Ask HN: Techniques for learning things quickly using coding agents?
2 points
7 hours ago
| 1 comment
| HN
Lately, I am finding myself using coding agents primarily for investigating and learning new topics, which include:

    * understanding how specific module works in codebase
    * overall architecture in the new codebase
    * onboarding faster in the new codebase
    * researching a topic
    * analyzing video content/meeting transcripts to quickly catchup with the topic
and so on.

My current workflow is just to point out to repository or folder with transcripts and ask agent to extract insights (assuming it will first analyze the content, then suggests ways and whats to extract) and reading bunch of output markdown files.

But I feel like this workflow can be improved, but no idea where to start, what are the cognitive methods and strategies.

How are you doing it? What strategies, skills do you employ to quickly learn new things

* in this context. learning == primarily onboarding in a new topic, because learning and being productive on anything takes time

lemonademan
5 hours ago
[-]
Moving from generating code to mapping code is already a major leap. But to break through the "feed data and read Markdown" plateau, you must transition from passive reading to active, structured prompting.

You can achieve this by moving from summarize to deconstruct. For example, standard agents default to generic summaries. You must force the agent to act like a senior engineer conducting a code review or a research lead. You can do this by prompting for architectural boundaries and friction points, not just descriptions, by asking "why", not "what".

You could also use tactical prompts for Code onboarding. An example prompt would be, "trace a single user request from the entry point API down to the database. List every module it touches and where mutations happen."

This is how I would do it if I were in your shoes, but there are certainly other methods that could be better suited for you. Either way, I hope this was at least useful in some way.

reply