Judge.me

Research

AI & Engineering

Whitepaper

Engineering Claude Code for Long-Lived, Multi-Session Projects

Keeping a project workable across resets and usage limits by treating state as files on disk, with disjoint roles for a Producer and a Composer.

G. Destefanis · UCL

Undated

A practical technical note, presented separately from our published academic papers.

Orient your reading

Keep the project coherent across sessions.

This technical note applies explicit state, version control, separation of concerns and reproducibility to long-running work with AI agents. It distinguishes subagent orchestration from independent sessions coordinated by a person.

Editorial context · The source paper remains the reference.

A long project cannot depend on a conversation remembering everything. This technical note applies explicit state, version control, separation of concerns and reproducibility to work that spans multiple AI-agent sessions.

Key takeaways

Key takeaways

  • Keep decisions, current state and resumable work in project files.

  • Distinguish one agent orchestrating subagents from independent terminals working side by side.

  • Make file ownership and handoffs explicit before running concurrent work.

Move project memory into files

The paper describes conversation context as finite, volatile and private to a session. A project folder gives a new session a durable starting point. Recording decisions while work proceeds reduces the need to reconstruct them after a reset.

Separate production from composition

A Producer maintains source material and project memory, while a Composer creates the deliverable. In the arrangement described, a person relays information between the two and file ownership stays disjoint. The purpose is to make each role’s authority and inputs legible.

Make concurrent work resumable

Independent terminals introduce coordination risks that differ from subagent orchestration. The note combines explicit ownership with version control, recorded inputs and a ledger for long-running jobs, so an interrupted task can resume without repeating completed work.

A practical engineering account

The document offers conventions, failure modes, a worked example and templates. It is a practical account of organising multi-session projects, rather than a measured comparison of tools. No publication date appears in the supplied document.

Cite this paper

Giuseppe Destefanis. (n.d.). Engineering Claude Code for Long-Lived, Multi-Session Projects. Technical note.

GD

University College London

Giuseppe Destefanis is an Associate Professor in UCL’s Department of Computer Science. His research spans empirical software engineering, mining software repositories and large language models. He uses data mining, machine learning and natural language processing to study how people, technical systems and economic incentives interact across digital platforms.