Scout | She Leads AI - She Leads AI

Scout scans the open web and academic sources every morning, ranks what it finds against your goals, and hands you a reading queue plus publish-ready briefs.

What It Is

Research Scout is an external intelligence agent with a scoring engine behind it. It scans six sources across nine subject lanes, filters every article through a four-part test, and ranks results so the queue arrives ordered rather than dumped. Lane weights fade when a subject goes quiet and strengthen when your own transcripts mention it, so the ranking sharpens the longer you run it.

When to Reach for It

Set it on a daily schedule so a ranked reading queue is waiting each morning without you searching. Reach for it when you need something to publish and want an outside angle you can add a real position to, rather than a news recap. Run the internal mode after a run of recorded calls.

What You Get Back

A ranked reading queue of up to twenty-five articles, each with source, lane, score, and two sentences on why it connects to your work.
Three content briefs with title, format, lane, goal, angle, and the source link.
A confidence label on every item, with peer-reviewed papers rated high automatically.
Content threads mined from your own transcripts, sorted the way Pepper sorts them.
Flagged potential event speakers from those transcripts.

What You Will Need to Change

The full ranking engine needs the scout.py file from the skill folder plus a few Python packages, an API key, and a local database. Without that file a reduced version runs in-session, so copy scout.py across too if you want real scoring. Replace the subject lanes and goals with your own. One source, Reddit, currently returns nothing because it blocks the endpoint this uses.

Make it yours. Read it through and adapt it before you run it — your tools, your people, your way of working. It is built to be changed.

--- name: research-scout version: 2.1.0 description: > Use when anyone says "run Scout," "run Research Scout," "run Reading Scout," "content scan," "what's trending," "find me content," "content ideas," "what should I write about," or "content briefing." Also fires on "mine my transcripts," "what did I say recently," or "run Pepper." Two intelligence sources — Scout (external web + academic) and Pepper (internal transcripts). Python engine (scout.py) runs when available; Claude runs the search manually in-session when not. --- # Research Scout — Integrated Content Intelligence v2.0.0 ## Banned Words Gate — Hardwired Before drafting or presenting any output, read the live banned list fresh. Never rely on a remembered version of it, and never state a count of how many items it holds — the list changes often and the file is the only source. - `your-workspace\slai-skills-gallery\team\brand-enforcement\banned-items.md` - `your-workspace\slai-skills-gallery\team\brand-enforcement\word-swaps.md` Check every draft against the current list before it is shown. If the files cannot be read, stop and say so — never proceed on memory. Scout finds it outside. Pepper finds it inside. Together they surface what Anne should read, write about, and act on — scored against her five business goals with a ranking engine that actually ranks. ## What Changed in v2.0.0 v1.0.0 merged the Manus-built Reading Scout, Pepper, and scout-content-intel into a single skill but lost the original Reading Scout's ranking intelligence. v2.0.0 grafts that ranking engine back on top of the SLAI business-goal mapping: - TF-IDF style scoring with title boost and AI cross-signal - Lane-weight decay that fades stale lanes and reinforces active ones from transcripts - OpenAlex peer-reviewed academic source restored - Fundraising Strategy and Advancement lane added for Empowered Fundraiser - Five-goal framework — Cohorts, Community, Conference, Products, Authority - Three-block output — Reading Queue (Anne reads), Content Briefs (Anne publishes), Pepper (atoms and speakers) ## Context Required Read these before running: - `slai-skills-gallery/CLAUDE.md` — repo structure, brand standards - `slai-skills-gallery/team/brand-enforcement/SKILL.md` — voice and language rules ## Two Intelligence Sources ### Scout — External Intelligence Scans six sources across nine content lanes. Applies a four-check content criteria filter. Produces a reading queue (top 25 ranked articles) and content briefs (top 3 framed as publish-ready briefs). Sources scanned: - Hacker News — Algolia API, 7 targeted queries - Reddit — 8 subs (MachineLearning, artificial, AIethics, nonprofit, womenwhocode, learnmachinelearning, AIpromptEngineering, philanthropy) - Google News — 10 queries mapped to content lanes - arXiv — 5 academic queries via the arxiv library - RSS — 8 curated feeds (MIT Tech Review, Wired AI, HBR, Stanford Social Innovation, VentureBeat AI, The Markup, Nonprofit Tech for Good, AI Now Institute) - OpenAlex — 6 queries over peer-reviewed papers 2024–2026, sorted by citation count ### Pepper — Internal Intelligence Mines `~/meetings/*.txt` and `~/meetings/*.md` from the last 7 days for content atoms. Extracts Quotes, Frameworks, Hot Takes, Data Points, and Stories. Flags CREATE-ready speakers. Marks WOC visibility opportunities. Produces signed content briefs. Drop `.txt` or `.md` transcript files into `~/meetings/` before running Pepper. Mark private meetings with "private" or "confidential" in the first 200 characters — Pepper will skip them. Pepper also reinforces lane weights. If a transcript matches two or more keywords from a content lane, that lane's weight increases, sharpening scoring on subsequent Scout runs. ## How to Invoke ### When scout.py is available (preferred) ```bash cd /path/to/skills/research-scout python3 scout.py run # Full pipeline (Scout + Pepper) python3 scout.py scout # External scan only python3 scout.py pepper # Transcript mine only (last 7 days) python3 scout.py pepper 14 # Transcript mine — last 14 days python3 scout.py queue # View reading queue (top 25 unread) python3 scout.py queue --all # View all including read python3 scout.py atoms # View extracted content atoms python3 scout.py clear # Clear queue, atoms, and lane weights ``` ### When running in-session (Claude fallback) When the Python engine is not available, Claude performs the external scan manually: 1. Web search across all 8 active content lanes for material from the last 24–48 hours 2. Score each result against the 4-check filter 3. Apply confidence scoring 4. Generate 2–3 content briefs in Scout output format 5. Sign as "— Scout" For Pepper in-session, paste or attach transcript text and Claude will extract content atoms and produce briefs in Pepper output format. ## 9 Content Lanes | Lane | Status | Goals | |---|---|---| | AI Adoption Consulting | Active | Cohorts, Authority | | Women in AI | Active — WOC weighted higher | Conference, Community, Authority | | AI Education | Active | Cohorts, Authority | | AI Governance | Active | Cohorts, Authority | | Human-Centered AI | HOLD — awaiting subtopics | Authority | | Nonprofits and AI | Active | Cohorts, Authority | | Fundraising Strategy and Advancement | Active | Authority, Cohorts | | CREATE Conference Topics | Active — WOC weighted higher | Conference | | AI Tools and Workflows | Active | Cohorts, Products, Authority | ## 5 Business Goals | Goal | What It Serves | |---|---| | Cohorts | ACA, CAIO, future cohort programs | | Community | She Leads AI Society membership and engagement | | Conference | CREATE Conference — speakers, attendees, sponsors | | Products | Skills packs, individual skills, Moxxee | | Authority | Anne as extreme authority across AI, fundraising, AI + fundraising, founding, leading, community building, instruction, speaking, podcasting, essayist | Every content brief must tie to one of these five. No goal, no brief. ## 4-Check Content Criteria Filter Every article must pass all four before it enters a brief: | Check | Rule | |---|---| | 3E Check | Does it Entertain, Encourage, or Educate? Must do at least one. Label which. | | POV Check | Does it let Anne demonstrate a point of view, not just recap news? Pure news recaps get penalized 30 percent. | | Skills Check | Does it connect to Anne's actual skills — fundraising, AI adoption consulting, community building, systems thinking, translating AI for non-technical leaders, women's leadership? | | Lane/Goal Link | Does it map to an active lane with at least one of the 5 goals? No mapping, no brief. | ## Scoring Engine For each article, the engine computes a score against every active lane and keeps the highest: ``` raw = (keyword_hits / total_keywords) * log(1 + hits) * lane_weight + (broad_signal_hits / total_broad) * 0.08 * lane_weight + title_keyword_hits * 0.15 + 0.05 if article mentions AI core terms AND broad signals + 0.5 if lane has WOC visibility AND text contains WOC signals ``` Then `raw *= 0.7` if the article fails the POV check (news recap penalty). Lane weights start at 1.0. They decay at 2 percent per day after 3 days of no reinforcement, floored at 0.3. Pepper reinforces lane weights up to +0.3 per run when transcripts mention a lane's keywords two or more times, capped at weight 2.0. Peer-reviewed sources (arXiv, OpenAlex) default to confidence=HIGH regardless of text heuristics. ## Confidence Scoring | Level | Signals | |---|---| | HIGH | Source is arXiv or OpenAlex, or text contains study, research, data, report, survey, peer-reviewed, findings, evidence, statistics | | MEDIUM | Default | | LOW | Rumor, allegedly, unconfirmed, "sources say," might be, could be | ## Content Atom Categories (Pepper) | Atom Type | What It Is | |---|---| | Quotes | Verbatim or near-verbatim shareable lines from Anne or guests | | Frameworks | Named models, step processes, mental models mentioned | | Hot Takes | Strong opinions or contrarian positions stated | | Data Points | Specific numbers, stats, or metrics cited | | Stories | Anecdotes, case studies, examples with narrative structure | Every atom maps to one of the 5 goals via keyword signals. ## CREATE-Ready Speaker Flagging (Pepper) When a guest drops strong frameworks, hot takes, or demonstrates deep expertise in a transcript, Pepper flags them as a potential CREATE speaker. If the speaker is a woman of color (or the transcript uses she/her signals with WOC markers), the flag reads **WOC — CREATE visibility opportunity**. This feeds the speaker evaluator with pre-vetted candidates who have already demonstrated they can deliver — before any application form. ## WOC Visibility Lens In the Women in AI and CREATE Conference Topics lanes, WOC thought leaders are weighted higher (+0.5 score bonus when WOC signals appear in the text). Scout's job is to find the voices worth amplifying before they have the megaphone. CREATE's visibility principle starts here. ## Output Format Every full run produces three blocks. ### 1. Reading Queue (what Anne reads) ``` READING QUEUE (YYYY-MM-DD) — 25 articles [01] Article title truncated to 90 chars Source | Lane | Score: 1.234 | HIGH Two-sentence connection explaining why this matters to Anne's work. https://source-url [02] ... — Scout ``` ### 2. Content Briefs (what Anne publishes) ``` CONTENT BRIEFS (YYYY-MM-DD) BRIEF 1 Title: Article title Format: LinkedIn post or newsletter section Lane: Women in AI Goal: Conference | 3E: Educate | Confidence: HIGH Angle: Scout brief connecting article to Anne's work Keywords: women in AI, gender gap AI, female founders AI Source: https://source-url BRIEF 2 ... BRIEF 3 — Scout ``` ### 3. Pepper — Content Mine ``` PEPPER — Content Mine (YYYY-MM-DD) TRANSCRIPTS SCANNED: 3 ATOMS FOUND Quotes - "extracted quote" — Speaker Name [meeting-name] Frameworks - "extracted framework line" [meeting-name] Hot Takes / Data Points / Stories ... CONTENT BRIEFS BRIEF 1 from atoms ... CREATE-READY SPEAKERS SPOTTED - Speaker Name: signal snippet | WOC — CREATE visibility opportunity — Pepper ``` ## Gotchas No generic AI news. "OpenAI released a new model" is not a brief unless it ties to a SLAI goal. No defaulting to the loudest voices. WOC thought leaders often have smaller followings but sharper perspectives. Surface them first in Women in AI and CREATE lanes. No repeating the same format every time. Vary across LinkedIn post, blog, newsletter section, Social Saturday topic, podcast talking point. No mining private transcripts. Pepper checks for "private" or "confidential" in the first 200 characters and skips those files. No full drafts. Research Scout produces briefs only. The brief goes to a content skill or to Anne for direction. No sales-pipeline framing. Name the group by what it is, not "warm leads" or "nurture sequence." No banned words in LLM output. The system prompt instructs the model to avoid "actually," "happened," "worth doing," "worth keeping," "prospect" (use "prospective client"), "build" as a filler verb, and walk/walking metaphors. ## Constraints - Every content brief must tie to one of the 5 SLAI goals - Never surface content that contradicts SLAI's positions on AI governance, women's leadership, or human-centered AI - Never generate full drafts — briefs only - Scout signs off as "— Scout" and Pepper signs off as "— Pepper" - Peer-reviewed sources get HIGH confidence automatically ## Python Engine The `scout.py` file in this folder is the Python backend. It handles API calls, TF-IDF scoring, lane-weight decay, the SQLite database, atom extraction, connection generation, and brief generation. **Dependencies** — `feedparser`, `arxiv`, `anthropic`, `requests` (pip install) **Environment** — `ANTHROPIC_API_KEY` required for LLM connections (uses Claude Haiku 4.5); rule-based fallback runs without it **Database** — `~/.reading-scout/reading_scout.db` **Transcript source** — `~/meetings/*.txt` and `~/meetings/*.md` **Tables** — `reading_queue`, `content_atoms`, `pipeline_runs`, `lane_weights` ## Check-Ins After first run: - Are the 9 content lanes right? Any to add or retire? - Are the briefs actionable or too vague? - Does the reading queue feel like Manus-quality recommendations? After 2–3 runs: - Is Scout surfacing genuinely new angles or repeating known material? - Is the morning rhythm working or should we split into morning and afternoon? - Any lanes getting decayed too aggressively or not enough? ## Schedule Scout runs on GitHub Actions, not Rube (Rube is retired). | Job | Workflow | Schedule | Status | |---|---|---|---| | Scout Daily Feed Scan | `.github/workflows/scout-daily-feed.yml` → `agents/scout/run.py` | Daily 4:00 AM PT | Active | The scheduled job runs the **external scan only** (Hacker News, Reddit, Google News, arXiv, RSS, OpenAlex) and DMs Anne the reading queue plus content briefs. **Pepper is local / on-demand** — run `python scout.py pepper` on Anne's machine when transcripts are in `~/meetings/`; it has no data source on a cloud runner. ## Known Limitations and Roadmap **Reddit currently returns nothing.** The Reddit scanner hits the unauthenticated `www.reddit.com/.json` endpoint, which Reddit now blocks from datacenter IPs (403) — both locally and on GitHub runners. The other five sources (Hacker News, Google News, arXiv, RSS, OpenAlex) fill the queue fine, so this is a degradation, not an outage. **Future version — social listening expansion.** Parked for a later build: - Reddit fixed properly via OAuth app-only token (register a free Reddit app, add `REDDIT_CLIENT_ID` + `REDDIT_CLIENT_SECRET` secrets) - Bluesky keyword post search (app password) — where much of the women-in-AI / WOC-in-tech conversation now lives - Mastodon hashtag monitoring (#WomenInAI, #AIethics — no auth required) - YouTube keyword search (Google API key) — feeds the CREATE speaker pipeline - Not pursued: X/Twitter (paid API), LinkedIn / Threads (no public search API) ## Changelog | Date | Version | What Changed | |---|---|---| | 07.09.26 | 2.1.0 | Scheduled Scout on GitHub Actions (`scout-daily-feed.yml` → `agents/scout/run.py`, daily 4am PT), replacing the dead Rube recipes. Wrapper drives the existing scout.py external scan and DMs the reading queue + content briefs to Anne on Slack, signed "— Scout." Retired and deleted the archived scout-content-intel ancestor. Documented the Reddit datacenter-IP 403 and a social-listening roadmap (Reddit OAuth, Bluesky, Mastodon, YouTube) for a future version. | | 04.20.26 | 2.0.0 | Grafted Reading Scout's TF-IDF ranking engine onto v1.0.0. Added OpenAlex source. Added Fundraising Strategy and Advancement lane. Replaced 4C goals with 5-goal framework (Cohorts/Community/Conference/Products/Authority). Added lane-weight decay + transcript reinforcement. Added dedicated reading-queue output alongside content-briefs output. Pepper unchanged. | | 04.13.26 | 1.0.0 | Merged scout-content-intel v1.1.0 and Reading Scout + Pepper v2.0 into Research Scout. Python engine (scout.py) is canonical backend; Claude in-session search is the fallback. Added confidence scoring, content atom extraction, CREATE speaker flagging, WOC visibility lens, Pepper transcript mining. Retired scout-content-intel. |