Skip to content

IMPLEMENT_PROMPT.md

Latest commit

 

History

History
103 lines (90 loc) · 5.85 KB

IMPLEMENT_PROMPT.md

File metadata and controls

103 lines (90 loc) · 5.85 KB
Raw

The Launch Prompt — Deploy claude-sdlc-kit Into a Project

This is the one prompt for installing this kit — there is no other. Paste it into a fresh Claude Code session opened at the root of the target project (any project living inside this folder/workspace, including a brand-new one). It requires Claude to read the kit's actual source files and reproduce a faithful, project-specific instance of them — never to reinvent the system from memory or a summary. Definition of done: real files with zero placeholders left in them, not a description of what could be done.

The Prompt

You are installing and activating claude-sdlc-kit in this project. The kit's source of truth lives at claude-sdlc-kit/ (locate it — sibling folder, path given by the user, or this repo's own copy). Do not improvise the system from memory or from the summary below — read the actual files. Work through these steps in order and do not stop until every deliverable in "Definition of done" exists.

1. Read the kit before touching anything

Read, in full:

  • claude-sdlc-kit/README.md and claude-sdlc-kit/CLAUDE.md (the template)
  • Every file in claude-sdlc-kit/.claude/agents/
  • Every file in claude-sdlc-kit/.claude/commands/
  • Every SKILL.md in claude-sdlc-kit/.claude/skills/*/
  • Every file in claude-sdlc-kit/docs/ (ARCHITECTURE.md, ONBOARDING.md, DECISIONS/0000-template.md, DECISIONS/0001-adopt-claude-sdlc-kit.md, CHECKPOINTS/README.md, HANDOVER/README.md, changelog/README.md)
  • claude-sdlc-kit/tooling/ (commitlint config, husky hook)

Build a complete picture of every agent's role/tools, every command's trigger, and every skill's methodology before step 2. If anything in this prompt conflicts with what the source files actually say, the source files win.

2. Audit the target project

Before copying anything, learn the real project: language(s), framework(s), package manager, actual working install/dev/test/lint/typecheck/build commands (run them — don't guess), directory layout, existing conventions, existing git history and branch model. If this project already has multiple independently-built apps in one repo (as this boilerplate does under src/pages/<app-name>/), decide — and record as an ADR in step 5 — whether the kit governs the whole repo or is scoped per app, before installing anything.

3. Install, adapted to what you found

Copy into the target project root, adapting content (not just paths) to what step 2 found:

  • .claude/agents/*.md — all thirteen agents, tool-access lists unchanged unless the target project genuinely lacks a capability they assume (say why if you trim one).
  • .claude/commands/*.md — all slash commands.
  • .claude/skills/*/SKILL.md — all skills.
  • tooling/ — wire up commitlint + husky if this is a Node project; otherwise find or note the equivalent enforcement point for the target stack (pre-commit, lefthook, CI step) and say so explicitly in CLAUDE.md's Known Gotchas — don't silently drop it.
  • docs/ARCHITECTURE.md, docs/ONBOARDING.md, docs/DECISIONS/0000-template.md, docs/CHECKPOINTS/README.md, docs/HANDOVER/README.md, docs/changelog/README.md, CHANGELOG.md seed.
  • CLAUDE.md — copy the template, then fill every <placeholder> with real values from step 2. A placeholder left in the delivered file is a failed delivery, not a TODO.

4. Produce the first real deliverables

Installing inert scaffolding is not the job — prove the system is live:

  • Run the code-audit methodology (7 phases, as documented in .claude/skills/code-audit/SKILL.md) and write docs/AUDIT.md. Every finding cites file:line or command output — no evidence, no finding.
  • Write docs/DECISIONS/0001-adopt-claude-sdlc-kit.md, following the template, recording the real date, real deciders (ask if unknown), and the repo-vs-per-app scoping decision from step 2 if relevant.
  • Write an initial docs/CHECKPOINTS/<YYYY-MM-DD-HHMM>.md establishing state zero, per .claude/skills/checkpoint/SKILL.md.

5. Verify before reporting done

  • .claude/commands/ has exactly the commands the kit ships (/audit /review /security /test /commit /pr /version /tag /changelog /release /checkpoint /resume /handover /onboard) and each resolves to a real command file.
  • Every command that names an agent has that agent's file present with matching responsibilities.
  • CLAUDE.md has zero remaining <...> placeholder tokens — grep for < bracket patterns to check.
  • Every command listed in CLAUDE.md's Commands table actually runs (you ran it in step 2 or just now).

6. Report

End with a short table: what was installed verbatim, what was adapted and why, and what's left for a human (secrets, CI platform wiring, remote git permissions — anything you can't or shouldn't do yourself).

Constraints

Never invent audit findings — every claim cites real evidence. Prefer small, reviewable diffs; if the target project already has content at a path the kit wants to write to (an existing CLAUDE.md, CHANGELOG.md, etc.), merge, don't clobber — ask if unsure. Ask before any destructive git operation. Never commit secrets; flag any you find. If context runs low mid-installation, write a checkpoint (per .claude/skills/checkpoint/SKILL.md) before compacting so the install can resume from where it stopped.

Definition of done

  • .claude/agents/, .claude/commands/, .claude/skills/ present and match the kit's source files (adapted, not reinvented)
  • tooling/ wired for the target stack, or its absence explicitly justified in CLAUDE.md
  • CLAUDE.md complete, zero placeholders, commands in it verified to actually run
  • docs/AUDIT.md exists with evidence-backed findings
  • docs/DECISIONS/0001-adopt-claude-sdlc-kit.md exists with real content
  • docs/CHECKPOINTS/ has one initial entry
  • Report delivered per step 6