Bilingual

"Software Engineering" Is Not Engineering
15 days ago
- Computer Science is not science and Software Engineering is not engineering because software design lacks objective metrics and ties to physical reality.
- Software is more akin to mathematics, allowing creation of virtual worlds with internal consistency, unlike engineering bound by physical laws.
- Debates in software often revolve around idioms, paradigms, and conventions, which are difficult to compare objectively due to reliance on human psychology and productivity factors.
- Programmer productivity and maintainability are key concerns, but empirical studies are scarce, and personal comfort with tools often dictates productivity.
- Existing computer science literature focuses on techniques and idioms, such as Boolean algebra, but fails to provide scientific methods for evaluating 'better' approaches.
- The flexibility of software enables infinite possibilities, but this also makes objective evaluation nearly impossible, leading to persistent, unresolved debates in the field.
Why you should not have Managers in your class name
15 days ago
- Naming classes and methods correctly is crucial as it influences code structure, future thinking, refactoring, and extension, with poor naming leading to technical debt.
- Using '-er' suffixes like 'Manager' often indicates poor design; for example, 'ConnectionManager' became complex with multiple responsibilities, whereas 'Connection' would have been clearer and simpler.
- Choosing precise names helps maintain clean code, simplifies extension and refactoring, and reduces complexity and bugs, so spending extra time on naming is recommended.
DSLs Enable Reliable Use of LLMs
7 days ago
- Domain-Specific Languages (DSLs) enhance LLM reliability by providing clear boundaries and constraints, reducing variability in code generation.
- DSLs act as a source of truth for software systems, enabling LLMs to serve as natural language interfaces and iterative partners in design.
- Abstractions and semantic models, like Tickloom for distributed systems, ground LLM prompts in existing vocabulary, improving accuracy and reducing hallucinations.
- LLMs operate in two phases: as brainstorming partners during DSL design and as reliable generators once the DSL is established.
- DSLs support deterministic validation (e.g., parsers or compilers), allowing autonomous agents to self-correct errors without human intervention.
- Examples include using LLMs with PlantUML for presentations and internal DSLs for testing distributed system scenarios, showcasing their versatility.
- Well-factored DSLs produce maintainable, readable artifacts that capture intent, reducing reliance on prompts as the primary source of truth.
Claude Is Not a Compiler
6 hours ago
- The author argues that LLMs like Claude exceed compilers by working vertically across all software layers, from strategy to machine code.
- Traditional software layers (executive to compiler) involve decision-making, but LLMs enable cross-layer integration without organizational overhead.
- A concrete example: building a distributed DNS server for exe.dev using iterative agent-driven design, differential analysis, and scar-tissue guidance over a week.
- The process contrasted 'vibe-coding' (delegating decisions) with 'vibe-engineering' (using LLMs to augment decision-making across layers).
- Software layers that provide insight remain valuable, but those offering mere convenience are becoming obsolete.
- The emerging paradigm shifts engineers' focus upward without abandoning deeper layer understanding, making vibe-engineering synonymous with engineering.

Hasty Briefsbeta

#software design

"Software Engineering" Is Not Engineering

Why you should not have Managers in your class name

DSLs Enable Reliable Use of LLMs

Claude Is Not a Compiler