scaffold

project-kit: Scaffold a New Project

Lay down the standard project setup, then help fill the CLAUDE.md skeleton.

The deterministic work is done by scripts/scaffold.sh (in this plugin's dir). This skill gathers the inputs, runs it, and then guides the human-judgment part (Domain Context, Architecture).

Workflow

1. Gather inputs (ask only what's not obvious):

   • Target directory (absolute path) and project name (defaults to the dir basename).
   • Profile: data (default — data-eng archetype with data/{bronze,silver,gold}), python (plain uv lib), or minimal (beads + docs only).
   • Whether to include .mcp.json (playwright/context7) — keep it for scraping/web projects, drop with --no-mcp otherwise.
   • CI is on by default (closes the portfolio's CI gap); drop with --no-github only if there's a reason.

2. Dry-run first. Run the script with --dry-run and show the planned actions:

   "$CLAUDE_PLUGIN_ROOT/scripts/scaffold.sh" <target> --name <name> --profile <p> --dry-run
   

   (If $CLAUDE_PLUGIN_ROOT is unset, use ${PROJECT_KIT_HOME:-$HOME/Code/project-kit}/scripts/scaffold.sh.)

3. Confirm, then run for real (drop --dry-run). The script is non-destructive: it skips existing files unless --force/--backup is passed, and won't re-run bd init if .beads/ exists.

4. Fill CLAUDE.md with the user. The template ships with <!-- fill --> prompts. Help the user complete:

   • Domain / Scientific Context — the real problem, the outcome/target, and especially the data provenance (own data? regional/Peruvian domain? public source?). This is the differentiator.
   • Architecture — stages and data flow.
   • Key Files — the table. Remove the Data Conventions section if the project isn't data-oriented.

5. Report what was created/skipped and the next steps the script printed.

6. After the first push (once the repo exists on GitHub), enforce the PR flow the scaffold assumes — a protected main with no unsatisfiable approval gate:

   "$CLAUDE_PLUGIN_ROOT/scripts/protect-branch.sh" <owner/repo>
   

   This requires a PR and gates merges on CI + resolved review conversations (no approval count — see SETUP-CHECKLIST.md §1 for the solo-dev rationale). Then enable Copilot/Sourcery auto-review in repo settings so those comments become the binding review.

Notes

• beads is per-project here: the script runs a fresh bd init (it auto-detects the issue prefix).
• After scaffolding, bd ready should work inside the new project for task tracking.
• The script needs bd and uv on PATH for those steps; it prints a hint and continues if either is missing.