Skip to content

Latest commit

 

History

History
144 lines (104 loc) · 10.4 KB

File metadata and controls

144 lines (104 loc) · 10.4 KB

Skill Evaluation — browser-debug-agent

Evaluated: 2026-08-19 Source: /Users/ddurzo/Development/ai/browser-debug-agent-skill Evaluator: skill-evaluation v2.1.0 Framework: Anthropic Skill Best Practices + Matt Pocock's writing-great-skills

Summary

Metric Value
Overall Score 92.32/100
Grade A
Category product-verification
Invocation model-invoked
Files 13 source files evaluated
Criteria scored / N/A 16 scored, 2 N/A

Scorecard

Axis 1 — Trigger

# Criterion Weight Score Notes
1 Invocation design 2x 90/100 Automatic invocation fits browser-visible implementation and debugging work; the three exits make autonomous execution useful (SKILL.md:3, SKILL.md:8). Trigger eval could not be run because this local skill is not registered in the current agent catalog.
2 Description quality 2x 88/100 Starts with action verbs, names concrete browser-debug branches, and states the existing-tool boundary without enumerating backend syntax (SKILL.md:3). It remains moderately long because several adjacent failure classes must route here.

Axis 2 — Structure

# Criterion Weight Score Notes
3 Steps vs. reference clarity 1x 94/100 Six ordered loop steps are separate from runtime and safety reference material (SKILL.md:34-105).
4 Branch-aware disclosure & pointers 2x 95/100 Backend detail is required before the first browser command; escalation, visual QA, and token-economy references have precise conditional pointers (SKILL.md:32, SKILL.md:88, SKILL.md:103, SKILL.md:111).
5 Conciseness 2x 95/100 Entrypoint is 129 lines, down from 377, while backend and failure-specific detail lives in four references (SKILL.md; references/).
6 Coherent scope 1x 92/100 Primary action is browser product verification and repair (SKILL.md:8-17). Installation, upgrade, CI, and V2 direction stay outside the loaded entrypoint (README.md:17-93, .github/workflows/validate.yml).

Axis 3 — Steering

# Criterion Weight Score Notes
7 Leading words 2x 90/100 “Evidence-led”, “first divergence”, and the three uppercase exits recur as compact process anchors (SKILL.md:3, SKILL.md:8, SKILL.md:55, SKILL.md:121-127).
8 Completion criteria & legwork 2x 96/100 Every operational step ends with an observable completion statement and the final definition accounts for passed, unrun, and blocked checks (SKILL.md:40, SKILL.md:51, SKILL.md:74, SKILL.md:82, SKILL.md:105, SKILL.md:121-127).
9 Gotchas section 2x 85/100 Failure traps are explicit—stale refs, forced clicks, arbitrary sleeps, backend mismatch, primary-profile attachment, and secret leakage—although they are co-located with their rules instead of under one Gotchas heading (SKILL.md:74, SKILL.md:80, SKILL.md:113-119).
10 Grounded in expertise 2x 95/100 Guidance reflects actual snapshot invalidation, Playwright runner/agent-CLI separation, CDP endpoint risk, renderer boundaries, regression radius, and geometry checks (references/browser-drivers.md:46-92, references/debug-loop.md:17-92, references/visual-qa.md:18-88).
11 Avoids railroading 1x 94/100 The skill defines task-fit defaults and capability boundaries while leaving native project commands, exact browser CLI, and regression radius adaptable (SKILL.md:23-32, SKILL.md:92-105).

Axis 4 — Pruning

# Criterion Weight Score Notes
12 No-ops (deletion test) 2x 90/100 Most sentences change execution or stopping behavior. Remaining low-value candidates are presentation guidance in the final-response sentence (SKILL.md:129) and some README restatement (README.md:11-15), neither affects the core loop materially.
13 Single source of truth 1x 89/100 Exit definitions now live only in SKILL.md:121-127; backend facts live in references/browser-drivers.md. The evidence hierarchy remains intentionally summarized in both SKILL.md:64-72 and references/token-economy.md:3-15, leaving minor duplication.
14 Relevance & sediment 1x 96/100 Obsolete assumed CLI shapes and the playwright-core false-positive detector were removed; current runtime names are isolated behind help probes (SKILL.md:21-32, scripts/detect-browser-backend.sh:1-83). Packaging-only material remains outside SKILL.md.

Conditional criteria

# Criterion Weight Score Notes
15 Setup flow 1x 97/100 Scored by judgment because this product-verification skill clearly benefits from portable setup. The flow covers ecosystem install, seven local hosts, global/project scope, no-op comparison, remote update, recoverable backup, and foreign-target refusal (README.md:17-93, scripts/install.sh:14-299).
16 Memory mechanism 1x N/A The evidence ledger is task-local diagnostic state, not durable business/runbook memory (SKILL.md:53-62).
17 Scripts & libraries 1x 98/100 Detector and installer have smoke coverage for JSON escaping, precedence, idempotence, global/local upgrade, backup, foreign collision, and forced recovery; CI runs them on every push/PR (tests/detector-smoke.sh, tests/install-smoke.sh:16-74, .github/workflows/validate.yml:10-23).
18 On-demand hooks 1x N/A No hook can enforce browser execution portably across the supported hosts; explicit validation commands are a better fit (README.md:116-123).

Trigger Eval

Skipped — sub-agents are available, but the target is a local project and is not registered in the current agent skill catalog. Running prompts without discoverability would measure installation state rather than description quality.

Prompts tested

# Prompt Expected Triggered Other skills
1 Fix the settings form: Save does nothing in the local app and I need you to verify the repair in a browser. should-trigger not run N/A
2 Reproduce this flaky checkout click and use console/network evidence to find the cause. should-trigger not run N/A
3 The mobile header overlaps the search field at 390px; patch it and prove the layout is correct. should-trigger not run N/A
4 Run the existing Playwright failure, fix the source, and rerun the exact browser scenario. should-trigger not run N/A
5 Check why this local route is blank even though the dev server returns 200. should-trigger not run N/A
6 Write unit tests for this pure date formatter. should-not-trigger not run N/A
7 Review this API schema for breaking changes without running the frontend. should-not-trigger not run N/A
8 Summarize this public web page for me. should-not-trigger not run N/A
9 Create a Figma component library for these tokens. should-not-trigger not run N/A
10 Optimize this SQL query and explain its execution plan. should-not-trigger not run N/A

Results

Metric Value
Should-trigger hit rate not measured
Should-not-trigger leak rate not measured
Other skills observed none

Observations

Use this prompt set after publishing or installing the skill into an isolated evaluation host. Do not score a 0/5 caused solely by absent discovery metadata.

Failure Modes Detected

Mode Evidence Root cause Defense
Duplication (minor) SKILL.md:64-72; references/token-economy.md:3-15 The core evidence order is useful inline, while the reference repeats and expands it for output-heavy tasks. Keep the inline order normative; prune the reference to additions that are unique when it next changes.
Weak steering risk (unconfirmed) Trigger eval skipped; SKILL.md:3 Invocation and runtime routing are well written but not empirically exercised in a discoverable host. Run the saved positive/negative prompt set after publication and tighten only on observed misses or leaks.

Prioritized Actions

1. Run the saved trigger regression after publication

Evidence: Trigger Eval section; SKILL.md:3

Fix: Install the released skill in an isolated supported host, execute all ten prompts independently, and revise the description only if observed routing warrants it.

2. Collapse the remaining evidence-order duplication

Evidence: SKILL.md:64-72; references/token-economy.md:3-15

Fix: Keep the seven-item diagnostic order in the entrypoint and make token-economy.md contain only additional shell/output tactics.

3. Forward-test runtime routing before adding bda

Evidence: SKILL.md:23-32; README.md:129-141

Fix: Record realistic existing-Playwright, Obscura-fast-path, and Chrome-specific tasks; verify that agents choose the intended installed runtime before freezing adapter semantics.

Bonus Patterns

Pattern Status Notes
Validation loops Present Exact reproduction and regression radius are mandatory (SKILL.md:84-105).
Output templates Present Compact evidence ledger and three exit formats (SKILL.md:55-62, SKILL.md:121-127).
Procedures over declarations Present Six-step evidence-led method adapts to the repository and runtime.
Defaults over menus Present Existing project tooling first, then compatible agent CLI/Obscura/CDP (SKILL.md:23-30).
Trace-checkable steering Present “Evidence-led”, “first divergence”, and uppercase exit states are distinctive.

Grade Scale

Grade Range Meaning
A 80–100 Production-quality, reference skill
B 60–79 Good skill, minor improvements needed
C 40–59 Functional but significant gaps
D 20–39 Needs substantial rework
F 0–19 Skeleton only, not production-ready

Generated by skill-evaluation v2.1.0, merging the Anthropic skill quality framework with Matt Pocock's writing-great-skills methodology.