New

---
name: new
description: Bootstrap a new project with the Spec-Driven Development Framework (Part 8)
disable-model-invocation: true
---

You are bootstrapping a new framework instance. This command is a thin binding:
the protocol lives in the edition document bundled with this plugin, you follow
it as written, and on any conflict between this file and the edition, the
edition wins. Never fork the protocol.

## 1. Guards and grounding

- Output contract for your FIRST message: it opens with the model expectation,
  before any question or observation. Session zero is architecture work and
  should run on the escalation tier of the model ladder (Part 2; `/model fable`
  under the current bindings, or `/model best` where the organization has no
  Fable access). It is not the day-to-day `opusplan` default. Say so; do not
  assume the user read the README, and do not let environment friction displace
  the line (a cold run has skipped it exactly that way). Continue either way.
- Greenfield guard: if the current directory is not empty and not a git
  repository, stop and ask before touching anything. An existing codebase
  belongs to /setlist:retrofit instead.
- Observe mode (`/setlist:retrofit --observe`, the verdict delta of
  `/setlist:upgrade`) does not apply here: a new project has no history to
  read, and this interview's proposal of every decision before a file is
  written is its equivalent.
- Load the protocol you are bound to. Run:
  `bash "${CLAUDE_PLUGIN_ROOT}/scripts/part.sh" 8`
  and follow that text (the Bootstrap Protocol) for everything below.

## 2. Gather conversationally (before the interview)

Ask for, in plain conversation:

- The project paragraph: what it is, the core user job, the one-sentence pitch.
- Constraints: budget, solo vs team, time, legal/privacy, anything mandated.
- The working mode: the user writes code too, or reviews only.

## 3. The Part 8 interview and decisions, exactly as written

- Run Part 8 Step 1 as one structured round: only genuine forks become
  questions, recommendations stated inline, and the stopping rule holds (one
  round is right, two is fine, three is procrastination). The round OPENS with
  "what has changed since this framework edition was written" (tooling,
  models, workflow), Part 8's mandated first question, never its last. Include
  the environment facts Step 1 demands: the opusplan check below, the QA
  tooling for this project type, the design surface question for UI projects,
  and the riskiest assumption worth a spike.
- Verify `opusplan` by LIVE PROBE before parking, the same probe the prompt
  path runs: attempt the selection once from this environment (a one-shot
  `claude -p --model opusplan` invocation that returns any clean reply
  verifies it resolves; in an interactive session, `/model opusplan` and read
  the response). Only if the probe cannot run or fails: leave the model line
  out of settings.json and park verification in STATUS.md with a named owner.
  Never park without attempting the probe.
- Verify `jq` in the same environment check (`command -v jq`). The stamp
  refuses before it writes anything when `jq` is missing or not working (it
  builds `.claude/sdd.json` with `jq`, and the three stamped GIT hooks fail
  closed without it; the session hooks report, or stay silent, and do not block), so
  check first rather than meet the refusal; a clean container is the common
  case, since jq is not in most default installs. Report the install command
  and let the user run it; install nothing yourself.
- Ask the `diagram_command` question in the same round, DEFAULT NONE. "Is there
  a command that prints this project's structure as one Mermaid block (an import
  graph, a schema dump)?" A yes records it in `.claude/sdd.json` beside
  `gate_command`, and its committed output under `docs/diagrams/generated/` is
  then required to equal what it prints at every close. A no records nothing,
  which is the right default: the framework ships no extractor, and a declared
  command that does not run refuses the close rather than degrading.
- Run Step 2 foundational decisions WITH the user at full depth. This is the
  product of the whole session; do not compress it. The core data model gets
  the most effort.

## 4. Phase 1: the mechanical stamp

Write the answers file to `.claude/stamp-answers.txt` in the project directory,
KEY=VALUE lines:

```
project_name=<name>
stack=<short stack id>
working_mode=<review-only | developer-writes>
ui=<yes|no>
opusplan_verified=<yes|no>
design_surface=<yes|no>
src_role=<src, or the agreed path>
tests_role=<tests, or the agreed path>
mode=new
```

Then run:
`bash "${CLAUDE_PLUGIN_ROOT}/scripts/stamp.sh" .claude/stamp-answers.txt .`

The stamp emits every framework-fixed file, including the four session hooks
(enabled; each documents its one-line disable in its header), the git hooks,
the trunk audit and the forge check beside it, `.github/workflows/
setlist-forge-check.yml` (the check as a required status check, `setlist forge
check`) and `.github/CODEOWNERS` (the four protected paths under an `@OWNER`
slot that phase 2 fills with the team that owns the enforcement layer; a
standing slot is a validate finding), the `gates` block in `sdd.json` (three
empty tiers `/scaffold` records beside `gate_command`), `specs/TEMPLATE.md`
extracted from the edition, the committed edition copy, and `AGENTS.md` beside
`CLAUDE.md` (a pointer to it and to the edition for agents that read
`AGENTS.md`, never a second copy of the golden rules). Do not regenerate
any stamped file by hand. Requiring the check on the trunk is the user's act at
their forge; name it in the hand-off rather than assuming it.

## 5. Phase 2: tailored generation (the 20 percent that is the product)

Write only what encodes the decisions just made with the user: the steering
docs' content, the founding ADRs (entry plus index row, typically five), the
first specs per Part 8 Step 4 (spike 0000 if warranted, 0001 to 0003, later
ones as DRAFT or parking-lot rows), `RUNBOOK.md` as Part 7 instantiated with
this stack's exact commands, and every `[PHASE 2 SLOT: ...]` marker the stamp
left behind. No slot marker survives this phase.

**Seed `docs/diagrams/` here, in phase 2, and never in the stamp.** Create the
directory and write two files from the decisions just made: `context.md` at L1
(the system, its users and the outside services it talks to) and the L2
container block at the bottom of `steering/structure.md` beside the core model.
Both under the `diagrams` skill's rules, each opening with the four-line header
from `${CLAUDE_PLUGIN_ROOT}/templates/root/DIAGRAM-HEADER.md` and
`Synced by: bootstrap`. Node names are the role paths the answers file recorded,
so they resolve against the tree from the first close.

Say plainly what creating the directory does, because it is a decision and not a
formality: its presence ARMS the close's diagram checks, so from the first spec
onward a close that says `updated` names its files and one that says
`no impact` while a diagram moved is refused. That is the intended state for a
project born under this edition. A user who wants to start without it deletes
the directory, and every reader behaves as it did before the diagram half
existed.

## 6. Hand off

Per Part 8 Step 5: point the user at RUNBOOK.md, name the next action in
specs/STATUS.md (normally `/scaffold`, then the spike or spec 0001), and note
that day-to-day Git runs through /setlist:checkpoint.

Warn explicitly in the handoff text: nothing is committed yet. Session zero
leaves the instance as uncommitted files by design (the first commit belongs
to /scaffold, which also runs git init when no repo exists), so until
/scaffold runs, one stray deletion loses the whole session. Run /scaffold
next, before anything else.

## Gotchas (field-observed)

- Do not improvise vocabularies the templates already fix. A generated
  DECISIONS.md index once carried the status "Accepted" where Part 4's legend
  is ACTIVE / SUPERSEDED-by-NNN / INFERRED; the template header states the
  legend, use exactly those words.
- Non-interactive bootstraps (`claude -p`) under default permissions cannot
  read the plugin root (`part.sh` lives outside the project directory), and
  the run stalls on an approval that never comes. The permission stance is
  the operator's, set outside the plugin (Part 7c): say what stalled and
  stop, never choose or suggest a mode that turns prompts off.