Agentic Project Workflow Template

A starter file structure for running a small multi-role team (Business Analyst, Designer, Developer) working alongside Claude, extracted and generalized from the Tallskugga project.

What this is

A set of process docs and Claude Code skills that define how each role works, what they read before starting, what they hand off to the next role, and how git/PR approval flows between them. No application code — this is the scaffolding, not the product.

This is human-led by design. A person makes the actual decisions at every stage — the agent's job is to surface options and accelerate execution around them, not decide unsupervised. Each stage in process/ has a "Human checkpoint" section for what that means concretely there; in practice it usually means grill-me runs before any other skill.

Nobody except the Developer needs to know git. Say you're done with something and Claude branches, commits, pushes, and opens the PR for you — skills/save-work.md for general output, skills/save-design.md for design screens. The git rules in rules/global.md describe what happens under the hood, not something every contributor has to operate directly.

How to use it for a new project

This is a GitHub template repository — use the "Use this template" button on GitHub for a brand-new, independent repo with no shared history, or see START_HERE.md for the from-scratch/CLI path. Either way, stage-guide takes it from there: Stage 0 (Setup) first, then Discovery, Requirements, Design, and onward as the project actually progresses. The [bracketed placeholders] in roles/*.md and rules/developer.md / rules/designer.md (tech stack, tone, constraints, screens) fill in progressively as those stages produce real answers — nothing gets guessed upfront.

What's already generic (no editing needed)

• .claude/skills/stage-guide/ — the entry point. Say "where are we" or "let's continue" and it orients you: current state, what's left, and available skills for whichever stage applies — then stops and asks what you want to do. Never decides unsupervised. Reads process/0X-stage/README.md live, so it never needs updating when a stage's process changes.
• process/ — the 8-stage software lifecycle (Setup → Discovery → Requirements → Design → Implementation → Review/QA → Release → Feedback Loop), each with a starting point, end point, steps, a human checkpoint, and which skills accelerate it. Each stage also has an artifacts/ folder for working material that doesn't fit anywhere more specific (interview notes, assumption lists, etc.).
• context/workflow-status.md — tracks progress per stage (not started / in progress / done) so a session — possibly a different person, possibly weeks later — can pick up where the last one left off. Read by stage-guide, not meant to be hand-maintained.
• rules/global.md, rules/ba.md — process rules, not stack-specific
• skills/write-design-brief.md, skills/save-design.md, skills/save-work.md — the brief-writing and git-handoff mechanics, so BA/Designer never need to run git themselves
• context/requirements-template.md, context/design-brief-template.md — fill-in-the-blank templates
• .claude/skills/ — 71 Claude Code skills: 6 engineering-workflow skills from mattpocock/ai-engineer-workshop-2026-project, 64 product-management skills from phuryn/pm-skills (MIT licensed), plus stage-guide itself. See process/skill-catalog.md for the full mapping to each lifecycle stage — there's deliberate overlap in places (e.g. two different PRD skills) so different approaches can be tested against each other.

What needs filling in per project

• CLAUDE.md — project name, description, approver (Stage 0 / Setup)
• context/project-context-summary.md — the actual decisions, once made (Stage 1 / Discovery)
• roles/developer.md, roles/designer.md, roles/ba.md — product summary, screens list, current status (progressively, as Discovery/Requirements/Design land)
• rules/developer.md, rules/designer.md — stack choices, visual direction, constraints (same — don't guess these at Setup)