Development Workflow — Two-Layer Planning Model
How strategy and execution are separated between humans and AI agents in Bitcoinology.
The Model
Key Principle
Nobody prescribes how to implement. Only what "done" looks like.
- Executives define milestones with verification criteria and human gates
- Agents read those milestones, study the codebase, create their own work breakdown, and execute
- Human gates are checkpoints where execution pauses for executive review
Two Layers Explained
Linear — The Executive View
Linear holds the strategic layer. Executives see projects, milestones, and high-level progress.
Each milestone has:
- Verification criteria — concrete test that proves it's done
- Human gate — executive must approve before next milestone starts
Beads — The Execution View
Beads (bd) holds the agent's work breakdown. The agent creates its own epics, tasks, and dependency graph.
The agent decides:
- How many tasks to create
- What dependencies exist between them
- What order to execute
- How granular to go
The /ship Loop
The Factory Droid runs the /ship command, which follows this cycle:
Milestone Verification Criteria
Each milestone has a specific test. The executive runs these to verify:
| Milestone | Verification | How to Test |
|---|---|---|
| Hybrid Search Live | Exact terms return accurate results | Search "UTXO consolidation" in the app |
| BYOK Key Management Live | User's key is used for LLM calls | Check OpenAI dashboard — query should NOT appear on our account |
| User Tiers Enforced | Free tier capped at 10/day | Send 11 search queries as free user |
| Oracle Agent | Complex queries decompose | Ask "How has Lightning narrative evolved?" — check sub-queries in trace |
| Citation Gate | No ungrounded claims pass | Review 10 responses — every claim should cite episode + timestamp |
| Deep Research E2E | Full pipeline works | Run flagship query, verify multi-step cited output |
| All 8 Workflows | Each workflow functional | Test each with 3 example queries from PRD |
| Contextual Embeddings | Retrieval improves | A/B comparison on 20 queries |
| Platform Features | Each feature works | Manual test: X login, export, audio, analytics |
Chain of Command
Tools
| Tool | Role | Used By |
|---|---|---|
| Linear | Project management, milestones, human gates | Executives |
| Beads (bd) | Task tracking, dependency graph, agent memory | Agents |
| Factory Droid | AI implementation agent | Automated |
| Claude Code | CEO / orchestrator | Strategic |
| GitHub | Code hosting, PRs, CI | Everyone |
| Vercel | Production hosting, auto-deploy from main | Automated |
Key Files
| File | Purpose | Audience |
|---|---|---|
AGENTS.md | Agent instructions, conventions, guardrails | Droid |
.factory/droids/ship.md | /ship command — autonomous implementation loop | Droid |
.factory/commands/ship.md | /ship trigger | Droid |
docs/gap-analysis.md | What to build — 13 gaps, 3 phases | Everyone |
docs/gap-analysis-technical-risk.md | Security & scalability findings | Everyone |
docs/gap-analysis-user-journey.md | UX gap findings | Everyone |
Related
- Gap Analysis: PRD vs Reality — the feature gap map
- Project Status — timeline and completed milestones
- Operating Model — team topology
Established 2026-02-24. The two-layer model ensures agents have autonomy to plan while executives retain control at milestone boundaries.