agentstrap

AgentStrap v2

Turn Claude Code into a disciplined development partner. A lightweight, EA-style methodology — structured docs, decision governance, and a multi-perspective review swarm — that carries a project from idea to shipped, and keeps that state alive across every session and device.

   idea ─▶ decisions ─▶ design ─▶ build ─▶ review ─▶ ship
   └──────── one governed, documented workspace ────────┘
            kept current across every session & device

It replaces "vibe-coding with amnesia" with a repeatable process: every decision is logged, every open question is tracked, every review is multi-angle — and nothing is lost when you switch machines.

---

The four pillars

Pillar	What it gives you
🗂️ Structured workspace	A domain-segmented docs vault (Foundations → Product → Design → Engineering → Operations → Business → Reference) — an EA-style backbone instead of scattered notes
⚖️ Decision governance	An append-only Decisions Log ⇄ Open-Questions register + PM-first working rules, so choices are deliberate and traceable
🔍 Review swarm	One command fans out 8 specialist reviewers (architecture, security, performance, …) → a single, deduped, prioritized report
🔄 Living continuity	Hooks keep your "where we left off" current every turn and restore it on any device — the discipline survives closing the laptop

Continuity isn't the headline feature — it's what makes the other three survive across sessions. A governed process is worthless if its state evaporates when you close the tab.

---

TL;DR

# 1. install (once per machine)
/plugin marketplace add https://github.com/mateim4/agentstrap-v2
/plugin install agentstrap@agentstrap-v2

# 2. set up a project (once per project) — run Claude from inside the project folder
/agentstrap:bootstrap

# 3. work the process: log decisions, track questions, audit, ship.
#    Continuity is automatic from here on.

---

What problem it solves

Pain	AgentStrap's answer
Projects start as ad-hoc chats with no structure or paper trail	A scaffolded, domain-segmented workspace from day one
Decisions get made in chat and forgotten / re-litigated	Append-only Decisions Log + Open-Questions register
Reviews are one-dimensional ("looks fine to me")	8-perspective audit swarm → consolidated P0–P3 report
Every project reinvents its own process	One reusable, opinionated methodology, installed once
Switch device → lost the thread, "where were we?"	Auto-handoff written every turn, restored on the other machine

---

The methodology

AgentStrap imposes a light, opinionated structure so a project is always organized the same way:

00 Foundations   principles · working rules · Decisions Log · Open Questions
10 Product       vision · wave-based discovery · requirements · user flows
20 Design        wireframes · design system · UX research
30 Engineering   architecture · data model · ADRs · integrations · runbooks
40 Operations    sprint board · backlog · milestones
50 Business      monetization · legal · go-to-market
90 Reference     external docs · research (cited)

Governance is a simple loop: open a question → record the lean → decide → migrate it to the Decisions Log with a date. Architecture is pressure-tested by the review swarm before it's trusted.

How this maps to TOGAF (and where it deliberately stops)

The domains and governance are TOGAF-inspired, not TOGAF:

AgentStrap	TOGAF analogue
00 Foundations	Preliminary Phase + Architecture Principles/Governance
10 Product	Phase A — Architecture Vision
50 Business	Phase B — Business Architecture
30 Engineering	Phases C/D — Information Systems + Technology Architecture
40 Operations	Phases F/G — Migration Planning + Implementation Governance
90 Reference	Architecture Repository
Decisions Log ⇄ Open Questions	Architecture decision log + governance
Audit swarm	Architecture compliance review

It intentionally omits TOGAF's heavy machinery — the formal ADM iteration cycle, stakeholder/viewpoint framework, capability & maturity models, the Enterprise Continuum. This is governance for a small team that wants rigor without the ceremony.

---

How it works (global engine + project state)

AgentStrap is two halves: