot be launched because required credentials/services are missing
• the analyst cannot visually inspect despite reasonable browser/computer-use escalation
• revisions are no longer producing meaningful improvement
• the analyst accepts the result

Parent orchestration pattern

1. define target and acceptance criteria
2. spawn generator
3. generator builds v1
4. spawn/prompt analyst with target + code/output + preview instructions
5. analyst performs code + visual + functional review
6. parent sends feedback packet to generator
7. generator revises
8. repeat until analyst accepts or stop condition is reached

Final response expectation

At the end, summarize:

• number of rounds
• final output/version
• visual checks performed
• functional/API checks performed
• biggest fixes made
• why the analyst accepted it
• remaining caveats, if any

Suggested companion reference

If needed, also read:

• references/loop-patterns.md