Back to Work

Engineering discipline for agentic coding

A public repository template that lets teams build quickly with coding agents while keeping product intent, review, security checks, and project learning in the workflow.

Role
Author of the method and template
Organisation
Maple Spark Labs
Timeframe
2026–present
Status
Published
Kind
Method
Technology and standards
  • Claude Code
  • Git
  • GitHub
Focus
  • AI-assisted development
  • Product intent
  • Code review
  • Project context
  • Documentation
  • Human review
  1. IntentState what the change is for and what must stay true.
  2. PlanChoose a small, reviewable implementation step.
  3. BuildUse AI to implement a small, reviewable change.
  4. ReviewReview the change and run the relevant checks.
  5. LearnUpdate the project context with decisions and lessons.
A development loop supported by repository instructions and commands. Following the loop remains the user's responsibility.

Problem

Coding agents can produce software faster than a project can preserve the reasoning behind it. Without shared context and review gates, speed increases drift: the agent loses the product intent, repeats old mistakes, and ships issues that should have been caught before the commit.

Constraints

The discipline had to live in the repository and fit ordinary development. It could guide an agent and a person, but it could not pretend that prompts enforce branch protection, replace security tools, or guarantee human judgment.

Approach

I encoded the product and engineering practices I use into repository instructions, durable project context, and a set of commands for agentic development.

The workflow starts by making implementation wait: explore the existing system, clarify the problem, and identify dependencies. It then creates a tracked plan, implements a reviewable change, runs a pre-commit review across security, code quality, infrastructure, and tests, and records the decisions and lessons the next change needs to inherit.

The point is not a better one-off prompt. It is scaffolding that puts an agent into the same disciplined cycle every time, while leaving the person responsible for what gets built and shipped.

Current state

The template is public on GitHub under an MIT licence. I use it in my own work, including Droplet Learn and this site.

Evidence

See claude-project-template (opens in a new tab) , Read I didn't change how I work with AI. I encoded how I think. (opens in a new tab) 

What I learned

The prompt is not the operating system. Product context, explicit review, and a durable record of decisions are what let agentic coding move quickly without losing direction.

Related work