System

Task-first coding-agent system

Boilerplate, rules, specifications, and feedback loops used to ship production SaaS with coding agents.

The “system” here is a workflow I refined while building with Cursor, Bolt, Lovable, and an internal TypeScript boilerplate: requirements first, conventions in rules, domain folders, bugs as tasks, new chat per task. It is how I currently run AI-assisted engineering at Sqaleup, and it is also a research object: evidence that specification and environment explain a large part of agent usefulness.

Problem

Coding agents can generate a large amount of code quickly, but speed does not remove product ambiguity. In early workflows I repeatedly saw agents spend time inventing stack decisions, guessing requirements, carrying stale conversation context, and solving the wrong version of a task.

System

I gradually moved toward a task-first engineering system: begin from a known TypeScript foundation, store durable conventions in project rules, write requirements and expected behaviour before asking for implementation, keep domain structure visible in the codebase, represent bugs as explicit tasks, and open a fresh conversation when the task changes materially.

Never start from scratch when a domain boilerplate already encodes decisions.
Write 10–20 minutes of concise requirements first. AI is fast pattern matching, not a mind reader.
Generate and maintain rules so the agent learns file placement, naming, and types.
Point it at existing UI and logic rather than asking it to invent a new house style.
Build in a feedback loop. Force analysis before shipping.
New chat per task beats one endless thread.

Autocomplete predicts code. An engineering agent tries to complete an objective. Completing an objective can require searching the repository, understanding dependencies, modifying several files, running commands, inspecting errors, and changing the implementation. That change requires more state, more context, more tools, and more verification. Permissions become part of the product. Specifications become part of the context pack.

Why it works

The workflow reduces the number of decisions the agent has to invent. The boilerplate communicates technical preferences. Rules communicate durable constraints. The task communicates intent. The codebase communicates existing architecture. Tests, previews, and runtime feedback close the execution loop.

Research significance

This system became evidence for a broader conclusion in my work: coding-agent capability is partly a property of the model and partly a property of the environment, specification, context, and feedback provided around the model.

Related experiments

EXP-008 covers vague prompts versus task-first requirements. EXP-011 covers infinite threads versus task-scoped conversations. EXP-012 covers blank repositories versus an opinionated engineering foundation.

Use in practice

This is not a standalone hosted lab product. It is the engineering workflow I use to build production SaaS and MVPs with coding agents, and it continues to evolve as the tools become more capable.

Abubakar, Imam. "Task-first coding-agent system." Imam Abubakar AI Research, 2025. https://imamabubakar.com/research/task-first-coding-system

This archive contains independent research, experiments, architecture studies, and technical observations. Some findings are exploratory and should not be interpreted as peer-reviewed scientific conclusions unless explicitly stated. Architecture studies of third-party products are based on publicly observable behaviour, published technical information, and engineering analysis. They are not claims of access to private source code or infrastructure.