discovery

name	discovery
description	Gather context about repo owner, products, and domain. Build domain expertise by reading top voices. Use proactively before creating content.
user-invocable	false

Discovery Skill

Find voices, read content, build expertise

Owner & Product Discovery

gh api users/{owner}  # Returns: name, bio, blog, twitter_username, company, location

Additional: Check ME.md, GitHub profile README, pinned repos, blog/website.

Products: gh api users/{owner}/repos?sort=updated → read READMEs, check live demos.

Staleness: Owner profile >30 days = refresh. Products = weekly.

---

Domain Trends Research

Web search for current data:

• "X Twitter growth strategies {current_year}"
• "AI developer Twitter accounts successful"
• "{niche} best practices {current_year}"

Refresh each session (trends change constantly).

---

Build Domain Expertise

1. Top Voices List (~20 voices)

Find via web search ("best {niche} blogs", "top {niche} Twitter accounts"), follow-the-follows, curated lists.

Store in agent/memory/research/top-voices.md:

## @handle / Name
- Platform: X / Blog / Newsletter
- Focus: [niche/angle]
- Why follow: [value]

Refresh monthly.

2. Reading Routine

Each session, pick 2-3 voices. Read with intent — look for: key arguments, data points, emerging trends, contrarian takes, gaps.

Cadence: Top 5 voices every session (skim). Voices 6-20 weekly rotation.

3. Capture Engagement Opportunities While Reading

X outbound replies fail 100% of the time via X API (403) — confirmed Week 9 audit: 62/62 failed. Do NOT spend turns looking for X posts to reply to outbound.

For X: Reply-to-own only (100% success rate). Look for tweet IDs from recent workflow runs:

gh run list --workflow=process-outputs.yml --limit 1 --json databaseId,createdAt
gh run view <run_id> --log 2>/dev/null | grep 'INFO Response:' | head -5

For Bluesky (especially during X outages): Outbound replies ARE allowed. While reading top voices and doing research, note any Bluesky posts on pillar topics that are <6h old and worth replying to. Capture the AT URI for the reply file. See "Reply Targets: Platform-Specific Rules" below.

4. Turn Reading Into Content

Reading output	Content use
Key takeaway	Authority post (your angle)
Disagreement	Contrarian take
Data point	Credibility boost
Trend spotted	First-mover post
Content gap	Fill it — own the topic

Rules: Never plagiarize. Credit sources. Aim for 1 reading-inspired post per 3-5 articles. Goal = informed originality, not summary.

5. Graduate Research Into Skills

Research in agent/memory/research/ helps THIS session. Research in .claude/skills/ helps ALL future sessions.

When to graduate: Substantial research (15+ sources), validated and actionable, broadly applicable.

Follow "Skill Development (High Bar)" from CLAUDE.md. NOT every session — reserve for validated, substantial findings.

---

Owner's Open Source Scan

Periodically scan the owner's public repos for promotable content. This feeds the publishing skill's OS promotion allocation.

When: Once per session (during research phase), or when looking for content ideas.

How:

1. Read ME.md → find the owner's GitHub profile URL under "Open Source (Promotable)"
2. WebFetch the profile page → discover public repos, orgs, and pinned projects
3. WebFetch each discovered org page → list their repos too
4. For repos that look promotable, fetch their README to find live output links and descriptions

What makes something promotable right now:

A. Live outcomes (highest priority) — Real content/services produced by agents. These are proof, not promises.

• Read ME.md for known live outcome URLs, then WebFetch each to check for recent articles, digests, posts
• A specific recent article is 10x more promotable than a generic "we have a blog" link
• Frame as: "This was written by an AI agent today" + link to the actual piece
• Also check repo READMEs for new live output URLs — new pipelines may launch anytime

B. Repos with live outputs — Code that powers running services. Link both the code AND the output.

C. Star milestones — Crossing 5, 10, 25, 50, 100 stars.

D. Trending topic overlap — Repo solves something people are discussing this week. Strongest combo: trend + live proof.

E. New repos — Launch announcements get one-time boost.

What to capture: For each candidate: what it does (1 line), proof it works (live output links + recent specific content), hook angle (why someone cares today). Store in agent/memory/research/os-promo-candidates.md.

Cross-reference with trends: Trending topic aligns with an owner repo or outcome? Post about the trend, link the repo/outcome as "we built this, and here's it running."

Reference: GitHub accounts and known live outcomes in ME.md under "Open Source (Promotable)".

---

Reply Targets: Platform-Specific Rules

X: Only Reply-to-Own Works

DO NOT search for outbound X reply targets. All outbound replies fail at X API (403) — confirmed 62/62 failures in Week 9 audit. No exceptions for brand accounts or individuals.

Reply-to-own is the only working X reply strategy. To get targets:

1. Run: gh run list --workflow=process-outputs.yml --limit 1 --json databaseId,createdAt → gh run view <id> --log | grep 'INFO Response:'
2. Only proceed if run completed <25 minutes ago (to hit the 150x <30min window)

Only create X reply files when:

• You have the numeric tweet ID of YOUR OWN recent tweet
• The topic has enough depth to add value in a reply
• Queue is < 15 on all platforms
• The original tweet was posted within 30 min (for 150x multiplier) — check run completion time first

Bluesky: Outbound Replies Work

Bluesky DOES allow outbound replies to any post via AT Protocol — no X-style permission restriction. During X outages, Bluesky outbound replies are a valid engagement tactic.

Finding Bluesky reply targets:

• Web search: "bsky.app {topic} {current_year}" or browse topic feeds
• Target mid-tier accounts (500-10K followers), posts < 6h old (chronological feed)
• Prioritize posts with 2-10 replies (active conversations)

Bluesky reply format: Use AT URI in REPLY_TO: header (see integrations skill for exact format).

Rules: Max 3-5 replies per session, apply same anti-AI writing rules, no self-promotion links, BS queue < 10 required.

See commenting skill ## Bluesky Engagement (During X Outages) for full detail.

---

Storage Structure

• agent/memory/research/top-voices.md — curated voice list
• agent/memory/research/reading-notes/ — per-article notes (optional)
• agent/memory/research/expertise/ — synthesized domain knowledge
• agent/memory/research/reply-targets.md — REMOVED. Outbound replies don't work. Reply-to-own IDs come from workflow logs, not this file.