Competitive intel with Grok Bot

Build a cloud-scheduled competitive intelligence system in Grokbot for a Product team.

## Goal
Every weekday morning, automatically gather competitive signals about a defined watchlist, filter hard for material news only, dedupe so nothing posts twice, and post a short digest to a Slack channel — even if my laptop is off.

## Inputs I will provide (or ask me for these first)
- Home company name (for framing only; do NOT include home-company news in the competitor digest)
- Competitor watchlist: [COMPETITOR_1], [COMPETITOR_2], [COMPETITOR_3], … (paid / material peers only — exclude freeware / open-source-only noise unless I say otherwise)
- Slack destination channel name (and resolve its channel ID)
- Preferred daily run time / timezone (default: weekdays 8:00 AM local)
- Lookback window (default: 14 days)

## Architecture (multi-agent — do not do gather + judge + post in one agent)
1. Orchestrator agent
   - Owns schedule, shared inbox, SQLite dedupe, Slack posting, and coordinating teammates
2. One collector agent PER competitor
   - ONLY gathers raw candidates
   - NEVER classifies materiality
   - NEVER posts to Slack
3. One Signal Classifier / vetting agent
   - ONLY classifies and formats keepers
   - NEVER scrapes
   - NEVER posts to Slack

### Collector responsibilities (each competitor)
For the lookback window, gather raw candidates from ALL of these (web search is first-class, not optional):
- LinkedIn company page posts via Apify (prefer harvestapi/linkedin-company-posts, no cookies; use date filters)
- LinkedIn key-people posts via Apify (prefer harvestapi/linkedin-profile-posts) — CEO/founder/CTO/notable PM, not only the company page
- General web/news search (launches, funding, pricing, M&A, layoffs)
- Product changelogs / blogs / app store / site pages as relevant
- WARN notices for layoffs
- EDGAR for material filings when the company (or parent) is public; otherwise treat as N/A privately — do NOT put empty WARN/EDGAR lines in the Slack post

For each competitor, resolve and store:
- Official LinkedIn company URL
- 1–3 key people LinkedIn profile URLs

Tag each candidate with: competitor, date, short raw summary, source_url, source_type
(linkedin_company | linkedin_profile | web | warn | edgar | changelog | other).

Write candidates to a shared inbox JSON path the orchestrator can read.

### Classifier responsibilities (high bar)
KEEP only:
- Major product launch or category-changing feature
- Pricing/packaging change that affects buyers
- Whale / named enterprise logo win (not generic “we signed a customer”)
- Funding, acquisition, sale, strategic partnership
- Layoffs, org reset, C-level hire/exit
- Rebrand, rename, positioning pivot

DROP always:
- Minor features, polish, bugfix, version-bump fluff
- Webinars, events, hackathon sponsorships, giveaways, marketing CTAs
- Generic hiring / culture posts
- Recycled talk about old launches
- Soft thought leadership with no business fact
- Non-whale customer anecdotes
- Duplicate amplifies of the same announcement

Output:
1) Kept bullets exactly: **Competitor — takeaway** · one sentence · [source](url)
2) Counts: Material: N · Filtered out: M
3) Short drop list (competitor + 1-line reason)
Never invent signals. Never narrate empty sources (“no WARN found”) in the Slack digest.

## Slack digest format
Post to the chosen channel.

Visual structure:
- Large title with emoji: 📡 Daily competitive digest — YYYY-MM-DD
- Bullets ONLY for material hits (no empty-source filler, no “Competitor X had nothing”)
- Quiet footer: Material: N · Filtered out: M
- If zero NEW material items after dedupe: skip the Slack post entirely (stay quiet)

Prefer Slack Block Kit (header + sections + context footer) if available; if the connector is markdown-only, use a # header + italic footer.
Note: posting via the Cursor Slack connector may show “sent via Cursor.” If I need a custom Slack bot identity instead, set that up as a separate path.

## Deduplication (critical)
Do NOT use Slack channel history as the dedupe source of truth.
Use a durable SQLite DB on the agent computer, e.g.:
- posted_signals(fingerprint UNIQUE, competitor, takeaway, source_url, posted_at, digest_date, slack_channel_id)
- run_stats(digest_date, material_count, filtered_count, posted_count)
Fingerprint = hash(canonical source_url + competitor + takeaway).
Each run: gather → classify → drop fingerprints already in DB → post only net-new → write new rows.

## Schedule
Create a Grokbot routine on weekdays at the agreed time.
Orchestrator kicks all collector agents → waits for handoffs → sends combined inbox to classifier → dedupes → posts if needed → writes SQLite → notifies me with the Slack link only when something posted.

## Setup order
1. Connect Slack and Apify
2. Confirm watchlist + Slack channel
3. Create 1 collector per competitor + 1 classifier with the responsibilities above
4. Stand up SQLite dedupe store
5. Dry-run once in chat
6. Post a test digest to Slack
7. Enable the weekday routine

## Success criteria
- End-to-end works with laptop off (cloud routine)
- Multi-source collection including LinkedIn people + web, not company LinkedIn alone
- High-bar filtering with Material/Filtered counts
- No duplicate Slack posts across days
- Succinct, source-linked bullets only