| name | setup-stack |
| description | Use this skill to run the entire stack-it pipeline end to end in one guided pass. It takes a project from nothing to a working, verified, documented tech stack by running each stage skill in order (identify-stack-slots, decide-stack, install-stack, scaffold-and-verify, document-stack), resumes from wherever the project already is, optionally sets up git and commits a checkpoint after each stage, and keeps a running task list so the user can see how much is left. Use it whenever someone wants their whole stack set up rather than just one step. Phrasings like "set up my stack", "set up my whole project stack", "help me start a new project from scratch", "pick, install, and verify my stack", "take me from zero to a working project", "bootstrap my project", or "run the whole stack-it pipeline" should trigger it. It pauses only where it genuinely needs the user (their tool picks in the decide stage, an install confirmation, a stack-fault re-pick, an OK before writing docs) and handles the loop where a broken tool sends verification back to re-decide. Reach for the individual stage skills instead when the user explicitly wants only one step (only choose tools, only install, only document); this orchestrator is for the full journey. It does not reimplement the stages; it sequences them. |
Set Up Stack
The orchestrator for the stack-it pipeline. It takes a project from nothing to a working, verified, documented tech stack by running the five stage skills in order, resuming from wherever the project already is and pausing only where a human is genuinely needed. Each stage is a real skill with its own SKILL.md; this skill's job is to sequence them, carry the .claude/stack-it/ artifacts between them, and handle the one loop where verification sends a bad choice back to be re-decided. Defer to each stage for how that stage works; don't reimplement it here.
The pipeline
| # | Stage skill | Produces / does |
|---|
| 1 | identify-stack-slots | .claude/stack-it/slots.yaml: the decision categories |
| 2 | decide-stack | .claude/stack-it/stack.yaml: concrete, pinned, vetted choices |
| 3 | install-stack | installs the locked stack; updates stack.yaml to match reality |
| 4 | scaffold-and-verify | builds a minimal slice and runs it green; updates stack.yaml on pin drift; escalates a stack fault back to stage 2 |
| 5 | document-stack | docs/stack.md and docs/stack-notes.md + the one-line AGENTS.md pointer + the README block documenting the stack |
Run each stage by following that stage's skill, not by re-deriving its behavior.
Step 0: Set up the workspace
Before the first stage, get two things in place that you maintain for the rest of the run.
Git. Check whether the project is a git repository (git rev-parse --is-inside-work-tree). If it isn't, offer to initialize one: git init plus a baseline .gitignore appropriate to the project (keep the .claude/stack-it/ artifacts; ignore build output, node_modules/.venv, and any .env* secrets). Ask once, batched with anything else you need up front; if the user declines, skip the commit steps below and just narrate progress. Working in git is strongly preferred: it lets you checkpoint after every stage (a clean rollback point that matters when scaffold-and-verify hits a stack fault and loops back), and the history is itself a resume signal. If the project is already a repo, note the current branch and whether the tree is clean before changing anything.
The running task list. Build the master list of everything the run will do and keep it current: it's how the user stays oriented through a long, sometimes-interactive process. Seed it with the five stages, then expand each into real sub-tasks as the artifacts appear: once slots.yaml exists, each open slot is a "decide" task; once stack.yaml exists, each entry is an "install" task and a "verify" task. Add and remove items as the stack changes (a deferred slot drops off; a stack-fault re-pick adds a task back). See The running task list below for how and when to show it.
The running task list
Maintain one list for the whole run and render it to the user between every stage, with clear done / current / remaining markers, so they always know how much is left and never have to wonder "are we done yet?". Keep it concise: group by stage and show counts where a stage has many items. Give each stage a short one-line purpose so the user understands what it does and why it's there, especially the first time you show the list, when the pipeline is still unfamiliar. Update it whenever the work changes (slots added or deferred, tools chosen, a stack fault sending a slot back). Derive it from the slots.yaml/stack.yaml entries plus the fixed stage sequence; don't invent a separate tracking file.
Example shape (adapt freely):
stack-it progress
[x] 1. identify slots: agent researches stack shape for your approval (15 slots: 13 required, 2 optional)
[>] 2. decide stack: agent swarm researches tooling choices, you pick (4 / 15 chosen)
[ ] 3. install: agent swarm pulls up-to-date setup steps, asks only when needed
[ ] 4. scaffold & verify: build a small example and verify the tooling works / fix
[ ] 5. document: document the finalized stack to docs/stack.md, docs/stack-notes.md, AGENTS.md and README
The handoff contract
Whenever you stop and return control to the user, end with one explicit line so it's never ambiguous whether you're waiting on them or about to continue. Use one of two shapes:
- Ready to proceed, no input needed:
Next: <stage or action>. Reply "go" (or tell me to change something) when you're ready.
- You need something first:
Before <next step>, I need <the specific thing>: <question>.
State the next step, not a menu of everything they could do. This applies at every pause the pipeline makes: the decide-stage picks, the install confirmation, a stack-fault escalation, and the OK before docs.
Step 1: Find the starting point (resume-aware)
Don't redo finished work. Before running anything, figure out where the project already is by inspecting:
.claude/stack-it/slots.yaml: have the slots been identified? (validate with ${CLAUDE_PLUGIN_ROOT}/scripts/validate_yaml.py --stage slots if present).claude/stack-it/stack.yaml: have the tools been decided and locked? (validate with ${CLAUDE_PLUGIN_ROOT}/scripts/validate_yaml.py --stage stack)- the project itself: are the locked tools actually installed (manifests and lockfiles present, dependencies resolved)? is there a verification slice, and does it pass? do both
docs/stack.md and docs/stack-notes.md already exist, each with its own <!-- stack-it:stack start/end --> pair and real content between the markers, along with the AGENTS.md pointer to docs/stack.md and the README block? A section in either file whose first line is Moved from CLAUDE.md by migrate; run document-stack to regenerate. or Placeholder written by migrate; run document-stack to fill this in. is a placeholder migrate left for document-stack, so the documenting stage hasn't run; one real section next to one placeholder is not documented either. An old-format project instead has a <!-- stack-it:stack start --> block in CLAUDE.md. That is documented on the old layout, which is not finished work: read CLAUDE.md from disk to spot it, since Claude Code strips HTML comments from loaded context. A block sitting there alongside an otherwise complete format-2 layout is a different animal: a leftover from a migration that did not finish, and not finished work either.
Map what you find to the first unfinished stage:
- no
slots.yaml → start at identify-stack-slots
slots.yaml but no stack.yaml → start at decide-stackstack.yaml but the tools aren't installed → start at install-stack- installed but no passing slice → start at scaffold-and-verify
- verified but undocumented → start at document-stack
- verified and documented, but on the old layout (a
<!-- stack-it:stack start --> block in CLAUDE.md, no docs/stack.md) → start at document-stack, which calls migrate to restructure before it refreshes
- all done, meaning
docs/stack.md and docs/stack-notes.md each with their own <!-- stack-it:stack start/end --> pair, neither generated section opening with a migrate placeholder or stub line, the plain-path pointer sentence to docs/stack-notes.md present inside docs/stack.md's markers, no <!-- stack-it:stack start --> block left in CLAUDE.md, the AGENTS.md pointer, the README block, and format: 2 in stack.yaml → say so; offer to re-run a specific stage. A missing stamp is not done, and neither is a missing notes file, a missing pointer sentence, a placeholder section in either file, or a leftover block: all of them route to document-stack, which calls migrate to finish the layout and then writes the content
When the state is ambiguous (a stack.yaml exists and some deps look installed, but you can't tell it's complete), tell the user what you found and confirm the entry stage rather than guessing. A brand-new project with none of these artifacts starts at the top.
Step 2: Run the stages in order
From the entry stage, run each subsequent stage in pipeline order. Between stages, every time: confirm the expected artifact is in place and valid before starting the next stage; checkpoint the work if git is enabled: commit the stage's artifacts and any files it produced with a conventional message (chore(stack-it): identify slots, feat(stack-it): decide stack, feat(stack-it): install stack, …) so each stage is a discrete, revertible point; and show the user the running task list plus a one-paragraph summary of what the stage produced and what's next.
Pause for the user only where a stage genuinely needs them; don't gate every step:
- decide-stack: their hard preferences and per-slot picks. Let the stage converse; that's its job.
- install-stack: confirmation before destructive, irreversible, or surprising commands. The stage handles this; don't add friction to routine installs.
- scaffold-and-verify: a stack-fault escalation (see Step 3).
- before document-stack: a quick "the stack's verified, write the docs now?" check.
Everywhere else, flow through: a successful, unambiguous stage rolls straight into the next with just a summary. The point of "guided" is that the user is informed and consulted at the real decision points, not interrupted at every one. Whenever you do pause, close with the handoff-contract line (see The handoff contract) so it's unambiguous whether you're waiting on them.
Step 3: Handle the verify → decide loop
scaffold-and-verify sorts each failure into a slice fault (it fixes), pin drift (it fixes and records to stack.yaml), or a stack fault: a chosen tool that fundamentally can't work. On a stack fault it stops and escalates rather than papering over. When that happens, the loop is:
- Surface the fault to the user with the stage's diagnosis (what failed, why it's the tool and not the slice).
- Re-enter decide-stack for just the affected slot(s), with the new information, to re-pick. It edits the same
stack.yaml.
- Re-run install-stack for the changed slot(s), then scaffold-and-verify again.
- Repeat until verification is green.
This loop is the whole reason the stages are separate skills: don't try to force a stack fault through inside verification. Keep the user in the loop on each re-pick; they own the stack decisions. Surface the fault and each re-pick with the handoff contract, update the task list (the affected slot goes back to "decide"), and checkpoint (commit) after each successful re-install-and-verify so every loop iteration is a clean rollback point.
Step 4: Finish
Once verification is green and document-stack has written docs/stack.md and docs/stack-notes.md, the AGENTS.md pointer, and the README block, give a final summary: the locked stack (tools and versions), what was installed, that the slice passed and which gates, where the docs live, and any caveats or pin drift recorded along the way. Point the user at the obvious next step (run the dev server, the test suite, the build) and flag anything still on them (accounts, secrets, env files). Make a final checkpoint commit for the docs if git is enabled, show the task list with everything done, and close with the handoff contract.
Scope and boundaries
This skill sequences the five stage skills; it does not reimplement their logic, and it respects each stage's own boundaries (it won't enter credentials, and it routes stack decisions back through decide-stack and the user rather than deciding alone). The pipeline is the default path, not a mandate: if the user doesn't want docs, skip document-stack; if they hand you a partial state, resume mid-pipeline per Step 1. When the user explicitly wants only one step, point them at that stage's skill instead of running the whole journey.
Bundled resources
${CLAUDE_PLUGIN_ROOT}/scripts/validate_yaml.py: Validate .claude/stack-it/ artifacts while detecting the resume point (--stage slots for slots.yaml, --stage stack for stack.yaml). Run python ${CLAUDE_PLUGIN_ROOT}/scripts/validate_yaml.py --help for usage.
