brief-glassdoor-signal

name	brief-glassdoor-signal
description	Produce a Glassdoor / employee-sentiment research brief — overall ratings, sub-ratings, tenure data with red-flag detection scoped to the relevant team, compensation data, theme analysis of top complaints and positives, recent direction inflections. Aggregate sentiment, not a recap of individual reviews. Use to gauge whether public employee discourse aligns or contradicts the company's outward image. Reads templates/research-briefs/glassdoor-signal.md and writes to <WORKING_FOLDER>/research/companies/<slug>/glassdoor-signal.md.
disable-model-invocation	false
allowed-tools	Read, Write, WebSearch, WebFetch, Bash(mkdir *), Bash(test *)

Brief — Glassdoor Signal

Aggregate sentiment, not a list of individual reviews. The point is patterns and inflections.

Inputs

$ARGUMENTS:

• Required: <company_name>.
• Optional: --slug=<slug>.
• Optional: --team=<engineering|sales|product|...> — scope tenure red flags + theme analysis to a team.

Procedure

1. Resolve config + paths

${CAREER_DATA_DIR}/config.json → WORKING_FOLDER. Output: ${WORKING_FOLDER}/research/companies/${slug}/glassdoor-signal.md.

2. Read template

${PLUGIN_ROOT}/templates/research-briefs/glassdoor-signal.md.

3. Gather

• Glassdoor company page (rating, sub-ratings, review count, CEO approval, recommend-to-friend).
• Glassdoor reviews list (latest 30 + historical sample).
• Levels.fyi / Blind / TeamBlind / Comparably for compensation cross-checks.

4. Compute headline numbers

From the company page directly. Trend: compare current rating to last 12 months if Wayback snapshots available.

5. Tenure red-flag check

• Compute median tenure overall.
• If --team provided, scope to that team's reviewers (use job-title field).
• Only flag sub-1-year medians scoped to a relevant team or to the whole company at <50 headcount. Sub-1-year medians at company-wide scale on a 5000-person org are noise.
• Quote 1–2 representative tenure-red-flag reviews if you flag.

6. Theme analysis

For the latest 30 reviews:

• Cluster negative reviews into 3–5 themes by frequency.
• Cluster positive reviews into 3–5 themes by frequency.
• For each theme, quote 1–2 representative reviewers verbatim with role + date.
• Look for patterns. Single-instance complaints ≠ themes.

7. Inflection detection

If reviews show a clear before/after — leadership change, layoffs, acquisition — note it with one quote on each side.

8. Compensation data

Pull stratified data (role × geography) from Levels / Glassdoor / Blind. Always cite source. Self-reports get "(self-reported)" tag.

9. Write output + update CRM

Append +glassdoor to the row.

Companion delegation

If social-feedback companion is installed, cross-check on Reddit / HN / Blind — social-feedback:check-discourse <company> employee sentiment. Use to validate or qualify Glassdoor themes.

Guardrails

• Don't over-index on outliers. Single 1-star review with personal grievance is not a theme.
• Tenure red flags are scoped. See decision rule above.
• No invented quotes. Every blockquote is verbatim from a real review with date.
• Self-reported comp data is tagged.
• Acknowledge low review volume (<20) explicitly — high-variance.

Failure modes

• Glassdoor page missing or fetch blocked → fall back to Indeed reviews + Comparably; mark with (non-Glassdoor source).
• Review volume too low (<10) → write ## Status: low-signal and avoid theme analysis.

Idempotency

• Re-run prompts before overwrite.
• --refresh silently overwrites.