How to Harness Large Language Models for Ten‑Fold Software Development Efficiency

To make large language models (LLMs) reliable partners that boost software development efficiency by up to ten times (instead of the current ~17%), you must feed them sufficient, relevant information.

The discussion builds on a previous piece about Software Engineering 3.0 and concretizes the inputs and outputs for each development activity.

1. Determining requirements

Ideally a single sentence like “build a simple e‑commerce site” would suffice, but in practice you need to provide:

The system’s existing functionality (full requirement documents) and select the portions relevant to the new feature.

Domain‑specific knowledge if the business area is specialized, preferably via a vertical LLM or curated training material.

Collaborative outputs should include:

A detailed, rigorous description of the new requirement (e.g., a user story with acceptance criteria).

A complete, well‑structured requirement document for the new feature.

Implementation tip: store the full requirement markdown in a Git repository; each new requirement becomes a commit or feature branch.

2. Software design and implementation

Typical LLM assistance is fine‑grained (code completion, comment generation). To let an LLM produce a full implementation for a specific domain you must supply:

The new requirement itself and relevant excerpts from the full requirement docs.

The current source code that will be modified, selected manually or automatically.

Clear naming conventions and accompanying design/description texts (repo README, file headers, interface docs, architecture documents) that help the LLM understand the code base.

Outputs of this collaboration are:

Commits that implement the requirement.

Updated design and description documents synchronized with the code changes.

3. Software testing

For effective testing the LLM needs:

The requirement item (user story + acceptance criteria) and related excerpts from the full requirement docs.

Existing test cases, scripts, and data that are relevant to the new test.

The LLM can generate or modify test cases/scripts/data, which after human review are executed automatically (e.g., UI automation). Future work may enable agents to execute textual test cases directly.

Outputs include a complete set of test cases/scripts/data for the new requirement stored in the Git repo, with incremental changes represented as commits or feature branches.

4. Summary

The article lists the inputs to provide LLMs and the collaborative outputs at each stage of software development, emphasizing that current practice is still far from the ideal ten‑fold efficiency and that substantial improvement space remains.

The method works best for new projects that adopt this workflow from the start; legacy projects may see low ROI when retrofitting.