## Sources 1. [TBM 427: The Bottleneck Strike Again!](https://cutlefish.substack.com/p/tbm-427-the-bottleneck-strike-again) 2. [What I'd Do As a Junior Candidate in Mid-2026](https://joereis.substack.com/p/what-id-do-as-a-junior-candidate) 3. [What Actually Helps Me Stop Hitting Claude Code Usage Limits](https://aimaker.substack.com/p/claude-code-usage-limits-token-saving) --- ### **TBM 427: The Bottleneck Strike Again! - by John Cutler** **Main Arguments** * **Engineering was rarely the true bottleneck** in product development, and the current narrative that "the bottleneck has moved" because of AI is an oversimplification of a complex system [1, 2]. * Product development is not a linear system with one "clogged pipe"; rather, it is a multi-faceted process involving **sensing, deciding, learning, aligning, and adapting**, not just writing code [1, 3]. * The "bottleneck is moving" meme is frequently used by consulting firms like McKinsey and Deloitte to justify **massive workforce reductions** by promising inflated AI productivity gains [4]. * If engineering truly were the only bottleneck, AI's ability to remove it would mean competitors could easily "take you to the cleaners" because the speed of raw code generation would drop to **zero competitive value** [5, 6]. **Key Takeaways** * **Constraints are distributed and co-evolving**; what appears to be a bottleneck is often just a temporary pattern emerging from a complex web of context, incentives, and dependencies [7]. * In human systems, the real constraints are often **policy, mindset, or coordination** rather than a single physical choke point [8]. * True competitive advantage shifts to organizations that can execute **"high-quality knowledge turns"**—the iterative cycle of forming hypotheses and testing them against messy human reality—rather than just generating artifacts [6, 9]. * Because AI technology and its impacts are changing so rapidly, anyone claiming to have "figured it out" is likely engaging in **deceptive marketing or is delusional** [2, 10]. **Important Details** * John Cutler references systems thinkers like **Donella Meadows and Alicia Juarrero** to illustrate that the "single constraint" model is often a simplifying lens rather than a literal truth [7]. * He cites **Eliyahu M. Goldratt’s Theory of Constraints**, noting that later work emphasized human systems' constraints over physical ones [8, 9]. * The current era of AI is compared to the introduction of the **printing press, electricity, or the internet**, where society is still grappling with fundamental shifts in knowledge and authority [11]. * Executive interest in AI productivity is often driven by a desire to expand **EBITDA and margins** by reframing labor costs like stock-based compensation as "too expensive" [4, 12]. --- ### **What Actually Helps Me Stop Hitting Claude Code Usage Limits - by Wyndo and Dheeraj Sharma** **Main Arguments** * Claude Code usage limits are often reached not just because of the number of messages sent, but because **every new message carries the weight of the entire conversation context** [13, 14]. * Significant "token waste" occurs when Claude is forced to carry **fragmented context**, discarded ideas, or unrelated files from earlier in a session [14, 15]. * The primary strategy for saving tokens is to **reduce unnecessary guessing** by the AI by helping it know exactly where to look and when a specific conversation is finished [15, 16]. **Key Takeaways** * **Clear the session (`/clear`) whenever a task changes** to stop carrying old, irrelevant context into new work [15, 17]. * **Compact the session (`/compact`) with specific instructions** when a session reaches 60-70% of the context window to preserve goals and decisions while dropping "slop" like verbose logs and old drafts [18-20]. * **Keep the `CLAUDE.md` file lean** and use it as a routing mechanism to point the agent to specific rule files only when they are needed for a particular task [21-23]. * **Match the model to the difficulty of the job**: use Haiku for mechanical cleanup, Sonnet for execution, and Opus for high-level planning and reasoning [24]. **Important Details** * Dheeraj recommends running `/compact` manually before it happens automatically so the user can **control what information is preserved**, such as success criteria and source URLs [18-20]. * Directly **pointing Claude at specific files or folders** prevents it from scanning the entire project and wasting tokens on irrelevant data [25, 26]. * **Unused MCP (Model Context Protocol) connectors** can add hidden costs; users should disable connectors like Canva, PayPal, or Notion if they aren't needed for the current task [27-29]. * Commands like **`/context` and `/usage`** are essential for identifying "leaks," such as automated workflows that consume more tokens than expected [30, 31]. * For users comfortable with the command line, **CLI plus skills** can be a lighter, more token-efficient alternative to always-on MCP connectors [29, 32]. --- ### **What I'd Do As a Junior Candidate in Mid-2026 - by Joe Reis** **Main Arguments** * The job market for junior software and data engineers in mid-2026 is **far more difficult than in 2022** due to rising interest rates and the rapid integration of AI into organizations [33, 34]. * Juniors must accept that **using AI is not up for debate**; they must leverage it to augment their tactical skills and perform heavy lifting like debugging, documentation, and prototyping [35, 36]. * AI has essentially **equalized the playing field**, as even senior professionals are scrambling to adapt, giving juniors an advantage because they have no "bad habits or dogma" to break [37, 38]. **Key Takeaways** * **Show your work through high-quality portfolios** that demonstrate problem-solving workflows, including loop engineering and evals, rather than just pushing AI-generated "slop" to GitHub [39]. * Focus on **foundational human traits** that are harder to teach than technical skills: Curiosity, Continuous Learning, Business Understanding, Communication, and Critical Thinking [40, 41]. * **Networking and mentorship** are critical for long-term career success; candidates should contribute to communities and build relationships without being transactional [38, 42]. * Broaden expertise beyond traditional data paths by learning **"unsexy" but essential topics** like data governance, modeling, security, and ontologies [37, 43]. **Important Details** * Reis notes that roughly **40% of people never read another book after college**, suggesting that simple persistence in learning is a major competitive advantage [37]. * Technical skills to augment with AI include **SQL, Python, databases, Git, and cloud fundamentals**, but users must still understand the underlying mechanics to review AI output critically [36]. * New essential technologies to explore include **MCP, RAG (Retrieval-Augmented Generation), and Vector Databases** [43]. * The source mentions that networking can pay off decades later, citing a personal example of linking two friends for a job based on **relationships built over 15-20 years** [42]. * Maxime Beauchemin is cited as an example of a senior leader pivoting toward **agentic workflows** with tools like Agor to automate complex tasks like legal reviews and bug bashing [44, 45].