Rosa Del Mar

Daily Brief

Issue 79 2026-03-20

Prompting As Engineering Asset And Context Management

Issue 79 Edition 2026-03-20 8 min read
General
Sources: 1 • Confidence: Medium • Updated: 2026-03-25 17:54

Key takeaways

  • Prompt engineering is iterative, and once a good prompt is found it should be made persistent (copyable or checked into Git) because additional verbosity can hit diminishing returns.
  • Eric expects to supplement the book with guidance on local LLM options beyond Ollama, including LM Studio.
  • AI evaluation can be done by saving model traces/outputs and using another model as a judge, but model comparisons can invert expectations if prompts are better tuned for a weaker model.
  • Teams can codify AI guardrails via persistent repository files (including cross-provider concepts like agents.md), but cross-provider compliance is not guaranteed.
  • OpenAI's API is usage-based and token-metered, and it is distinct from the ChatGPT subscription product billed monthly.

Sections

Prompting As Engineering Asset And Context Management

  • Prompt engineering is iterative, and once a good prompt is found it should be made persistent (copyable or checked into Git) because additional verbosity can hit diminishing returns.
  • Teams can check a reusable prompting-context file into Git so everyone uses the same boilerplate without retyping each session.
  • Models can perform better on technologies with extensive public documentation and open development history because training data includes those artifacts.
  • Internal standards such as naming conventions can be injected into LLM context via repo files checked into Git or via boilerplate instructions pasted at chat start.
  • Prompt specificity improves usefulness because LLMs do not reliably infer unstated preferences; providing explicit technical context, examples, and output constraints increases the likelihood of a useful response.
  • System messages can provide persistent context across chats, reducing the need to repeat instructions such as preferred technologies, roles, and response format.

Learning Path And Enablement Scaffolding

  • Eric expects to supplement the book with guidance on local LLM options beyond Ollama, including LM Studio.
  • The AI Networking Cookbook is organized from foundational AI setup to advanced topics including MCP, building copilots, and evaluating cost/performance tradeoffs.
  • The book intentionally uses OpenAI as the primary provider early on to reduce beginner overwhelm while teaching transferable concepts like temperature and system messages.
  • All code examples from the book are available in a public GitHub repository and are commented so they can be followed without the book.
  • Community content moved from networkanimation.community to networkautomation.com, which offers free signup notifications and a Discord for real-time Q&A.
  • The intended audience is a network engineer with basic networking knowledge (around Network+ or CCNA) and basic comfort reading Python.

Model Selection And Evaluation Pragmatism

  • AI evaluation can be done by saving model traces/outputs and using another model as a judge, but model comparisons can invert expectations if prompts are better tuned for a weaker model.
  • Competent LLM results often depend on selecting an appropriate model (not the cheapest default) and iterating prompts until the request is correctly specified.
  • Eric currently prefers Claude Sonnet 4.6 with thinking mode for coding tasks and often uses Gemini for browser-based workflows where it can reference open tabs.
  • Comparing outputs across multiple models is most worthwhile for high-stakes decisions or systematic evaluation and is usually unnecessary for routine tasks once a model is good enough.
  • Eric expects that for low-risk coding-assistant tasks, output quality can be high enough that extensive double-checking may feel less efficient than trusting the assistant.

Guardrails Interoperability And Versioned Governance

  • Teams can codify AI guardrails via persistent repository files (including cross-provider concepts like agents.md), but cross-provider compliance is not guaranteed.
  • Combining global and project-local guardrails can constrain LLM behavior, with more specific local instructions overriding broader defaults.
  • Guardrail files used to guide assistants are typically plain-English Markdown rather than a strict structured format.
  • A proposed safer workflow is to start in planning/ask mode with a coding assistant, then summarize agreements into a file checked into a centralized repository for team reuse and change history.

Cost Pricing And Initial Access Friction

  • OpenAI's API is usage-based and token-metered, and it is distinct from the ChatGPT subscription product billed monthly.
  • Eric estimated that about $20 of OpenAI API credit was sufficient for his book examples and was not exhausted during his work.
  • Many AI providers now require credit card verification or paid trials rather than broad free tiers, increasing upfront requirements to get started.

Watchlist

  • Eric expects to supplement the book with guidance on local LLM options beyond Ollama, including LM Studio.
  • Eric reported hearing that Mac minis are in high demand for local LLM use and that this popularity has contributed to shortages.
  • Teams can codify AI guardrails via persistent repository files (including cross-provider concepts like agents.md), but cross-provider compliance is not guaranteed.

Unknowns

  • Which specific AI providers changed free-tier and verification requirements, and when did those changes occur?
  • What token usage, model choices, and task mix underlie the approximately $20 API credit datapoint, and how variable is it across common network-automation tasks?
  • What are the measurable performance thresholds for local models (latency, throughput) across different hardware configurations and with/without GPU offload?
  • How consistently do major assistants read and obey repository guardrail files (including agents.md), and what failure modes exist when they do not?
  • What empirical success-rate improvements come from structured-output instructions (JSON/YAML) versus unstructured prompting for automation pipelines?

Investor overlay

Read-throughs

  • Rising local LLM experimentation could lift demand for compact, efficient desktop hardware, implied by reported high demand for Mac minis tied to local LLM use.
  • Growing use of repo-native guardrail files and structured prompting could increase demand for developer tooling that standardizes prompts, context, and governance across teams.
  • Token-metered API usage and reduced free tiers may increase sensitivity to cost controls, evaluation tooling, and prompt efficiency practices to manage variable operating expense.

What would confirm

  • Sustained reports of shortages or elevated demand for small form factor machines specifically attributed to local LLM workloads, plus continued additions of guidance on local runtimes beyond a single tool.
  • Broader team adoption of persistent prompt assets and repo guardrail files, with evidence of workflows standardizing structured outputs and versioned context artifacts in repositories.
  • More emphasis on metered usage management such as trace capture, model judging, and prompt tuning practices adopted to reduce token spend and improve reliability.

What would kill

  • Local LLM usage interest fades or shifts away from desktop hardware, with no ongoing signals that small machines are constrained or in unusually high demand for this purpose.
  • Guardrail files remain nonportable across assistants and fail silently often enough that teams abandon repo-native governance rather than standardize on it.
  • API access friction and free-tier reductions do not materially change behavior, with experimentation remaining driven by subscriptions rather than token-metered API usage.

Sources