Articles/Platform Integration

GSD Core: A Phased Build With Two Human Gates

Platform IntegrationTool Spotlight
By Vanya SmytheOctober 19, 20264 min readVerified Oct 19, 2026
gsd(primary)
gsdgsd-coreclaude-codespec-driven-developmentcontext-engineeringcommissioningmit

What GSD is

GSD is short for 'Get Shit Done'. It is a set of slash commands, agent definitions and file templates that runs inside an AI coding agent. It makes the agent work in planned phases, not in one long chat.

Lex Christopherson created the original repository on 14 December 2025 under the MIT licence. He publishes as TÂCHES, and the README describes him as a solo developer. On 20 October 2026 the repository had 64,269 stars and 5,430 forks. Its last commit was dated 31 May 2026, and the repository is archived. Its npm package, get-shit-done-cc, stopped at version 1.42.3 on 17 May, after 223 published versions.

The move to GSD Core

The archived README now carries one message: GSD 'continues as GSD Core'. GSD Core has been renamed 'Git. Ship. Done.' It is published by the Open GSD organisation, which has its own site at opengsd.net. It is also under the MIT licence, and it was created on 22 May 2026.

On 20 October GSD Core had 10,685 stars, 763 forks and 292 open issues. Its latest push was the day before, and release v1.18.0 was dated the same day. Releases arrive roughly weekly.

The archived repository does not say whether the fork was a handover. I searched on 20 October and found nothing published that settles the question. The single sentence in the README is the whole record.

GSD Core installs with npx @opengsd/gsd-core@latest on Node 22 or later. It supports Claude Code, OpenCode, Codex, Copilot, Cursor, Windsurf and several other agents. It needs no API key beyond the agent's own.

The loop and its files

A project starts with /gsd-new-project. This command asks questions and sends out four researcher agents in parallel. A synthesiser merges their findings, and a roadmapper splits the work into phases. The command writes PROJECT.md, REQUIREMENTS.md, ROADMAP.md and STATE.md into a .planning folder in the repository.

Each phase then runs four commands:

  • /gsd-discuss-phase puts the undecided questions to a person and records the answers in CONTEXT.md.
  • /gsd-plan-phase runs a researcher, a planner and a plan-checker in a loop until the plans pass.
  • /gsd-execute-phase gives each plan to an executor. Where the plans allow it, executors run in parallel groups, and 'each task gets its own atomic commit'.
  • /gsd-verify-work runs a verifier, takes a person through acceptance testing and writes fix plans for anything that fails.

Finally, /gsd-ship opens the pull request.

The context-rot claim

The project's central claim is that it 'solves context rot'. Context rot is the loss of quality that sets in as a long AI session fills up. GSD runs research, planning and execution in separate subagents that each start fresh, so that 'each executor starts with a clean 200k-token context' while the main session stays small.

The mechanism is plausible and the files are real. Neither repository publishes a benchmark, however. The fair description is a system designed to limit context rot, not one shown to cure it.

The README is also open about cost. It says the research, plan-check and verifier agents 'add tokens and time'. It offers quality, balanced and budget model profiles and does not give a figure.

Why I decline the skip-permissions advice

The README recommends running Claude Code with --dangerously-skip-permissions, so that the loop runs without interruption. For a firm's repository I think that is the wrong choice. An agent that can run any command without asking can also delete, push or fetch anything its shell can reach. An atomic commit does not make up for a credential sent somewhere it should not go.

A commissioning tool built with it

The solution applies the loop to a small battery-storage integrator that is building an internal tool for commissioning reports. The tool reads the test CSVs, checks them against the firm's acceptance limits and drafts the report.

The engineer answers discuss-phase from the limits and the report template that the firm keeps in the repository. The engineer also goes through acceptance testing in verify-work before anything ships.

The design adds four controls around GSD. It keeps the agent in a sandboxed container with its permission prompts on. It pins the GSD Core version. It runs continuous integration on the shipped pull request before the merge. It starts on the balanced profile and watches the cost of each phase.

See the design, its diagram and the demo at the solution page.

I think the loop is sound. I do not follow its advice on permissions.

Subscribe to engineering insights

Get notified when we publish new technical articles.

Topics

Unsubscribe anytime. View our Privacy Policy.