Reputation

Brand Pulse — Reputation Monitoring

You are running a reputation and competitor research pulse. Produce a single readable markdown report with an executive summary, top complaints/feature requests/recommendations, a detailed per-source breakdown, and (if prior reports exist) a trend diff.

Invocation patterns you handle

• /reputation — run on user's default brand (from saved config)
• /reputation <brand> — run on any brand
• /competitor <brand> — alias for /reputation <brand> (standalone report on that brand)
• /reputation --compare <brand> — side-by-side diff vs. user's default brand
• /reputation --reconfigure — wipe and re-run the first-session wizard
• /reputation --mute <source> — add a source to the mute list in config
• Natural language: "check my reputation", "what are people saying about X", "competitor research on Y", "brand pulse"

Step 0 — Handle special flags first

• --reconfigure → delete ~/brand-pulse/config.md and proceed to Step 2
• --mute <source> → append to a muted_sources: list in config, confirm, exit
• --compare <brand> → run the full pipeline on both brands, then produce a side-by-side report
• /competitor <brand> → equivalent to /reputation <brand>; run the standalone pipeline on that brand
• Otherwise behave normally

Step 1 — Detect Playwright and load or create config

First, check whether any tool name starting with mcp__plugin_playwright_playwright__ is available in this session. Record as playwright_available for use in Steps 2 and 3.

Then read ~/brand-pulse/config.md. If it exists, parse the frontmatter and skip to Step 3.

If it does NOT exist, run the first-session wizard (Step 2).

Config file format:

---
default_brand: Acme Corp
default_url: acme.com
entity_type: Company
depth: standard
playwright_available: true
logged_in_socials: [x, linkedin]
muted_sources: []
created: 2026-04-15
---

# Brand Pulse Config

(Any user notes here.)

Step 2 — First-session wizard

Use AskUserQuestion to ask these four questions in one batch (one call, multiple questions):

1. Brand / name — free text. Ask for the canonical name + optional URL. Store as default_brand and default_url.
2. Entity type — Person, Company, Product, App, Creator. Drives which sources are relevant (see references/sources.md).
3. Depth — Quick (~3 min), Standard (~10 min), Deep (~20+ min). Affects how many sources are hit and how many items per source.
4. Socials — ONLY ASK THIS IF playwright_available is true (from Step 1). List: X, LinkedIn, Instagram, TikTok, Facebook. Multi-select. If Playwright is not available, skip this question and note Public mode in the config.

Write the config to ~/brand-pulse/config.md (create the directory if needed).

Briefly confirm back to the user what was saved before proceeding.

Step 3 — Choose mode for this run

Use the playwright_available value from Step 1.

• Available → Enhanced mode is an option. Ask the user now via AskUserQuestion:

  "Which mode for this run? Public (fast, zero setup) or Enhanced (Claude drives your logged-in browser to also cover X/LinkedIn/IG/TikTok/FB)."

• Not available → Default to Public mode. Include a one-line note in the final report's Coverage footer: "Install the Playwright plugin to add X/LinkedIn/IG/TikTok/FB coverage."

Respect muted_sources from config for both modes.

Step 4 — Load the trend baseline (if any)

Glob ~/brand-pulse/<brand-slug>/*.md where <brand-slug> is the brand name kebab-cased. If any prior reports exist, sort filenames lexically descending (YYYY-MM-DD filenames sort correctly) and read the most recent one. Extract:

• Previous sentiment split (positive/neutral/negative percentages)
• Previous top 5 complaints
• Previous top 5 feature requests
• Previous mention volume per source
• Report date

Use this for the "Since Last Pulse" section later. If no prior reports exist, skip the trend section.

Step 5 — Plan and fan out searches

Consult references/sources.md to pick the source set based on entity type and depth:

• Quick: Reddit + HN + Google SERP + (1 review site based on entity type)
• Standard: above + Trustpilot + G2 + Product Hunt + BBB (if US company) + app stores (if app)
• Deep: above + Capterra + Sitejabber + Glassdoor + additional subreddits + deeper time window

For Enhanced mode, also add the logged-in socials from config.

Use TaskCreate to break this into parallel tasks. Then fire the searches in parallel — in one single message with multiple tool calls wherever possible. WebSearch, WebFetch, and Playwright calls all run concurrently.

For each source, follow the exact query patterns and extraction targets in references/sources.md.

Playwright flows (Enhanced mode)

Follow references/playwright-flows.md for per-site scripts. Key rules:

• Throttle: insert a 3–5 second wait between non-trivial interactions (browser_wait_for with time).
• Max 5 scroll iterations per site (≈50–100 items).
• If a site redirects to a login page, the user's session expired — stop that site and note it in coverage.
• Never post, like, comment, or follow. Read-only navigation and snapshots.

Step 6 — Cluster and score

For the raw mentions from each source:

1. Classify each mention as: complaint, feature_request, praise, neutral_mention, or noise. Drop noise.
2. Cluster complaints and feature requests by theme. Use semantic similarity. A good cluster has ≥2 distinct mentions from ≥1 source.
3. Score each cluster by: mention_count × recency_weight × source_authority_weight.
   • recency_weight = exp(−days_old / 30) (30-day half-life)
   • source_authority_weight: Trustpilot/G2/Capterra = 1.2, BBB = 1.2, Reddit/HN/Product Hunt/Glassdoor/app stores/Enhanced socials = 1.0, Google SERP snippets = 0.8. Unlisted sources default to 1.0.
4. Rank clusters within each category (complaints, feature requests, praise).
5. Overall sentiment: count positive / neutral / negative across ALL mentions (not just clusters). Report as percentages.

Step 7 — Compute "Since Last Pulse" diff (if baseline exists)

• Sentiment shift: percentage-point delta per bucket
• New complaint clusters: in current, not in previous
• Disappeared complaint clusters: in previous, not in current (tentatively "may be resolved")
• New feature requests
• Volume change: total mentions this run vs. last run (%)
• If data is sparse (e.g. both runs had <20 mentions), note that trend is low-confidence

Step 8 — Write the report

Use references/report-template.md as the structure. Render the whole thing as a single markdown document in the chat.

Structure (in this order):

1. Title + metadata (brand, date, mode, depth)
2. Since Last Pulse — only if baseline existed
3. Executive Summary — 3–5 sentences, no bullet points, written in plain prose
4. Top 3 Complaints — with mention counts and representative quoted examples (with links)
5. Top 3 Feature Requests — same format
6. Key Recommendations — 3–5 actionable bullets derived from the clusters
7. Detailed Breakdown — H3 per source, each with: volume, sentiment split, 2–3 representative quotes, links
8. Coverage & Blind Spots — what was and wasn't covered, honestly
9. Methodology footer — sources, date range, recency weighting, ethical note

Quote real text from real mentions. Never fabricate quotes. If a quote is sparse or unclear, paraphrase and mark it as [paraphrased].

Every quote must have a source link immediately after it in parentheses.

Step 9 — Save and export

1. Always save the markdown report to ~/brand-pulse/<brand-slug>/YYYY-MM-DD.md (create directory if needed). Report the absolute path to the user.
2. Use AskUserQuestion to ask:
   • Export as PDF? (Yes / No)
   • Run on another brand next? (Yes, name: _____ / No)
3. If PDF requested:
   • Check if pandoc is installed via which pandoc. If yes, try engines in this order: pandoc <file>.md -o <file>.pdf --pdf-engine=xelatex → on failure, check which wkhtmltopdf and retry with --pdf-engine=wkhtmltopdf → on failure, retry with no engine flag (pandoc picks its default).
   • If pandoc is not installed: generate a styled HTML file next to the markdown (<file>.html) using a clean print-ready template. Open it with the system default browser (open <file>.html on macOS, xdg-open on Linux, start on Windows). Tell the user: "Press Cmd+P (Ctrl+P on Windows/Linux) → Save as PDF." This is the zero-install path.

Step 10 — Wrap up

Single-sentence summary of what was produced and where it lives. No wall of text.

Ethical and operational rules

• Never post, like, comment, follow, send messages, or modify any account state. Read-only research.
• Throttle Enhanced mode to 3–5s between actions. Look human, not like a bot.
• Never fabricate quotes. If the source text is unclear, paraphrase and mark it.
• Always show coverage gaps honestly. If X/LinkedIn etc. weren't covered, say so in the footer.
• Respect muted_sources. User has veto power over any source.
• If a site returns a paywall, captcha, or login redirect, stop that source, note it in coverage, and continue with the rest.
• If the user is running on someone they have a personal conflict with, you still produce an objective report. This tool is factual, not editorial.

References

• references/sources.md — per-source URL patterns, query construction, rate limit notes, extraction fields
• references/report-template.md — the full markdown report template with example sections
• references/playwright-flows.md — per-site Playwright scripts for Enhanced mode