SDLC Studio

Version 2.0.0 | MIT Licence

A full software-development-lifecycle skill for AI coding agents - one folder of instructions, templates, and tested scripts that takes a project from requirements to verified code:

PRD → TRD → Personas → Epics → User Stories → Plan → Code → Tests → Verify

SDLC Studio is a standard Agent Skill (SKILL.md format), so the same install works in Claude Code, OpenAI Codex, Gemini CLI, opencode, and GitHub Copilot.

The idea (for beginners)

Most AI coding jumps straight from a vague prompt to code, then drifts as the project grows. SDLC Studio adds the steps a real team uses, so the agent always has the context it needs to stay on track. Each step writes a plain markdown file under sdlc-studio/ in your project. You stay in control: review each artefact, then run the next command.

Start a brand-new project with prd create, or point it at existing code with prd generate. If you ever lose your place, /sdlc-studio status shows where you are and /sdlc-studio hint tells you the single next thing to do.

Install

One line, any platform:

# macOS / Linux - Claude Code, globally (the classic)
curl -fsSL https://raw.githubusercontent.com/DarrenBenson/sdlc-studio/main/install.sh | bash

# Every tool you have installed
curl -fsSL https://raw.githubusercontent.com/DarrenBenson/sdlc-studio/main/install.sh | bash -s -- --target auto

# Windows
irm https://raw.githubusercontent.com/DarrenBenson/sdlc-studio/main/install.ps1 | iex

Where it installs

--target claude,codex picks tools; --local installs into the current project; --list-targets shows the map and what is detected. The generic .agents/skills directory is read by four tools, so --target agents covers Codex, Gemini, Copilot, and Cursor in one copy (Claude Code does not read it). Native alternatives also work: gh skills install DarrenBenson/sdlc-studio sdlc-studio (Copilot) or gemini skills install https://github.com/DarrenBenson/sdlc-studio.

Stale-copy sweep: after installing, the installer refreshes every other sdlc-studio copy it finds in the known locations (reported as old -> new; it never touches a directory that is not this skill), so no stale version lingers when you use several tools. Opt out with --no-sweep / -NoSweep. Full detail: docs/INSTALL.md.

Verify

Start your tool in any project and run /sdlc-studio status (Claude Code) or confirm the skill is listed (/skills in Codex/Gemini). You should see the pipeline dashboard.

Quick start

/sdlc-studio status            # Where am I? (four-pillar dashboard)
/sdlc-studio hint              # The single next thing to do
/sdlc-studio prd create        # New project: interview -> PRD
/sdlc-studio prd generate      # Existing code: extract the PRD from it
/sdlc-studio epic              # Break the PRD into Epics
/sdlc-studio story             # Break Epics into User Stories with AC
/sdlc-studio code plan         # Plan the next story
/sdlc-studio code implement    # Build it
/sdlc-studio code verify       # Verify against acceptance criteria

Run /sdlc-studio help for the full catalogue (bugs, change requests, RFCs, personas, test specs, test automation, GitHub sync, and every flag).

What you get

Two modes for every document. create interviews you (greenfield); generate reverse-engineers a migration blueprint from existing code (brownfield) - detailed enough to rebuild the system, validated by running tests against the real implementation.

Status that polices itself. Artifacts carry canonical statuses; reconcile detects and fixes drift from a file census; acceptance criteria can carry executable Verify: lines that reconcile --verify actually runs. A blind-review gate checks that a plan's tasks logically satisfy every AC before implementation.

Determinism in scripts, judgement in the model. Eight stdlib-only Python helpers (census, status, validation, ID allocation, repo indexing, AC verification, GitHub sync, plan/lessons management) with 180 unit tests do the mechanical work; the model does the thinking.

Agentic execution. epic implement --agentic analyses the story dependency graph and hub-file overlap, then runs safe waves of parallel implementation agents (Claude Code only), with quality gates at every wave boundary and a lessons file that makes each wave smarter than the last.

Cross-project memory. A lessons registry ships with the skill (lessons recall before big decisions; lessons add --global to promote what you learn), and each project accumulates its own pitfall file that agentic waves inject into every prompt.

Two-way GitHub sync. CRs, stories, and epics sync with GitHub Issues through the gh CLI; merged PRs trigger completion cascades.

Output structure

sdlc-studio/
  prd.md  trd.md  tsd.md  personas.md
  epics/      EP0001-*.md  + _index.md
  stories/    US0001-*.md  + _index.md
  plans/      PL0001-*.md  + _index.md
  bugs/       BG0001-*.md  + _index.md
  test-specs/ TS0001-*.md  + _index.md
  crs/  rfcs/  reviews/  workflows/
  .local/                  # gitignored: caches, reports, lessons
tests/                     # generated test code (project root)

Requirements

• Python 3.10+ for the bundled scripts (pure stdlib, no pip installs)
• gh CLI (authenticated) only for the GitHub sync commands
• Whatever test runners your AC verifiers invoke (pytest, vitest, go...)

Upgrading from v1.x

Re-run the installer - it replaces the skill in place and sweeps other locations up to the same version. No project migration is needed: the artifact schema is unchanged (schema_version: 2), so existing sdlc-studio/ directories keep working. What changed in v2 is the skill itself: open Agent Skills frontmatter, $CLAUDE_SKILL_DIR script paths, consolidated reference docs, new helpers and CI guards - see CHANGELOG.md.

Roadmap

Candidates for v2.1 (recorded from the AI-DLC review; not committed): task-level dependency DAGs for finer-grained parallel waves, review iteration history as first-class artifact fields, and a lightweight artifact-relationship graph for impact analysis.

Troubleshooting

• /sdlc-studio not found - confirm the skill directory exists for your tool (install.sh --list-targets), then restart the session; Claude Code also live-reloads project-level skills.
• Installer download fails - check network/proxy; you can install manually: git clone this repo and copy .claude/skills/sdlc-studio/ into your tool's skills directory.
• Commands run but nothing happens - run /sdlc-studio status; if the pipeline is empty the next step is prd create or prd generate.
• Uninstall - ./install.sh --uninstall (same --target/scope you installed with); preview with --dry-run.

Contributing and development

Dev instructions live in AGENTS.md (read natively by Codex, Copilot, Cursor, Gemini; Claude Code imports it via CLAUDE.md). Lint with npm run lint (markdown, style, links, frontmatter spec, version consistency, line budgets); test with npm test. Behavioural eval scenarios live in evals/. See CONTRIBUTING.md.

Documentation

• docs/INSTALL.md - full installer reference
• /sdlc-studio help - command catalogue (also .claude/skills/sdlc-studio/help/help.md)
• .claude/skills/sdlc-studio/reference-doctrine.md - the operating doctrine for running any project with this skill
• CHANGELOG.md | SECURITY.md | SUPPORT.md