2025 guide to Context Engineering for Software Engineers
How context engineering techniques like RAG, tuning the system prompt, web search, and MCP tools create a big impact, backed by experience (I share my context-engineered code review process)
Everyone is waiting for the next exponential leap. The arrival of GPT-5 showed we are not making exponential improvements or reaching a super-intelligent model soon. The jump from one version to the next is gradual. The engineers getting ahead are not just swapping today’s model for tomorrow’s slightly smarter one.
The real performance bottleneck for companies isn't the model’s intelligence, it's the quality of the context you give it. The breakthrough comes from context engineering: designing the entire information ecosystem for your AI and integrating it into your workflows. Stop waiting for a smarter AI and start building the systems that give you a 10x career leap today.
Companies reward force multipliers, not solo coders. Context engineering lets you automate repeated cognitive work, reduce review friction, and own outcomes that move the business forward.
⭐ In this post, you'll learn
Why context engineering is a high-leverage skill
The techniques that make AI output faster and more accurate
How to apply context engineering to real engineering workflows
How to turn context engineering results into promotion-ready evidence
Why context engineering is skill to focus on in your company
Even the best LLM has zero awareness of your company’s architecture, business rules, or team culture. It does not know your microservices, your coding standards, or the historical context from a six-month-old Slack thread. Without this, its output will be generic and wrong in subtle ways.
Imagine this: You've been at your company for three years, and now a new hire comes in with higher IQ. You'll outperform this new hire on its first day. You ship faster because you know the approval flows and hidden dependencies. You prevent regressions because you remember past incidents and edge cases. You reduce cycle time because you know exactly where to get answers without waiting. You navigate cross-team dependencies because you know who to involve and when.
You do not create impact in your team because you are the smartest engineer. You create impact because you understand the business domain.
The same applies to AI. A model without context is like a new hire without onboarding. Feed it the same environment you operate in and it becomes a high-speed extension of your brain. That is the essence of context engineering, giving AI the same insider advantage you have.
Context engineering techniques
Context engineering isn't “just prompting better.” It is using a set of repeatable techniques to give the AI the right information, in the right structure, at the right time.
1. System instructions: role and persona prompting Set a clear role or persona for the model before it responds. For example, "You are a principal software architect explaining event-driven systems to a mid-level engineer". This shapes the tone, detail, and perspective of the response.
2. Retrieval Augmented Generation (RAG) Combine the model with external knowledge sources that you put in a vector DB. Instead of relying solely on the model's internal data, retrieve relevant documents and feed them into the prompt so answers are accurate and current. You don't need a PhD for this, this vectors DB can be a PostgreSQL with the pgvector extension.
3. Short and long term memory patterns Use different memory scopes intentionally. Short term memory is the conversation context within this chat. Long term memory, like ChatGPT's persistent memory, retains facts and preferences over time for personalization.
4. Web Search Grounding Enhance answers with real-time search results. Ask the model to search for the latest information, then ground its reasoning in that retrieved data.
5. Prompting Explore multiple prompting techniques to control model output.
Constraint-based and negative prompting: Specify what the model must do and what it must avoid.
Prompt layering and lazy prompting: Break complex instructions into layers, providing the next layer only when needed.
Prompt pipelines and chaining: Feed the output of one prompt into the next for multi-step tasks.
Zero-shot prompting: Give the task with no examples, relying on model knowledge.
Few-shot prompting: Include a few examples to guide the output style and accuracy.
6. MCP tools Extend model capabilities with external tools, APIs, or plugins. These can perform calculations, access private data, or run specialized tasks the model alone cannot do, or it wouldn't be efficient to do it with AI instead of code.
7. Reasoning: chain of thought, tree of thought, and decomposition Guide the model to reason step-by-step. Chain of thought walks through linear reasoning, tree of thought explores branching possibilities, and decomposition breaks problems into smaller, manageable parts.
8. Workflow orchestration Coordinate multiple AI agents or tools in a sequence. Each agent can handle a specific task, passing results to the next to build complex workflows. They can communicate with protocols like A2A (Agent to Agent, by Google) or ACP (Agent Communication Protocol, by IBM)
9. Input/output context Clearly define how input is structured and how output should be formatted. For example, "Input is JSON with fields X and Y. Output a Markdown table summarizing the key values with columns A and B".
The most useful apps are the ones that allow you to use as many of these context engineering techniques as possible:
A practical example
Here is what the shift looks like in a real workflow I implemented at work.
The old way:
Open all the files in a PR, cross-reference with the backlog in JIRA and the design doc, rely on memory for coding standards, then spend 45 minutes writing comments. Every step depends on your ability to juggle context in your head.
The new way (this is only an MVP):
Aggregate your context in minutes. I use a simple greasemonkey script to copy all PR files into my clipboard at once, along with ticket descriptions from the PR descritpion and relevant design docs from my browser bookmarks. This is a "manual RAG" step that takes under two minutes because my team doesn't have yet our documents in a vectors DB.
Then I load my prompt. It includes the role (“Staff Engineer”), the coding standards, the expected output format, and reasoning instructions. I give it examples of good comments in our style, and I embedded inline our coding standards. I even tell it to check external libraries against public vulnerability data searching the web. That prompt lives in my prompt library, which is my long-term memory for code reviews.
Once the AI outputs its review in a Markdown table, I become the final filter. I discard low-value comments, keep the high-value ones, and add my own. The review is faster, more consistent, and leaves a paper trail I can reuse.
The image below reminds me of the human flaw we want to avoid with AI. We want AI to prevent all the comments as possible, so the human can focus in anything that's really important.
Convert context engineering wins into promotion evidence
Managers do not care about “I used AI.” They care about measurable impact: higher velocity, lower risk, and fewer cross-team blockers. You have to turn context engineering into proof.
Collect before-and-after metrics. Keep links to your prompt library, demo videos of your scripts in action, and decision log entries showing minutes saved. Build your promotion doc like a case study: problem statement, approach, measurable result, sample artifacts, and endorsements from stakeholders.
Scale your influence. Publish your context vault internally, run a one-hour workshop, and create templates others can reuse. Pair with two engineers, adapt your playbook to their work, and record their results. Add those wins to your promotion evidence. The more you are seen as the person improving how others work, the stronger your case becomes.
🎯 Conclusion
Pick one boring, repetitive, or ambiguous task this week. Design a context pipeline for it, ship a minimum viable automation, measure the minutes saved, add it to your decision log, and bring the results to your next 1:1.
If you don't know how to do this, just check the 2nd and 3rd section in this article. Those are the tools and you just need to do an MVP, like the example about how I applied context engineering to my code reviews.
Ship the automation, show the delta, and ask for more responsibility in your next performance review.
🗞️ Other articles people like
👏 Weekly applause
These are some great articles I’ve read last week:
Your STAR Method Is Broken (Principal Solutions Architect, AWS) by
and . Using the STAR format effectively can improve your interview results.Second Year As a Solopreneur. Here's What I Did: by
. Building in public and seeking feedback is the only way to build a successful product.Why You Can't Afford to Ignore TypeScript? 🌟 by
. Investing in TypeScript imrpoves code quality and reduces bugs.Your First 90 Days as a Tech Lead by
. Your first 90 days as a tech lead should focus on building trust, understanding team dynamics, and setting clear expectations.HTTP Headers to Build 10X APIs 🔥 by
. Understanding HTTP headers is crucial for software engineers.The Art Of Productive Impatience by
. Productive impatience works to overcome inertia in both projects and career growth. Take ownership of your outcomes.
🙏 One last thing before you go:
I am always working to make this newsletter even better.
Could you take one minute to answer a quick, anonymous survey?
See you in the next email,
Fran.
Nice article, I often wonder about something that I call context-wallet, a portable context that we can plug on different model - I expanded a little bit here
https://www.tostring.ai/p/brag-brain-retrieval-augmented-generation?r=clbp9
Love this AI deep dive!
Thanks for mentioning my post 🙏