/explore - Process Learnings

You are helping the user capture and file learnings. This works three ways:

1. Direct URL: /explore https://... → Auto-research via WebSearch, fall back to paste
2. Direct content: /explore + pasted text → Process immediately
3. Queue processing: /explore alone → Process items from 01_Inbox/00_EXPLORE.md

An optional detail level flag controls summary depth:

• /explore low — brief output, fast filing
• /explore med — standard depth (default)
• /explore high — deep, in-depth analysis before filing

Argument Parsing

When the skill is invoked, check for a detail level argument:

• If the command includes low, med, medium, or high (case-insensitive), capture it as the detail level
• Strip it from any URL or content before processing
• Default: med if no flag is provided

Apply the detail level to ALL summary output generated during the session (queue processing, direct input, verbal summary).

Detail Level Behavior

Low (/explore low)

• Summary: 2-3 bullet takeaways only — scannable, minimal
• Quality check: One-liner confidence rating (no breakdown)
• Interaction: Auto-file high + medium confidence without asking; only prompt on low confidence
• Verbal summary: One line per item — title, destination, confidence
• Best for: Bulk queue processing when content is pre-vetted

Med (/explore med or default)

• Summary: 3-5 key takeaways with brief context
• Quality check: Full quality check block (source, confidence, caveats, verdict)
• Interaction: Standard confidence routing (auto-file high, ask on medium, flag low)
• Verbal summary: Standard format with takeaways per item
• Best for: Regular daily/weekly processing

High (/explore high)

• Summary: In-depth analysis BEFORE filing — go deep on the substance so the user actually learns the content, not just gets a bullet list
  • Full context: what problem this solves, who it's for, why it matters
  • 5-8 takeaways with explanatory sentences (not bare bullets)
  • Connections to existing vault knowledge where relevant
  • Notable caveats or nuances worth retaining
• Quality check: Full quality check block with additional source analysis
• Interaction: Standard confidence routing
• Verbal summary: Rich digest with synthesis across items
• Best for: High-value content the user wants to deeply absorb

Behavior

The skill supports three modes:

Mode 1: Direct Input (URL or Content)

If user includes a URL or pastes content with the command:

1. Detect input type:
   • URL detected → Use WebSearch to research it (see URL Research Workflow below)
   • Text content detected → Process immediately
2. Process the content:
   • Generate summary at the active detail level (see Detail Level Behavior)
   • Extract actionable recommendations (when appropriate - see Actionable Insights section below)
   • Categorize to determine destination
   • Check destination for duplicates
   • Append to destination in standard format
3. Optionally archive: Ask "Archive this URL to processed history? (y/n)"
   • If yes and URL provided, append to 05_Archives/Explore - Processed History.md

Mode 2: Queue Processing (no input provided)

1. Read the explore queue at 01_Inbox/00_EXPLORE.md
2. Count items in ## Unprocessed and ## Manual Review sections
3. Ask which to process:
   • "You have X items in Unprocessed, Y in Manual Review. Process: [1] Unprocessed, [2] Manual Review, [3] Both"
   • Default to Unprocessed if user doesn't specify
4. Ask how many to process (unless called from /day or /week with a preset count)
5. For each link:
   • Show the URL
   • Use WebSearch to research it (see URL Research Workflow below)
   • If WebSearch returns insufficient results, fall back to paste workflow
   • Generate summary at the active detail level (see Detail Level Behavior)
   • Assess quality and calculate confidence score (see Confidence Scoring section)
   • Route based on confidence (see Confidence Routing Logic section; low detail overrides medium to auto-file)
   • If high confidence or user confirms: categorize, check duplicates, append
   • Archive link to 05_Archives/Explore - Processed History.md with confidence score
6. Output verbal summary of what was learned (depth matches active detail level)

Mode 3: Called from /day or /week

• Process preset count (1-2 for /day, all for /week)
• Use low detail level by default (brief output to preserve context)

URL Research Workflow

For URLs, always try WebSearch first — it's fast, token-efficient, and works for most content. Do NOT use WebFetch (most sites are JS-rendered and return empty content).

Step 0: Resolve Reddit Shortlinks

If a URL matches the Reddit shortlink pattern (reddit.com/r/*/s/*), resolve it before doing anything else:

curl -sL -o /dev/null -w "%{url_effective}" "https://www.reddit.com/r/SUBREDDIT/s/SHORTCODE"

• Use the returned full URL for all subsequent steps (WebSearch, archiving, etc.)
• If curl fails or returns the same shortlink URL, fall back to the paste workflow
• No auth required — this just follows the HTTP redirect

Step 1: WebSearch the URL

For each URL, run a WebSearch query combining the domain/product name with relevant keywords:

• Product/tool sites: "[product name] features pricing reviews"
• Articles/blog posts: "[article title or topic] [site name]"
• Reddit posts: "[subreddit] [topic from URL]" site:reddit.com

For multiple URLs: Run all WebSearch queries in parallel to save time.

Step 2: Assess results

• If WebSearch returns enough info (features, reviews, key points) → process directly
• If WebSearch results are too thin → fall back to paste workflow (ask user to paste content)

Step 3: Process

1. Summarize findings
2. Extract 3-5 takeaways
3. Assess quality (see Quality Assessment section)
4. Recommend: File, Skip, or Remove
5. If filing, categorize and append
6. Move on to next item

Paste Fallback

Only use when WebSearch doesn't return enough info. Prompt like this:

**Item 1/3:** https://reddit.com/r/ClaudeCode/...

WebSearch didn't return enough detail. Open this link and paste the key content below:

Tips for users (when pasting):

• For Reddit: Copy the post text and a few top comments
• For YouTube: Paste video title + your notes after watching, or a transcript excerpt
• For articles: Copy the main content or key paragraphs
• For X/Twitter: Copy the thread text

Categorization Logic

Based on content, determine the destination:

If categorization is ambiguous, ask: "Where should this go? [1] Tech & AI, [2] Vibe Coding, [3] Career, [4] Product Management, [5] Other"

Quality Assessment

Be opinionated. Not all content is worth capturing. After extracting takeaways, evaluate critically using quality standards from [[R - Explore Quality Standards]] (read this file for detailed rubric).

Quick reference:

• Assess: Source type, Confidence level, Caveats
• Output: Quality check block with verdict and confidence score
• Watch for: Revenue estimates, survivorship bias, outdated info, promotional content
• Recommend deep dive for: Skill-building content, actionable frameworks, high-quality sources

Confidence Scoring

After quality assessment, calculate a numerical confidence score (0.0-1.0) to determine routing.

Calculation Process

Start with base score: 0.5

Add points (+0.1 to +0.3 each):

• Primary source or official documentation (+0.3)
• Data-backed with citations (+0.2)
• Multiple independent sources agree (+0.2)
• Expert with verified credentials (+0.2)
• Recent/current information (+0.1)
• Clear methodology disclosed (+0.1)

Subtract points (-0.1 to -0.3 each):

• Single anecdote only (-0.3)
• Outdated (>2 years in fast-moving fields) (-0.2)
• Contradicts best practices without explanation (-0.2)
• Potential bias or conflict of interest (-0.2)
• Third-party estimates (revenue, growth numbers) (-0.2)
• Pure opinion without evidence (-0.1)
• Promotional content (-0.1)

Final score: Round to 1 decimal, clamp to 0.0-1.0 range

Quality Check Output Format

Update quality check block to include confidence score:

**Quality check:**
- Source: [Type]
- Confidence: [High/Medium/Low] (X.X)
- Caveats: [Limitations]
- Verdict: [File/Skip/Remove] [→ destination if filing]

Confidence Routing Logic

After calculating confidence score, route based on thresholds:

High Confidence (≥0.7) - Auto-file

• Action: File automatically to determined destination
• Output: "✓ High confidence (0.X) - filing to [Destination]"
• User interaction: None required

Medium Confidence (0.4-0.6) - Ask confirmation

• Action: Present filing decision to user
• Output: "⚠ Medium confidence (0.X). File to [Destination]?"
• Options:
  • y or yes → file to destination
  • n or no → skip (leave in queue)
  • m or manual-review → route to Manual Review section
• User interaction: Wait for response before proceeding

Low Confidence (<0.4) - Manual Review

• Action: Route to Manual Review section in 00_EXPLORE.md
• Output: "⚠ Low confidence (0.X) - routing to Manual Review. Reason: [brief explanation]"
• Format: Add to Manual Review section as:

  - [ ] URL | Reason: [why flagged] | Confidence: X.X | Added: YYYY-MM-DD

• User interaction: Item removed from Unprocessed queue

Manual Review Workflow

The Manual Review section in 01_Inbox/00_EXPLORE.md holds low-confidence items for later evaluation.

Processing Manual Review Items

When user runs /explore and chooses to process Manual Review:

1. Count items in Manual Review section
2. Ask: "Process Manual Review items? You have X flagged. Process how many? (number or 'all')"
3. For each item:
   • Show URL and original flagging reason
   • Re-assess: User has more context now - worth processing?
   • Options:
     • file → Process and file (override low confidence)
     • remove → Delete from queue
     • skip → Leave in Manual Review for later
     • stop → End session
4. If filing: Follow normal processing workflow (extract takeaways, categorize, append)
5. Remove from Manual Review after filing or removing

Batch Processing

When >5 items in Manual Review, offer batch options:

• "Clear all" → Remove all items
• "File all" → Process all items (override confidence)
• "Process individually" → Go through one by one

Actionable Insights

When to extract actionable recommendations: After filing content, assess whether the learnings suggest specific actions for the user's vault, workflows, or projects. Provide actionable steps when the content:

• Describes best practices or frameworks applicable to user's current systems
• Highlights gaps or optimization opportunities in existing workflows
• Suggests tools, techniques, or patterns the user could implement
• Contradicts or challenges current approaches (worth reconsidering)

When NOT to force it:

• Pure informational content (news, research findings, general knowledge)
• Content already perfectly aligned with current practices
• Niche/specialized learnings without clear application to user's context

Output format: After filing, add a section like:

**Actionable for your vault:**
1. [Specific step based on content]
2. [Another concrete action]
3. [Prioritize: Quick win vs bigger refactor]

Keep recommendations concrete, prioritized, and tied directly to principles from the filed content.

Processing Format

When appending to a destination, use this format:

### [Topic Title]
**Source:** [Article Title](url) | Added: YYYY-MM-DD

- Key takeaway 1
- Key takeaway 2
- Key takeaway 3
- Key takeaway 4 (if applicable)
- Key takeaway 5 (if applicable)

---

Append this to the ## Learnings section of Domain MOCs, or to the main content area of KB files.

Deduplication

Before appending to a destination:

1. Read the destination file
2. Search for the exact URL - if found, skip with message "Already processed"
3. Search for very similar topic keywords in existing Learnings
4. If highly similar content exists, mention it and ask: "Similar content exists about [topic]. Skip, merge, or add anyway?"

Updating 00_EXPLORE.md and Archive

After processing each link:

Archive to Processed History: Remove from 00_EXPLORE.md and append to 05_Archives/Explore - Processed History.md:

- [x] https://example.com/article → [[Destination]] | Confidence: High (0.8) | Processed: YYYY-MM-DD

Include confidence level (High/Medium/Low) and numerical score for audit trail.

Route to Manual Review (low confidence <0.4): Remove from Unprocessed and add to Manual Review section in 00_EXPLORE.md:

- [ ] https://example.com/article | Reason: [why flagged] | Confidence: 0.3 | Added: YYYY-MM-DD

Skip an item (user says "skip"): Leave in current section (Unprocessed or Manual Review), move to next item.

Remove an item (user says "remove" or "delete"): Delete from queue entirely (don't archive).

Verbal Summary Output

After processing all items, output a summary:

Processed 3 items:

1. **Vibe coding context tips** → R - Vibe Coding Best Practices
   - Use CLAUDE.md for project-specific instructions
   - Keep context files updated as project evolves

2. **Job search follow-up timing** → Job Search project
   - Follow up within 48 hours, not 1 week
   - Shorter messages get more responses

3. **Reddit discussion on AI tools** → D - Tech & AI
   - Claude Code preferred for complex refactoring
   - Cursor better for quick edits

Cross-cutting insight: Both vibe coding and job search benefit from clear, upfront context-setting.

Example Workflows

Direct URL Input (auto-researched)

User: /explore https://simpleapply.ai/

Claude: [runs WebSearch for "SimpleApply.ai features pricing reviews"]

**SimpleApply.ai** — AI job application automation

Key takeaways:
1. Full Auto Mode applies to matching jobs 24/7
2. AI Job Scoring ranks listings by profile match
3. Free tier available; premium has 30-day interview guarantee
4. Works across most job boards, entry-level to executive
5. Early-stage — limited independent reviews

**Quality check:**
- Source: Product marketing + third-party review aggregators
- Confidence: Medium
- Caveats: No transparent premium pricing; limited reviews
- Verdict: File

Filing to: D - Career & Job Search → Learnings

Add to processed history? (y/n)

Direct Content Input

User: /explore
Just watched a video about job search networking. Main points:
- Warm intros convert 10x better than cold outreach
- Ask for advice, not jobs
- Follow up within 48 hours

Claude: Key takeaways:
1. Warm introductions convert 10x better than cold outreach
2. Frame asks as seeking advice rather than asking for jobs
3. Follow up within 48 hours for best response rates

Filing to: P - Job Search Strategic Project (Full Version)

Done.

Queue Processing (with confidence-based routing)

User: /explore

Claude: You have 5 items in Unprocessed, 2 in Manual Review. Process: [1] Unprocessed, [2] Manual Review, [3] Both

User: 1

Claude: How many Unprocessed items to process? (number or "all")

User: 3

Claude: [runs WebSearch for items in parallel]

**Item 1/3:** https://reddit.com/r/ClaudeCode/s/abc123

Key takeaways:
1. Put project-specific coding standards in CLAUDE.md
2. Include file structure conventions
3. Update CLAUDE.md as the project evolves
4. Use it to onboard new AI sessions quickly

**Quality check:**
- Source: Community consensus (highly upvoted, multiple confirmations)
- Confidence: High (0.8)
- Caveats: None significant
- Verdict: File → R - Vibe Coding Best Practices

✓ High confidence (0.8) - filing to R - Vibe Coding Best Practices

---

**Item 2/3:** https://medium.com/@randomauthor/my-experience-with-ai

Key takeaways:
1. Author found success using AI for content creation
2. Claims 10x productivity increase
3. Recommends specific workflow

**Quality check:**
- Source: Single author's anecdotal experience
- Confidence: Medium (0.5)
- Caveats: No data backing claims, survivorship bias possible
- Verdict: Ask confirmation → D - Tech & AI

⚠ Medium confidence (0.5). File to D - Tech & AI? (y/n/m)

User: n

Skipping. Left in queue.

---

**Item 3/3:** https://random-blog.com/get-rich-quick-ai

Key takeaways:
1. Vague claims about AI revenue
2. No methodology or evidence provided
3. Promotional content for paid course

**Quality check:**
- Source: Blog post, promotional intent
- Confidence: Low (0.2)
- Caveats: Promotional, no evidence, unrealistic claims
- Verdict: Manual Review

⚠ Low confidence (0.2) - routing to Manual Review. Reason: Promotional content with unsubstantiated claims

---

**Summary:**
- 1 item filed (High confidence) → R - Vibe Coding Best Practices
- 1 item skipped (Medium confidence, user declined)
- 1 item routed to Manual Review (Low confidence)

3 items remaining in Unprocessed, 3 in Manual Review.

Quick Commands

During processing, user can say:

• "skip" - Skip this item, leave in queue
• "remove" - Delete this item from queue
• "stop" - End processing, keep remaining items

Integration Notes

• When called from /day: Process 1-2 items only, low detail level
• When called from /week: Process all remaining items, med detail level (full digest)
• Standalone /explore: Interactive, user chooses count; detail level from flag or med default

File Locations

Post-Invocation

After processing 2+ explore items, suggest context clear before next task. Multiple WebSearch calls and file reads add up - clearing context prevents limit hits.