Chapter 7: Extension Patterns

You are a research analyst synthesizing information from multiple sources.

<task>
Create a comparative analysis of the papers below, identifying agreements, contradictions, and gaps.
</task>

<output_structure>
<section name="executive_summary">
3-sentence overview of the state of research on this topic
</section>

<section name="consensus_findings">
For each finding where 3+ papers agree:
- Finding: [statement]
- Supporting papers: [list with citations]
- Strength of evidence: [Strong/Moderate/Weak]
</section>

<section name="contradictions">
For each major disagreement:
- Question: [what they disagree about]
- Position A: [view 1 with citations]
- Position B: [view 2 with citations]
- Possible explanation: [why they might differ]
</section>

<section name="research_gaps">
- Methodological: [what methods are missing]
- Contextual: [what populations/settings understudied]
- Theoretical: [what mechanisms unexplored]
</section>

<section name="quality_assessment">
For each paper:
- Citation: [Author Year]
- Sample size: [N]
- Design quality: [High/Medium/Low with brief reason]
- Generalizability: [High/Medium/Low with brief reason]
</section>
</output_structure>

<quality_requirements>
- All claims must cite specific papers
- No speculation beyond what papers state
- Flag when evidence is weak or conflicting
- Use academic tone but remain accessible
</quality_requirements>

<input_documents>
Paper 1:
"""
[PASTE PAPER 1 ABSTRACT OR FULL TEXT]
"""

Paper 2:
"""
[PASTE PAPER 2 ABSTRACT OR FULL TEXT]
"""

Paper 3:
"""
[PASTE PAPER 3 ABSTRACT OR FULL TEXT]
"""
</input_documents>

<constraints>
- Response length: 1,000-1,500 words
- Format: Markdown with proper headers
- Citations: Use (Author Year) format
- Do not introduce information not in the papers
</constraints>

Generate the analysis.

<persona>
You are a senior software engineer conducting a security-focused code review.
</persona>

<task>
Review the code below for security vulnerabilities, following OWASP Top 10 guidelines.
</task>

<review_structure>
<security_assessment>
For each vulnerability found:
- Vulnerability: [OWASP category]
- Severity: [Critical/High/Medium/Low]
- Location: [file:line or function name]
- Exploit scenario: [how an attacker could exploit this]
- Fix: [specific code change needed]
</security_assessment>

<secure_patterns_used>
List security best practices already implemented:
- Pattern: [what's done well]
- Location: [where it's implemented]
- Benefit: [what attack this prevents]
</secure_patterns_used>

<recommendations>
Priority order:
1. [Critical fixes - must do before deployment]
2. [Important improvements - should do this sprint]
3. [Best practices - nice to have]
</recommendations>
</review_structure>

<review_criteria>
- SQL injection vulnerabilities
- XSS attack vectors
- Authentication/authorization flaws
- Insecure data storage
- Insufficient logging
- Dependency vulnerabilities
</review_criteria>

<code_to_review>
[Python code here - see example below]

Example vulnerable code:
  from flask import Flask, request, render_template
  import sqlite3

  app = Flask(__name__)

  @app.route('/user/<user_id>')
  def get_user(user_id):
      conn = sqlite3.connect('users.db')
      cursor = conn.cursor()
      query = f"SELECT * FROM users WHERE id = " + user_id
      cursor.execute(query)
      user = cursor.fetchone()
      return render_template('user.html', user=user)
</code_to_review>

<constraints>
- Be specific: cite exact line numbers
- Provide working code fixes, not just descriptions
- Assume production environment with real user data
</constraints>

Perform the security review.

SECURITY ASSESSMENT

Vulnerability 1: SQL Injection (OWASP A03:2021)
- Severity: CRITICAL
- Location: Line 9 (query construction)
- Exploit scenario: Attacker sends /user/1%20OR%201=1 to dump entire user table
- Fix:
  ```python
  query = "SELECT * FROM users WHERE id = ?"
  cursor.execute(query, (user_id,))

<Callout type="warn">

**When to Use XML Structuring:** Prompts with 5+ distinct requirements, multi-document analysis, complex output formats (reports with many sections), or situations where small misunderstandings cause big problems.

</Callout>

---

## Pattern 3: Few-Shot Workflow Chains

**Use Case:** Teach complex multi-step processes with examples

### The Concept

Combine few-shot learning with workflow chaining. Show the AI an example of the complete process, then have it replicate it.

### Complete Example - Research Workflow (Abstract → Analysis → Implications)

```text
You are a research analyst following a structured analysis workflow.

I'll show you one complete example of the workflow, then ask you to apply it to a new paper.

<example>
<input_paper>
Title: "Remote Work and Innovation" (Smith 2024)
Abstract: We surveyed 500 software teams over 18 months, finding that remote teams produced 22% fewer patent applications but 31% more incremental improvements to existing products. Suggests trade-off between radical and incremental innovation.
</input_paper>

<step_1_extraction>
**Research Design:** Survey, longitudinal (18 months)
**Sample:** N=500 software teams
**Key Variables:** Remote vs. office, patent applications (radical innovation proxy), product improvements (incremental innovation proxy)
**Main Finding:** Remote work decreases radical innovation (-22%) but increases incremental innovation (+31%)
**Statistical Significance:** Not stated in abstract
</step_1_extraction>

<step_2_analysis>
**Implications for Theory:**
Challenges assumption that remote work is universally positive or negative. Suggests innovation TYPE matters, not just quantity.

**Limitations:**
- Software teams only (generalizability?)
- Patent applications as innovation proxy (misses non-patentable innovation)
- Correlation not causation (selection effects?)

**Practical Significance:**
Large effect sizes (22%, 31%) if causal. Suggests companies should match work location to innovation goals.
</step_2_analysis>

<step_3_recommendations>
**For Researchers:**
- RCT needed to establish causality
- Expand to other industries
- Mechanism unclear (collaboration barriers? Deep work time?)

**For Practitioners:**
- Use office time for brainstorming/ideation (radical innovation)
- Use remote time for execution/refinement (incremental innovation)
- Hybrid model: alternate based on project phase

**Follow-Up Questions:**
- What's the optimal remote/office ratio for different innovation goals?
- Does team size moderate the effect?
- How long do effects take to emerge?
</step_3_recommendations>
</example>

Now apply this exact workflow to the new paper below:

<new_paper>
Title: "AI Coding Assistants and Developer Productivity" (Johnson et al. 2024)
Abstract: Randomized controlled trial with 200 professional developers. Treatment group used GitHub Copilot for 3 months. Found 26% faster task completion but 14% more bugs in initial code. Bug rate normalized after code review. Suggests speed-quality trade-off with AI assistance.
</new_paper>

Follow the same structure:
1. Extract research design details
2. Analyze implications and limitations
3. Generate recommendations for researchers and practitioners

Output:

<workflow_template>
Step 1: [First action - extract/identify/categorize]
Step 2: [Second action - analyze/compare/evaluate]
Step 3: [Third action - synthesize/recommend/decide]
</workflow_template>

<example_complete_workflow>
<input>[Example input]</input>
<step_1_output>[What Step 1 produces]</step_1_output>
<step_2_output>[What Step 2 produces]</step_2_output>
<step_3_output>[What Step 3 produces]</step_3_output>
</example_complete_workflow>

Now apply to:
<new_input>[Your actual content]</new_input>

# Weekly Competitive Intel Workflow

**Time:** 30 minutes total
**Frequency:** Every Monday morning
**Tools:** Web browser + Claude

## Process

| Step | Action | Time | Output File |
|------|--------|------|-------------|
| 1 | Manual research on 3 competitors | 10 min | raw_intel_[date].md |
| 2 | Extract structured data (Prompt 6) | 5 min | competitor_analysis_[date].md |
| 3 | Compare week-over-week (Custom prompt) | 5 min | competitive_changes_[date].md |
| 4 | Generate exec summary (Structured output) | 3 min | exec_summary_[date].md |
| 5 | Create action items (Decision analyzer) | 7 min | action_items_[date].md |

## Automation Opportunities

- **Now:** Template all 5 prompts (done above)
- **T1.2 (Claude Projects):** Upload all files to Project, auto-compare
- **T2.x (API Integration):** Fully automate collection + analysis

## Success Metrics

- Week 1 baseline: 90 minutes manual analysis
- Week 4 target: 30 minutes with prompts (67% time savings)
- Quality: Action items lead to measurable responses

# Literature Review State Tracker
**Project:** AI Impact on Labor Markets
**Started:** 2025-01-15
**Target:** 25 papers reviewed, synthesis complete by 2025-01-20

## Progress

**Papers Reviewed:** 3 / 25
- [x] Chen & Rodriguez (2023) - Remote work productivity
- [x] Smith (2024) - Remote work innovation
- [x] Johnson et al. (2024) - AI coding assistants
- [ ] [Remaining 22 papers]

## Emerging Themes

**Theme 1: Automation vs. Augmentation**
- Papers: Johnson 2024
- Key insight: AI tools augment rather than replace when used with human review
- Gap: No studies on full automation scenarios

**Theme 2: [To be developed]**

## Key Findings Tracker

| Paper | Main Finding | Effect Size | Quality |
|-------|--------------|-------------|---------|
| Chen 2023 | Remote → +13% productivity | Medium | High (RCT) |
| Smith 2024 | Remote → -22% radical innovation | Large | Medium (Survey) |
| Johnson 2024 | AI → +26% speed, +14% bugs | Large | High (RCT) |

## Research Gaps Identified

- [ ] Long-term effects (>6 months)
- [ ] Non-knowledge-worker sectors
- [ ] Interaction effects (AI + remote work)

## Next Steps

**Tomorrow's session:**
- Review papers 4-8
- Update emerging themes
- Check if "Automation vs. Augmentation" theme holds

You are a research analyst continuing a multi-day literature review.

<project_context>
You've been helping analyze papers on AI's impact on labor markets. We're on day 2 of a 5-day review process.
</project_context>

<current_state>
"""
[PASTE ENTIRE STATE TRACKER DOCUMENT FROM ABOVE]
"""
</current_state>

<todays_task>
Review the next 5 papers (papers 4-8) and:
1. Extract key findings using the same format as the tracker
2. Identify if they support existing themes or reveal new ones
3. Update the research gaps list
4. Flag any contradictions with previous papers
</todays_task>

<papers_to_review>
Paper 4: [abstract]
Paper 5: [abstract]
Paper 6: [abstract]
Paper 7: [abstract]
Paper 8: [abstract]
</papers_to_review>

<output_format>
**UPDATED KEY FINDINGS TRACKER:**
[Extended table with new papers added]

**THEME UPDATES:**
Theme 1 (Automation vs. Augmentation):
- [Update based on new papers]
- New supporting papers: [list]
- Contradicting papers: [list if any]

Theme 2 (NEW if emerged):
- [Description]
- Supporting papers: [list]

**NEW RESEARCH GAPS:**
[Add any new gaps identified]

**CONTRADICTIONS FOUND:**
[Any papers that disagree with previous findings]

**PROGRESS SUMMARY:**
- Papers reviewed: 8 / 25
- Themes identified: [count]
- On track for [date] completion: Yes/No
</output_format>

Generate the update.

# [Project Name] State Tracker

**Project Goal:** [What you're trying to achieve]
**Timeline:** [Start date] to [End date]
**Current Phase:** [Which stage you're in]

## Progress Metrics

**Completion:** [X] / [Total] [units]
- [x] Completed item 1
- [x] Completed item 2
- [ ] Remaining item 1
- [ ] Remaining item 2

## Accumulated Knowledge

**Key Insights So Far:**
1. [Insight 1 with supporting evidence]
2. [Insight 2 with supporting evidence]

**Decisions Made:**
- [Date]: [Decision] - [Reasoning]
- [Date]: [Decision] - [Reasoning]

**Open Questions:**
- [ ] [Question 1] - [Status: researching/blocked/answered]
- [ ] [Question 2] - [Status: researching/blocked/answered]

## Data Collected

[Table or structured list of data accumulated]

## Next Session Plan

**When:** [Date/Time]
**Goal:** [What to accomplish]
**Inputs needed:** [What you'll need ready]
**Expected outputs:** [What you'll produce]

## Resumption Prompt Template