Begin

begin: from fresh copy to working project

Run this once. When it's done, STATE.md carries real state and this skill never triggers again.

Which of the two entrances is this? Read it off the tree before doing anything, and say which one you read: a fresh copy holds nothing but Groundwork, while an existing project has its own .git history, source files, and a README that is not this one. Both take the route below; the existing project keeps everything it already has, and the steps say where the two differ. One skill, two doors: decision 0018 records why.

1. Clean the copy

Do this housekeeping silently: it is plumbing, not progress. Report on it only if something went wrong or needs the owner's action; a list of deleted template files is noise to them.

• If a MASTER_PROMPT.md or MASTER_PROMPT.local.md exists at the root: it is Groundwork's own

origin brief, not part of any project. Delete it. (Fresh copies no longer carry it: it is gitignored at source. This clears it from older copies made while it was still tracked.)

• CHANGELOG.md at the root is Groundwork's release history, not this project's. Note which

version this copy came from (its newest entry) in STATE.md at step 4, then empty the file down to its heading so this project's first release writes into it.

• Delete any non-archived spec folder: docs/specs/[0-9]* directories are in-flight Groundwork

maintainer work, never the new owner's. Keep the worked example in docs/specs/archive/007-pickup-slots/ and the TEMPLATE*.md files.

• docs/product/BRIEF.md and docs/specs/archive/000-baseline/ hold Groundwork's own scope and

the record of what the framework itself had already shipped. Counted against this project they would report someone else's work as done here. Put the blank brief back (cp docs/product/TEMPLATE-BRIEF.md docs/product/BRIEF.md) and delete the baseline folder.

• docs/PRODUCT.md is Groundwork's own record of the surfaces this framework renders, which the

design method would read as this project's product truth. Delete it: the method's own init writes this project's, in the same place, from what the interview below already captured. docs/DESIGN.md stays: sections 1 and 2 ship with Groundwork and are input to that method.

• docs/operations/deploy.md holds Groundwork's own release route, which would read as this

project's if it stayed. Put the blank runbook back (cp docs/operations/TEMPLATE-DEPLOY.md docs/operations/deploy.md). deliver fills it at first delivery.

• docs/state/DEBT.md carries Groundwork's own debt rows, which maintain would harvest and

report as this project's. Put the blank ledger back (cp docs/state/TEMPLATE-DEBT.md docs/state/DEBT.md). The first accepted defer: marker fills it; on the existing-project door below, the check total lands here as the starting position.

• docs/compliance/REGISTER.md holds Groundwork's own compliance answers, which would read as

this project's if they stayed. Put the blank register back (cp docs/compliance/TEMPLATE-REGISTER.md docs/compliance/REGISTER.md). Step 4 fills its header from the interview; comply fills the rest before first delivery. The two files beside it, docs/compliance/COMPLIANCE.md and docs/compliance/AI-LITERACY.md, are shared knowledge and stay as they are.

• Reset "denylist" in checks/config.json to []: its entries guard the origin repo's

retired wording, never this project's. Keep styleBans intact (those are generic).

• Keep the numbered records in docs/decisions/: they document the system this project just

inherited (why the rulebook, skills and checks work the way they do). This project's own decisions continue from the next free number. Naming a range here would go stale the next time the framework records one.

• Strip the data-derive attributes from the stat strip in index.html (leave the numbers and

the page alone). They tie those numbers to a gate that counts this repo, and in a copy the numbers describe the framework, not the project: the first decision this project records would otherwise turn the gate red.

• Verify prerequisites: git --version and node --version (Node ≥ 20). Missing → tell the

owner exactly what to install, then stop.

2. Interview the owner

Existing project: the code answers first. Its README, manifests, directory layout and recent commits already hold what the product is, who runs it and what it is built on. Read them, play the answers back for a one-line confirmation each, and spend the interview on what only the owner knows: why it exists, who it is for, what is deliberately out, and what "done" means from here. Asking an owner to describe software they have been running is the fastest way to lose their trust.

Material first. Ask whether the owner already has anything written: a PRD, project description, pitch, notes, or a rough idea dump. Take it now (pasted text or a file; convert non-Markdown files with the ingest skill) and treat it as the primary source: extract answers to the questions below from it, play each extracted answer back for a one-line confirmation, and interview only for the gaps. A vague idea is a valid starting point: capture what the owner does know, record the rest in BRIEF.md as explicitly open, and let scope sharpen it (step 6 routes there).

For whatever the material leaves open, interview one question per message, never a batch: announce upfront how many questions are coming ("question 1 of 5"), keep that counter on every question, and go as deep per topic as it takes to make the answer concrete before moving on. Capture answers verbatim where wording matters. Only what is genuinely theirs to decide. Everything else you decide later, and say so.

1. Product: name; one sentence: what it does, for whom, why now.
2. Users: who, in what situation, on which devices.
3. Scope: the 3-7 capabilities version one must have. Then explicitly: what is not in it.
4. Constraints: deadline, budget sensitivity, integrations, data residency, personal data

(yes/no: feeds the compliance register), any AI features (EU AI Act relevance).

1. Success: how the owner will know it worked.
2. Ownership: who signs off on scope changes (default: the owner you're talking to).

Discovery, proportional to project class. The answers above size the project: personal (the owner is the user), team (others inside one organization depend on it), or organization (a client, systems of record, a compliance regime, or real money involved). The class usually becomes clear at questions 2 and 4; extend the announced counter when it does. Class names which of BRIEF's discovery rows to ask about - the row comments there carry what each row holds, and every row the class does not ask gets n/a () per BRIEF's own header rule. Never interrogate a hobby project:

• Personal: ask nothing extra.
• Team: ask the current situation and cost of doing nothing, systems & integrations,

and rollout & adoption.

• Organization: ask every discovery row, concretely, one topic at a time.

Do not ask about stack (that's the stack skill, argued on merit) or design details (the design skill asks those when the time comes).

3. Challenge the idea

The interview captured the idea; this step tests it, before any template absorbs it. Load the critical-thinking skill and run its moves on the product idea itself - whether this is the right product to build is the highest-leverage decision of the whole lifecycle, and it gets pushback before it gets paperwork. The moves that always bite here: the named alternative (which for a whole product includes buying or configuring something that exists, and doing nothing), the load-bearing assumption, and the falsification question. The skill owns the method and the disagree-and-commit rule; this step fixes only the moment it runs.

Record the outcome in one or two lines: idea confirmed → a dated note in BRIEF.md's Product section; idea changed, or the owner overruled a real concern → a decision record in docs/decisions/.

4. Write it down

• Fill docs/product/BRIEF.md: numbered SC-items, explicit out-of-scope, constraints, and

the discovery rows at the interviewed depth. Write each SC-item in the owner's own words, as something they would recognize without translation: the progress overview quotes these lines back to them (node checks/progress.mjs).

• Existing project: write one baseline record in docs/specs/archive/000-baseline/spec.md, from

docs/specs/TEMPLATE.md, stating what already shipped before the brief existed and which SC-items it covers. Without it the overview quotes a running product back to its owner as nothing done, and the alternative is inventing retroactive specs nobody wrote. Everything from here forward gets a real spec at the size spec picks.

• Seed the glossary docs/product/CONTEXT.md with the domain terms the interview surfaced;

the file's template comment defines the entry format.

• Fill the handoff block in docs/state/STATE.md: status active, phase prepare, Now ▶ next step,

and the Groundwork version this copy started from (step 1). Without that line nothing can say which framework this project holds, which is what makes a later improvement findable.

• Fill the header of docs/compliance/REGISTER.md: set "Personal data processed", "AI features"

and "Applicable regimes", dated. Either first field a yes → comply runs before first delivery.

• Rewrite README.md for this product: the top half (what, for whom, how to run once it

exists) replaces Groundwork's pitch, and "Start a project" goes with it (that section is copy-the-template onboarding, done by now). Keep "How it works", "Requirements", "What lives where" and "Handing over" intact for successors; the License section is the owner's call (comply sets it). Existing project: its README already belongs to the product, so leave the top half alone and add "What lives where" and "Handing over" to it instead.

5. Set up the machinery

First, if a .git/ directory exists the copy still carries Groundwork's commit history (it was cloned, not made with "Use this template", degit, or a ZIP). The project must start from its own root, so remove it: rm -rf .git. In an existing project that .git is the owner's own history: keep it, skip git init below, and run the rest on the branch they are on.

If .claude/skills is not a symlink to ../.agents/skills (degit and some ZIP tools break it), restore it: ln -sfn ../.agents/skills .claude/skills. On Windows without symlink support: set "skipSymlinkCheck": "" in checks/config.json, the reason as the value, and point the tool at .agents/skills/ directly.

git init -b main
node checks/check.mjs --install-hooks   # wires core.hooksPath → checks/hooks (versioned)
node checks/check.mjs                   # must be green before the first commit
node checks/progress.mjs --register     # adds this project to the owner's cross-project overview
git add -A && git commit -m "chore: initialize project on Groundwork" \
  -m "Traces-to: explicit request: project initialization (begin)"

Existing project: run node checks/check.mjs before installing the hooks, and read its output as a measurement rather than a verdict. Real code turns it red on contact (length caps, typography in text the project already had), and adoption cannot mean cleaning a whole codebase before any work is possible. Take it in this order: fix what is unsafe to leave, secrets in tracked files first; give every file the project already had in docs/ its row in docs/README.md, which is a real fix and takes a minute; mark the rest at the site with the escape the check already offers and a reason (checks:allow-length: opening a comment, checks:allow-style), never by adding paths to an exclusion list in checks/config.json; and log the total in docs/state/DEBT.md. Then install the hooks and commit, with git init skipped. Report the numbers to the owner: what it flagged, what you fixed, what is now marked. If that count is large, it is the finding worth discussing before anything else gets built.

The design method, when this product has an interface. Groundwork does not make interfaces itself: making one runs on impeccable, installed per project, and design carries the Groundwork side of it. Whether this product has an interface is already visible in the interview (question 2 named the devices, question 3 the capabilities): play that read back as one yes/no question. It is asked here and nowhere else.

• Yes → node checks/design-method.mjs --install. Report the single line it prints, installed

version included, so the owner knows which release this project holds. The payload is gitignored like a dependency, so it does not enter the commit above; the enforcement line at the top of node checks/check.mjs is what says it is there.

• No → skip it, and record in STATE.md that this product has no interface and therefore no

design method. A later screen starts by installing it.

The install refuses before writing anything when it cannot finish: no npm, no network, or a Node below the floor impeccable itself states. Report its line as it comes, name in one line what is unavailable until it installs (design work would otherwise run on model defaults), put that in STATE.md under "Blocked on:", and carry on. Setup does not fail on it, and nothing else in this step depends on it.

If the owner has a remote (GitHub gets CI from .github/workflows/ci.yml; another host needs its equivalent: port it before first delivery), wire it and push. If not, note in STATE.md that CI is a deliver precondition still to be wired. On GitHub, also enable private vulnerability reporting on day one, so SECURITY.md's reporting channel exists before anyone reads the policy: gh api --method PUT 'repos/{owner}/{repo}/private-vulnerability-reporting' (not on GitHub or no gh → skip; deliver's first-release check covers it).

6. Hand off

Update STATE.md, then propose exactly one next step, normally:

• Scope needs sharpening → scope
• Scope is solid → stack, then architect, then design.

Report: what you recorded, what you decided yourself, and that single next step. ⚓

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Tradebaas
• Source: Tradebaas/Groundwork
• License: MIT
• Homepage: https://tradebaas.github.io/Groundwork/

Install and usage instructions live in the source repository linked above.