Tips for Using LLMs in Software Development

Background

Context is everything when working with LLMs. Here are five practical levers that determine whether your agent session succeeds or wastes a lot of tokens.

This article explores the minimum amount of context you need to provide to an LLM in order to perform a given software development task. When answering a user's prompt, an agent harness (like Claude Code or Cursor) can perform additional work such as reading files, exploring trial solutions, and updating existing source code. We want the LLM — together with its harness — to succeed at whatever task we provide, but we need to be thoughtful about the context we give it.

Context 1: Project Folder

For the LLM to efficiently reach application source code, database schemas, table structures, and markdown documents, it is best to start the agent harness inside the project folder, or just by specifying the project folder for the conversation. The harness will allow the LLM to list directories, search for keywords within the project, and examine what libraries are being used (e.g., `requirements.txt`). 

Context 2: Task Description

Describe clearly to the LLM what needs to be done. Where possible, provide both the objective and the method to perform the task. The more specific you are, the less time the agent spends guessing at intent.

Context 3: Definition of Done

For some tasks, the definition of done is simply that the application logic is implemented. A stronger task definition also includes a validation mechanism — for example, agreement of results within a tolerance of ±1E-8 compared to a reference result. In that case, you should also provide the location of the reference result files (whether CSV or XLSX), as well as the location of your application's output (a database table or a directory path).

If the structure of the reference result and the application result differ, a mapping between the two terminologies must also be provided so the LLM can reconcile them correctly.

Context 4: Limiting Conditions

This is optional, but it is worth setting boundaries to prevent the agent harness from burning tokens endlessly on unproductive paths. One alternative is for the user to press the stop button when too much effort is being spent with no good result. Sometimes it is better to pause and ask the LLM to report its current status, so you can decide what the next best action should be.

Context 5: Session Continuity

After roughly 10 turns of conversation — which may involve no user input at all, just the results of Python script executions or grep searches — the context load grows significantly. Every prompt, response, and file read is tokenized, and the full history is sent on every subsequent turn. While there is some caching, context usage always trends upward.

A practical solution is to ask the LLM to write a session summary to a markdown file before ending the session. This effectively compresses the history, removing lower-value tokens from the next iteration. You can then start a fresh conversation, point the LLM to the summary file, and continue where you left off.

It also helps to keep the task definition in its own dedicated markdown file — stored in a `specs/` or `docs/` directory — so that any new session can read both the task context and the current session status before getting to work.

Conclusion

In short, a well-structured context reduces wasted turns and keeps your agent on track. Start with these five elements — project folder, task description, definition of done, limiting conditions, and session continuity — and refine based on what your specific project needs. The investment in setting up context upfront pays for itself quickly in faster, more reliable results.


Comments

Post a Comment

Popular posts from this blog

Long running process in Linux using PHP

Background To do stuff, I usually create web-based applications written in PHP. Sometimes we need to run something that takes a long time, far longer than the 10 second psychological limit for web pages. A bit of googling in stack overflow found us this  http://stackoverflow.com/questions/2212635/best-way-to-manage-long-running-php-script , but I will tell the similar story with a different solution. One of the long running tasks that need to be run is a Pentaho data integration transformation. Difficulties in long running PHP scripts I encountered some problems when trying to make PHP do long running tasks : PHP script timeout. This could be solved by running set_time_limit(0); before the long running tasks. Memory leaks. The framework I normally use have a bit of memory issues, this can be solved either by patching the framework (ok, it is a bit difficult to do, but I did something similar in the past ) or splitting the data to process into several batches. And if ...
Read more

Migrating Joget 3 Instance

Image
On some occassions we need to move joget 3 workflow service from one server to another. We need to migrate two part : the server application files, and the database. Our aim is to migrate joget server with data intact. Identify running instance First we need to determine where joget is installed. In an ideal situation this information is stored in CMDB. But since we have yet to see Joget - Cloud foundry package, and we haven't deploy CMDB enterprisewide, this step is necessary. Look  for java processes in the source machine : ps auxw | grep java In my machine we got these output : 501      26733  0.0 22.2 759176 457552 ?       Sl   Apr26  18:20 /home/admin-dev/p/jdk1.6.0_21/bin/java -XX:MaxPermSize=128m -Xmx512M -Dwflow.home=./wflow/ -Djava.util.logging.manager=org.apache.juli.ClassLoaderLogManager -Djava.util.logging.config.file=/home/admin-dev/p/joget-linux-3.0.0/apache-tomcat-6.0.18/conf/...
Read more

Copying Big Oracle tables Using Apache Spark

Background Sometimes we need to copy table data from one database to another. Logically the best way to do this is to do database specific export (expdp in oracle lingo) and import in the destination database (impdp in oracle). But sometimes there are deficiencies in this method such as unable to do parallel process for single table, and requirements of DBA access in the database. This post shows how to do table copy using Apache Spark and Apache Zeppelin. Preparations In order to allow Apache Spark access to oracle jdbc connections, we need to add dependency to ojdbc6.jar. To do this, write this paragraph in Zeppelin : %dep z.reset() z.load("/path/to/ojdbc6.jar") Basic Approach The most basic approach to copy table data is to retrieve data in one query and save the  resulting records in the target database. val tableNameSrc = "TABLENAME" val tableNameTrg = "TABLENAME" import java.util.Properties Class.forName("ora...
Read more