Audit Agent Experience Skill#79
Conversation
Spawns parallel Claude subagents against a target docs/SDK/SKILL.md from a one-sentence prompt, captures structured traces, and renders a graded HTML report scoring Setup Friction, Speed, Efficiency, Error Recovery, and Doc Quality. Includes narrative cross-agent review to surface convergent hallucinations and silent workarounds the JSON self-report misses. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
|
@cursor review |
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 3 potential issues.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 8f64258. Configure here.
| "Vertical automation — fintech, insurance, healthcare, legal, supply chain, GTM" | ||
| ], | ||
| "researched_at": "2026-04-23" | ||
| } |
There was a problem hiding this comment.
Unrelated file committed in wrong PR
Medium Severity
skills/event-prospecting/profiles/browserbase.json is unrelated to the Audit Agent Experience skill being added in this PR. It belongs to a separate event-prospecting skill (which has no SKILL.md or other files). It contains Browserbase-specific competitive intelligence (competitors, customers, use cases) that doesn't belong in this changeset. This is especially notable given the SKILL.md explicitly instructs "Never pre-fill a product, URL, or company" and "Ignore environment signals."
Reviewed by Cursor Bugbot for commit 8f64258. Configure here.
| .dimension-bar-fill.grade-c { background: var(--grade-c); } | ||
| .dimension-bar-fill.grade-d { background: var(--grade-d); } | ||
| .dimension-bar-fill.grade-f { background: var(--grade-f); } | ||
| .dimension-rationale { font-size: 0.875rem; color: var(--muted); } |
There was a problem hiding this comment.
Dead CSS for explicitly-disabled dimension breakdown section
Low Severity
Thirteen CSS rules for .dimension, .dimension-head, .dimension-bar-fill, etc. are defined in the template but are unused dead code. The SKILL.md explicitly states "do not render a per-dimension breakdown section" and no corresponding HTML elements or placeholders exist in the template body. This adds unnecessary weight to every generated report.
Reviewed by Cursor Bugbot for commit 8f64258. Configure here.
| Classification heuristic: | ||
| 1. If the same trace earlier contained a successful `llms.txt` WebFetch whose output mentioned this URL → `FROM LLMS.TXT` | ||
| 2. Else if the same trace earlier contained any WebFetch/Bash output that mentioned this exact URL → `FROM PREV PAGE` | ||
| 3. Else if the subsequent tool_result has `err: true` with 404 content → `GUESS · 404` |
There was a problem hiding this comment.
Field name mismatch in URL provenance classification
Medium Severity
The URL provenance classification heuristic in SKILL.md checks for err: true on tool_result events, but the trace schema defined in subagent-brief.md instructs subagents to use error: true. This field name mismatch means the orchestrator would fail to detect 404'd URLs, misclassifying GUESS · 404 results as TRAINING PRIOR — undermining the most valuable doc-discoverability signal in the report.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 8f64258. Configure here.
1 participant
Spawns parallel Claude subagents against a target docs/SDK/SKILL.md from a one-sentence prompt, captures structured traces, and renders a graded HTML report scoring Setup Friction, Speed, Efficiency, Error Recovery, and Doc Quality. Includes narrative cross-agent review to surface convergent hallucinations and silent workarounds the JSON self-report misses.
Note
Low Risk
Primarily adds new skill documentation/templates and a static prospecting profile, with no executable code changes beyond what future skill runners may follow.
Overview
Introduces a new
audit-agent-experienceskill, including a detailedSKILL.mdplaybook for running parallel subagent onboarding audits (config prompts, credential-handling guidance, trace parsing, scoring rubric, and report generation flow).Adds supporting assets for the skill: an HTML report template (
assets/report-template.html) plus reference docs (references/*) defining prompt variants, subagent trace schema/brief, and scoring rules, and includes a new MITLICENSE.txt. Also adds aBrowserbaseprospecting profile JSON underskills/event-prospecting/profiles/.Reviewed by Cursor Bugbot for commit 8f64258. Bugbot is set up for automated code reviews on this repo. Configure here.