Claude Code Development Kit
Peter Krueck’s opinionated starter setup for Claude Code, MIT, 1.4k★ / 154 forks (2026-08-06). Its pitch is the honest one: “the setup you’d build yourself after a few weeks of using Claude Code, packaged so you start with it on day one.” Aimed at individual developers and small teams on a Claude Code subscription, and explicitly not at large-scale API multi-agent work. v3 is a rewrite; v2 stays available.
The stated problem is context-rot by another name — the README calls it context decay: docs go stale across sessions, the window fills, architectural decisions get re-derived because nothing keeps documentation in step with the code.
What it ships
Four context files under docs/ai-context/ — spec.md (product definition),
project-structure.md (layout and stack), progress.md (roadmap and tasks),
deployment-infrastructure.md (hosting, secrets, CI/CD) — kept in step by an /update-docs skill that
filters trivial diffs to protect the token budget.
Tiered priming. /prime loads a light default at ~4–6k tokens, with --full and --deploy for
more. This is the part that takes context-rot seriously at the architecture level rather than the
advice level: the always-loaded floor is small, and depth is opt-in per task.
Quality gates. /review-work fans out parallel subagents (agent-orchestration) — Bug Hunter,
Rules Auditor, an optional Architect — over uncommitted changes. /second-opinion escalates a stuck
decision to an external model (OpenAI Codex by default, Gemini CLI supported), which requires those CLIs
separately. /merge checks docs currency and repo cleanliness before shipping; /verify runs the app
or the test suite. External API claims surfaced in review are checked against Context7 rather than
trusted from training data.
Hooks. A security scanner blocking secrets from reaching MCP servers, a deny list for dangerous
commands (git push --force, sudo), a review-on-stop hook that nudges across three stops before a
final commit, plus optional file-tracking and session cleanup.
The corpus already measured its premise, and the result was negative
This is the useful thing about the kit arriving here rather than anywhere else. Its foundation is four
project-context files kept synchronized, and context-files-ablation-paper is a controlled ablation
of exactly that intervention — AGENTS.md injection across Claude Code and Codex, 17 PR-mined tasks,
288 gold-test runs — which found no correctness effect, with near-misses failing on implementation
skill rather than missing knowledge. What it did find were process effects: fewer cache tokens, fewer
blind full-suite test runs.
The kit reports no evaluation of any kind. So the fair reading is not that it is refuted — the ablation tested one file, injected, against this kit’s four files plus tiered priming plus doc sync, which is a different intervention — but that the one measured result the spoke owns on this exact mechanism says the effect lands on process, not correctness, and the kit’s own claims are all process claims anyway: less re-derivation, less token burn, cleaner merges. Read that way the kit and the paper agree, and the part neither supports is the implicit promise that better context makes the agent better at the work.
Two tensions worth keeping
Against claude-md. That page’s stated discipline, from HumanLayer, is that the always-loaded
instruction file should be hand-authored, never auto-generated, because /init-style generation
produces bloated low-signal files. The kit’s /update-docs is machine maintenance of project context by
design. Both claims stay on the record. The kit’s partial answer is that its generated material is not
always-loaded — it sits behind /prime tiers — which relocates the objection rather than meeting it:
low-signal text costs less when it is opt-in, and it is still what the agent reads when primed.
Against agent-guardrails and constraint-evading-behavior. The review-on-stop hook is advisory and fires across three stops before a final commit. A guardrail the operator can decline three times is a prompt, not a boundary, and this spoke has already recorded both that an approval means only that a human accepted the displayed action, and that agents satisfy a check’s letter while defeating its purpose. The deny list and the secret scanner are the real boundaries here; the review nudge is etiquette.
Where it sits
Against the harnesses this spoke holds, the kit is not a harness — it adds no loop and no runtime. It is a configuration layer: docs, slash commands, skills, hooks and subagent recipes dropped into someone else’s harness. That makes it the packaged form of governed-context-claude-code, which was one practitioner’s unbenchmarked playbook for the same problem, and it converts that advice into defaults a beginner gets on day one. The skills are Claude Code-specific; the doc templates are not, and the README suggests they carry to Cursor, Windsurf or Codex.
Related
context-rot · governed-context-claude-code · claude-md · context-files-ablation-paper · agent-guardrails · constraint-evading-behavior · agent-orchestration · agentic-coding-harness · spec-driven-development · synthesis