P.H.R.E.D.
Project Hub for Retrieval, Execution, & Delivery
A complex mobile app, two people, and a production MVP that would have been nearly impossible to coordinate without the right structure. This is the operating methodology that made it workable: the stack, the source-of-truth system, the workflow patterns, and an orchestration agent that keeps AI-assisted work connected instead of scattered.
Shared as a public case study in practical AI orchestration, not AI theater.
Applied case study
Context
The Need
The project that motivated this was complex. Confidential product, small team, limited budget, aggressive timeline.
The problem with AI-assisted work isn’t output. It’s that the output goes everywhere. Product decisions in one doc. Design exploration in another. Tasks created without context. AI conversations that go nowhere. Nothing connects.
The answer was to build the operating system before accelerating the work. Project OS gives AI tools enough structure to actually help: planning, tasks, handoffs, QA, documentation. And it keeps humans in control of scope, privacy, and decisions that can’t be undone. The product details stay confidential. The methodology doesn’t.
Overview
The system
This is the stack. These are the tools and patterns in active use.
Google Workspace
Working docs, meeting notes, brainstorming, shared context
Asana
Roadmap, Gantt, tasks, dependencies, milestones, weekly status
Figma
Design source of truth, components, tokens, platform guardrails, handoff
GitHub
Versioned methodology docs, implementation notes, issues, commits, and proof of work
Claude Code
Agent-assisted planning, repo work, architecture review, Asana automation
Claude Design
Design exploration, creative iteration, concept review, visual direction
Codex
Setup support, code assistance, implementation review, structured file edits
Slack
Planned team communication and agent notification surface
MCP
Controlled bridge between AI agents and external tools
Project OS
Source-of-truth patterns, rules, reusable prompts, approval gates, and context management conventions
Phred
Orchestration-agent pattern: reads approved context, proposes actions, flags risks, and prepares handoffs for human review
System map
How the operating layer works
Phred doesn’t replace human judgment. He makes sure fewer things fall through the cracks before the work reaches you.
Information architecture
What counts as truth?
Not every document carries the same weight. The system helps AI-assisted workflows know the difference between a committed decision and a half-finished note.
This is what stops an AI tool from treating a brainstorm like a committed requirement.
Workflow
Design to code
Explore & design
Review & hand off
Build & ship
Design exploration, approved design, engineering, and QA stay connected. The confidential specifics don’t need to be here for the workflow to make sense.
Orchestration
Meet Phred
Phred has a job. Read the right context, identify what changed, propose the next action, and flag what needs a human. No broad authority. No autonomous decisions. A well-defined role for a well-defined tool.
What Phred can do
- +Summarize current project state from approved sources
- +Read selected methodology and project-context docs
- +Draft proposed tasks for review
- +Flag missing source-of-truth links
- +Identify blockers and stale tasks
- +Prepare weekly status drafts
- +Extract decisions and open questions from notes
- +Prepare design-to-code handoff summaries
- +Convert QA findings into proposed tasks
What Phred cannot do
- ×Approve MVP or roadmap scope changes
- ×Change permissions or billing
- ×Access secrets
- ×Send external communications
- ×Move future-phase work into current-phase scope
- ×Approve privacy or security decisions
- ×Merge production code
- ×Treat brainstorms as requirements
Integration map
Touchpoints
Planning
- Project OS
- Google Workspace
- Asana
Project OS defines what's approved. Everything else (docs, tasks, meetings) connects to that foundation.
Design
- Figma
- AI Design Exploration
- Figma Dev Mode
One place for visual decisions. AI speeds up exploration. A structured handoff makes sure engineering gets what it needs.
Engineering
- GitHub
- Claude Code
- Codex
- React Native / Expo
The repo holds methodology docs, notes, and proof of work. AI coding tools assist with implementation; humans review before anything ships.
Automation
- MCP
- Asana
- Phred PM's all Asana/Meeting tasks
- Weekly status workflow
Phred routes context between systems. MCP keeps the integrations controlled. What surfaces is what needs a decision.
Quality Control
- QA planning across mobile platforms
- Event simulation mode
- UAT
QA plans run against approved requirements. Simulation and UAT run before production. The confidential specifics don't need to be here.
Controls
Why the guardrails matter
AI gets more useful when the rules are explicit. Project OS tells agents what’s approved, what’s open, what’s off-limits, and when to stop and ask.
- ✓Current-phase versus future-phase separation
- ✓Approval gates before scope changes
- ✓Privacy and permissions rules
- ✓Platform-specific implementation guardrails
- ✓Offline, stale-state, or failure-mode requirements where relevant
- ✓Simulation and UAT requirements
- ✓Design system discipline
- ✓No secrets in docs or prompts
- ✓Human approval required for high-risk changes
What’s missing
Future methodology extensions
RAG / Retrieval Layer
- Better context control: less raw content dumped into models, more structured retrieval
- A proper retrieval layer for approved docs instead of injecting full markdown into every prompt
Specialized Sub-agents
- Repetitive process support
- QA review
- Requirements auditing
- Platform-specific guardrail review
Takeaways
What this proves
Systems thinking
Built the operating model before the product. Rework is expensive. Context loss is more so.
AI orchestration
Connected AI to real product, design, engineering, and QA workflows. Human approval stayed in the loop where it mattered.
Creative operations
Converted ambiguity into structure, tasks, handoffs, and decision paths.
Technical fluency
The tool stack is real: Claude Code, Codex, MCP, GitHub, Figma, Asana, React Native / Expo. Each one chosen for a reason.
Product judgment
Kept humans in control of privacy, permissions, scope, and anything that can't be undone.
Public methodology note
Reusable by design
This methodology is reusable. The patterns (source-of-truth structure, approval gates, agent boundaries, decision logs) can apply to any team that needs AI-assisted work to stay connected and under control.
What’s not here: the specific product, company, code, designs, or business details. Those are confidential. Everything else is fair game.
This is the kind of AI work I’m interested in.
Not prompts for the sake of prompts. Not AI theater. Practical systems that help small teams move faster without losing control of the product, the roadmap, or the standards.
Built by Randall Fransen. Applied during work on a confidential mobile app project. Product specifics stay confidential.