GSD Core

A framework that sits between a person and whichever coding agent they run, and drives it through a disciplined loop instead of a chat. Each milestone repeats five steps, one phase at a time: discuss the implementation decisions before anything is planned, plan and have the plan checked against the context it will run in, execute the plans in parallel waves so that every executor starts from a clean window, verify by walking through what was built, then ship. The problem it is aimed at is context rot — the quality loss that accumulates as an agent fills its window — so research, planning and execution happen in fresh subagents and the main session is kept thin. All state is file-based under a planning directory, as human-readable Markdown and JSON: project brief, requirements, roadmap, state and configuration, with no database and no server, so it survives a context reset and can be committed. Installation is a single command, after which the installer asks which runtime to target and whether to install globally or into one project.

Who built itAn organization account rather than a person, which is why the attribution splits so unevenly. The repository has 6,046 commits and 50 listed contributors: trek-e, whose commits carry the name Tom Boucher and the address trekkie@nomorestars.com, holds 3,742 of them; glittercowboy holds 945; davesienkowski 133, Tibsfox 127, jeremymcs 122, github-actions[bot] 95, 0xdhx 82. 5,829 of the 6,046 commits are linked to an account and 217 are not. Of the 3,143 co-author trailers, the largest single entry is not a model at all — it is sim, on 540 — while 2,309 of them, spread across 19 differently named Claude models, credit one, the biggest being Claude Sonnet 4.6 on 465.

1. 01

   A history older than the repository that holds it

   The metadata says the repository was created on 2026-05-22 under the open-gsd organization, but the oldest commit it carries is dated 2025-12-14 and is titled “Initial commit: Get Shit Done - meta-prompting system for Claude Code”, so the project began under another name and the history came with it. The rename was not left as a repository setting: it survives in the tree as installer migration 003-rename-get-shit-done-to-gsd-core.cts, an archived changeset file, and docs/cleanup-get-shit-done-cc.md — which is what a rename has to become once users already have files installed on their disks. The default branch is next, not main. In ten months the repository took 6,046 commits: 101 in December 2025, then never fewer than 276 in a month, peaking at 1,028 in the month it was created. Around that sit 10,038 stars, 719 forks, 43 watchers and 186 open issues, on a tree of 3,630 files and 76 MB. The newest commit, on the last day of the record, is a fix that labels a user’s typed arguments in every argument-taking command template.

2. 02

   The process is itself the largest artifact in the repository

   The slogan is Git. Ship. Done., and the project runs that loop on itself, in public. There are 98 architecture decision records under docs/adr/, each numbered after the issue that produced it, from a 1,405-byte note to a 126,874-byte design for one owner of the planning semantic model; the rest of the long ones cover enforcement by construction, an embeddable orchestration engine, and one owner per workflow verdict. The index above them is generated by a script and gated by an 87 KB test, so a new record that is not indexed fails the build. Above the records sit epics split into numbered phases: epic #5056 is designed by ADR-5057 and delivered as at least seven phases, and each phase arrives as its own issue, its own pull request, and a fail-first test written to fail against the current code and closed before the fix lands. The pull request bodies cite the ADR section and the phase entry by name, and a phase that changes behaviour ships with a changeset entry. The repository also keeps its own output: .gsd/phase/feat-3677-quick-batch-hardening-acceptance/ holds a 29 KB design document, a test matrix and 12 KB of acceptance evidence.

3. 03

   Thirty-one workflows, one for each rule that stopped being a convention

   .github/workflows/ holds 31 workflows, and their names read as a list of decisions about what to stop enforcing by hand. pr-target-validator greets a pull request opened against main from an ordinary fix branch with a comment saying most pull requests should target next, then lists the four exceptions: release branches, hotfix branches, critical-fix branches for a production outage, and back-merges. pr-template-format requires one of three typed templates and refuses a body without the matching heading; when a contributor’s patch was otherwise sound, the automated objection was answered inside the hour by retargeting the branch and adding the heading, with a note that the patch itself was unchanged. auto-branch creates the branch and comments the two commands that check it out. require-issue-link, changeset-required, docs-required, duplicate-check, duplicate-sweep, dismiss-unauthorized-pr-approvals, auto-close-unsolicited-prs, branch-cleanup, stale and a version-gate fill out the set. One of them was itself broken: the Dependabot auto-merge job declared only pull-requests: write, so gh pr merge --auto failed on every dependency pull request until the job was granted contents: write as well.

4. 04

   The test suite is the biggest directory in the tree

   tests/ is 1,247 files and about 29 MB, larger than the source it tests, with 997 of those files sitting directly in the directory and the rest split between 146 fixtures, 35 helpers and a 55-file QA set. The runner is a 119 KB script. Beside the ordinary test files there are property tests, nine unit.test.cjs files and an adversarial/ fixture directory with separate corpora for frontmatter, roadmap, security and TOML parsing. Static enforcement is the second wall: 32 custom ESLint rules live under eslint-rules/, among them no-adhoc-markdown-parsing, no-source-grep, no-tautological-assert, no-magic-sleep-in-tests, no-posix-mode-bit-assert and no-rendered-text-length-assert, rules that exist because each of those mistakes had already been made. A further 55 lint-* scripts, 22 gen-* generators and seven check-* guards sit in scripts/ beside their committed allowlist files. Several are ratchets with a committed ceiling file, so the number may only fall: a mutation-score ratchet, an allowlist ratchet, a test-file-count ratchet, and a history file that records every CI job’s wall clock against its timeout and is republished by one rolling pull request.

5. 05

   Budgets measured in bytes, and only ever tightened

   The most distinctive engineering here is aimed at prompt text rather than at code. Workflow files are loaded verbatim into the agent’s context every time the matching command runs, so they are held to a byte budget enforced by a test, in tiers of 90,000 bytes for the three top-level orchestrators, 54,000 for large workflows and 38,000 for a focused one, and the ceiling may only fall. The reasoning is written down: bytes rather than lines, because line count over-penalizes prose and under-catches token-dense tables and code blocks; the unit adopted from a vendor’s truncation limit, but not that vendor’s number; and the caveat that extraction only helps when the extracted file is read at the step that needs it, since prose moved into a file that is still eagerly imported shrinks the measurement rather than the context. The same pressure produced workflows/discuss-phase/, where the parent became a dispatcher and per-flag bodies moved into nine mode files, and .compact.md variants for 29 of the agent definitions. A second budget governs skills: a flat listing of 86 skills costs roughly 2,150 tokens on every single turn, so six namespace routers costing about 120 tokens were layered above them, using pipe-separated keyword tags on the strength of published routing research.

6. 06

   What is forbidden, and what the users find anyway

   An unusual share of the repository is about prohibition. Twenty files sit in .out-of-scope/, each a written refusal — other runtimes in the core, a human-readable rendering for plan documents — with a 72 KB record arguing for enforcement by construction rather than by review. The managed hooks are where that lands: eleven registered hooks, including a 53 KB secret-read guard that blocks Read, Grep and Bash from touching secret files and replaced a deny rule the installer used to write, a prompt-injection guard, and guards that scan tool output for injected instructions and refuse edits to unread files. The bugs that reached users are the instructive ones. A cosmetic status line spawned git status on every render without opting out of git’s optional index lock, so a decoration could fail another commit; the fix sets GIT_OPTIONAL_LOCKS=0 at both seams that read. Command templates spliced typed arguments into instruction prose, so a flag was read as ordinary text and ignored; the same text went into a double-quoted shell snippet, so an argument containing a command substitution ran in the user’s shell. The community, meanwhile, reads the integration branch source: one contributor filed three bugs within seconds of one another — an include that resolves to prose on one runtime, three assumptions in a bootstrap document that fail on others, and a symlink that aborts a command.