ADSI Agent Kit v3.0

ADSI = AI Design Slop Index. ADSI is a production-oriented product/UI quality gate for detecting observable design slop symptoms: unrealistic data, passive workflows, inconsistent status logic, weak accessibility, generic visual clichés, layout brittleness, and missing production states.

ADSI does not infer whether a UI was made by AI. It scores whether the UI behaves like a real product.

What changed in v3

ADSI v3 upgrades v2 into a broader production stack:

• Python-first ADSI core.
• Static heuristic scanner.
• Browser collector via Playwright: screenshot, DOM, computed styles.
• Contrast checker from computed styles.
• SARIF reporter for code scanning annotations.
• Before/after ADSI diff.
• MCP-compatible stdio server.
• Antigravity /adsi [do] [target] workflow.
• Agent adapters for Antigravity, Cursor, Claude Code, Codex, Hermes-style multi-agent runners, and OpenClaw.
• Optional Node Playwright companion package.
• Stage-by-stage quality audit log.
• 10 unit tests and example reports.

Install locally

cd ADSI-Agent-Kit-v3
python -m pip install -e .

Optional browser capture dependency:

python -m pip install playwright
python -m playwright install chromium

Optional Node browser collector:

cd packages/adsi-browser
npm install
npx adsi-browser-collector --url http://localhost:3000/admin --out artifacts/adsi/admin

Antigravity usage

Install into your workspace:

adsi init /path/to/repo --platform antigravity

Then in Antigravity, use:

/adsi [do] [target]

Examples:

/adsi scan dashboard admin
/adsi fix dashboard admin
/adsi collect http://localhost:3000/admin
/adsi contrast artifacts/adsi/admin/computed-styles.json
/adsi diff reports/before.json reports/after.json

The Antigravity adapter writes:

.agents/workflows/adsi.md
.agents/skills/adsi.md
.agents/skills/adsi/SKILL.md
AGENTS.md
ANTIGRAVITY.md

CLI usage

Scan code or HTML

adsi scan ./src --product "My Product" --screen "Admin Dashboard" \
  --output reports/adsi-audit.json \
  --report reports/adsi-audit.md \
  --sarif reports/adsi-audit.sarif

Capture browser artifacts

adsi collect http://localhost:3000/admin --out artifacts/adsi/admin

Outputs:

artifacts/adsi/admin/screenshot.png
artifacts/adsi/admin/dom.html
artifacts/adsi/admin/computed-styles.json
artifacts/adsi/admin/capture.json

Check contrast

adsi contrast artifacts/adsi/admin/computed-styles.json \
  --output reports/adsi-contrast.json \
  --fail-on 1

Validate audit JSON

adsi validate reports/adsi-audit.json

Score an existing audit

adsi score examples/corelasi_dashboard_audit.v3.json --breakdown

Use as a release gate

adsi gate reports/adsi-audit.json --max-score 20

Exit code is 1 when the gate fails.

Compare before/after

adsi diff reports/before.json reports/after.json \
  --output reports/adsi-diff.json \
  --fail-on-regression \
  --max-score 20

Run MCP-compatible server

adsi mcp

Exposed tools:

• adsi_scan
• adsi_score
• adsi_validate
• adsi_diff

Install agent adapters

adsi init /path/to/repo --platform all

Platform options:

• antigravity
• cursor
• claude
• codex
• openclaw
• hermes
• generic
• all

Recommended production workflow

1. Run adsi scan as deterministic first-pass audit.
2. For running apps, run adsi collect and adsi contrast.
3. Ask your coding agent to fix only top ADSI areas.
4. Run lint/typecheck/test/build.
5. Re-run adsi scan.
6. Use adsi diff to verify improvement.
7. Use adsi gate --max-score 20 before production claims.

Scoring bands

Score	Band	Release decision
0-19	low_slop	pass
20-39	mild_moderate_slop	warn
40-59	high_slop	block
60-100	severe_slop	block

Repository map

Path	Purpose
adsi/	Python package and CLI engine
schemas/	ADSI audit JSON schema
rubric/	ADSI v3 scoring rubric
rules/	Heuristic rule catalog
adapters/	Platform-specific agent instruction files
packages/adsi-browser/	Optional Node/Playwright collector
docs/V3_ARCHITECTURE_AND_PACKAGE_STRATEGY.md	Design and package decision
docs/STAGE_AUDIT_LOG.md	Stage-by-stage quality audit
prompts/	Audit and repair prompts
examples/	Example inputs and baseline audit
reports/	Example generated reports
tests/	Unit tests

Limitations

• Static scan is deterministic but not a substitute for product-context review.
• Browser collection requires Playwright.
• MCP server is a minimal stdio foundation; it intentionally avoids extra SDK dependencies.
• Contrast is computed from captured browser styles, so it is only as accurate as the rendered state you capture.