Building AI Developer Tools & Perplexity's Memory System

Lenny Rachitsky - How to Build Your Own AI Developer Tools with Claude Code#

Source: https://www.lennysnewsletter.com/p/how-to-build-your-own-ai-developer Credibility: High (detailed technical walkthrough with live demonstration)

What happened: Lenny published a technical guide showing how to build custom AI developer tools using Claude Code—demonstrating that non-engineers can now create production-ready automation tools in hours, not weeks.

Key workflow patterns demonstrated:

Building custom automation tools:

• Tool creation timeline: 2-4 hours to build custom CLI tools that integrate with APIs, process data, and automate workflows
• No traditional coding required: Describe what you want the tool to do; Claude Code generates implementation, tests, and documentation
• Production-ready output: Tools handle error cases, include logging, and follow best practices without manual specification

Specific examples built:

• API integration tool: Automated data sync between project management tools (Linear, Notion, Asana)
• Custom linter: Code quality checker enforcing team-specific patterns beyond standard linters
• Deployment script: One-command deployment pipeline with pre-flight checks and rollback capability

The architecture pattern: Tools are standalone CLI utilities using Node.js or Python. They integrate with existing systems via APIs and run locally or in CI/CD. Claude Code generates the entire implementation including:

• Argument parsing and validation
• API authentication and error handling
• Logging and debugging output
• Tests and documentation

Why it's faster than traditional development:

• No boilerplate: Claude Code generates project structure, dependencies, and configuration
• Context awareness: Reads existing codebase to match conventions and patterns
• Error recovery: When initial implementation fails, Claude iterates to fix issues autonomously

Why it matters for PMs: This continues the "vibe coding" pattern from Feb 8 but focuses on internal tooling. PMs who can describe workflows can now build automation without engineering resources. The 2-4 hour timeline means prototyping is essentially free—test ideas before committing team bandwidth. For PMs managing technical teams, the question becomes: which internal tools should we build quickly with AI versus waiting for engineering capacity?

Critical questions:

• What's the maintenance burden when tools break and the PM who built it has moved on?
• How do you ensure security and reliability when non-engineers deploy automation?
• At what complexity does "build with Claude Code" become slower than traditional development?
• Does this create fragmentation—every PM building their own tools rather than standardizing?

Action you could take today: Identify one recurring workflow you currently do manually—data sync, report generation, deployment automation. Spend 30 minutes describing it to Claude Code and see if it can build a working prototype. This calibrates what's actually achievable versus what sounds good in theory.

Aravind Srinivas - Perplexity's Memory System Goes Agentic#

Source: https://x.com/AravSrinivas/status/2019913158589329497 Credibility: High (CEO announcement with technical detail)

What happened: Aravind Srinivas revealed Perplexity's memory system now operates agentic workflows—actively using remembered context to personalize research, not just passively recalling facts.

Key architectural insight:

What changed from passive to agentic memory:

• Passive memory (previous): System stores facts about you ("works in fintech", "prefers technical depth") and surfaces them when relevant
• Agentic memory (new): System actively applies context to shape research strategy—different search paths, source selection, and synthesis based on your history

How agentic memory works: When you ask a question, Perplexity's agent:

1. Reviews your memory profile (topics, depth preference, past research)
2. Plans research strategy based on your patterns (skip basic explanations, prioritize technical sources)
3. Executes research with personalized heuristics
4. Synthesizes findings matching your style

The "more agentic for each user" claim: The system doesn't just remember your preferences—it actively changes its behavior per user. Same question from two users triggers different research strategies based on their memory profiles.

Why this matters vs. simple personalization: Traditional personalization: "You like technical content, here's technical content." Agentic personalization: "You research AI safety papers, so I'll check arXiv first, cross-reference with top labs' blogs, and synthesize with emphasis on methodology not conclusions."

Why it matters for PMs: This demonstrates the next evolution of AI products: from stateless (every conversation starts fresh) to stateful (remember context) to agentic (actively use context to change behavior). For PMs building AI products with memory, the question shifts from "what should we remember?" to "how should memory change our agent's strategy?"

Critical questions:

• How does the system prevent overfitting to past behavior—locking users into narrow research patterns?
• What's the feedback mechanism when agentic memory makes wrong assumptions about what you want?
• Privacy model: how is memory stored, who can access it, and can users audit what's remembered?
• Does personalization create filter bubbles—narrowing information exposure over time?

Action you could take today: If you're building AI products with memory, map your current memory architecture: passive storage (save context) versus agentic application (use context to change behavior). Identify one workflow where actively applying memory would improve output quality, not just relevance.

Simon Willison - How StrongDM's AI Team Builds Serious Software Without Coding#

Source: https://simonwillison.net/2026/Feb/7/software-factory/ Credibility: High (detailed technical analysis from respected developer, includes architectural patterns)

Note: This item was already covered in detail in the Feb 8 update. See that update for complete analysis of StrongDM's "software factory" approach, technical architecture, and organizational patterns.

Deduplication note: Full coverage appeared in Feb 8 update analyzing vibe coding as organizational pattern. Not repeating here to avoid duplication.