One AI Development Workflow, Any Coding Assistant
Teams adopting AI coding tools keep hitting the same problem: the project’s knowledge
ends up inside one vendor’s configuration. Rules live in .cursorrules, or a Claude
hook, or a Copilot instruction file. Switch assistants, or just add a second one, and
the project forgets everything it knew.
Stackkeel takes the opposite position. The development process belongs to the repository, not to any tool reading it.
The three layers
Section titled “The three layers”1. One instruction file: AGENTS.md. The AGENTS.md standard
is read natively by GitHub Copilot, Cursor, Codex CLI, Gemini CLI, Zed, and dozens of
other tools. Stackkeel’s root AGENTS.md carries everything: the architecture, the
invariants, the six-phase pipeline, commit grammar, and workflow rules. CLAUDE.md
is a one-line pointer to it. Per-app shims add only local details like ports and
commands.
2. One skill format: Agent Skills. Repeatable procedures (personalize the kit,
start a feature, prepare a release, process feedback safely) are
Agent Skills: SKILL.md files with YAML frontmatter,
readable by roughly forty tools including Claude Code, Codex, Copilot, Cursor, and
Gemini CLI. They live in .claude/skills/, the most widely scanned location, with a
symlink at .agents/skills for tools that only read the neutral path. A skill is a
runnable checklist; an assistant without native skill support can simply follow it.
3. Enforcement outside every assistant. This is the layer most setups skip. Instructions are suggestions; gates are guarantees. Stackkeel’s binding rules live in git hooks and CI:
| Rule | Enforced by |
|---|---|
| Commit message grammar with defect-tracking trailers | commit-msg hook + CI re-validation |
| A work log exists before code changes | pre-commit hook + CI |
| Typecheck, lint, tests, and tripwires pass before push | pre-push hook running pnpm kit:verify |
| No secrets, stale instructions, missing audit calls, or stray style tags | tripwire suite in CI |
| An unpersonalized template copy cannot ship | kit:check --strict in CI |
Claude Code users additionally get session hooks for fast in-editor feedback. Their absence in another tool changes nothing about what can merge.
What this looks like day to day
Section titled “What this looks like day to day”Open the repository in Claude Code and the personalization check, feedback counts,
and review cadences surface automatically at session start. Open the same repository
in Cursor or Codex and AGENTS.md tells the assistant to run pnpm kit:status,
which prints the same information. Either way, the same six-phase workflow applies,
the same work-log discipline is enforced at commit time, and the same CI verdict
decides what lands.
Model independence is also insurance. Assistants are improving quickly and pricing shifts often. A repository that encodes its own process can follow the best tool at any moment without a migration project.
Try it
Section titled “Try it”The getting started guide works identically regardless of which assistant you point at the repository. If you use several, point them all at it; they will read the same playbook.