Search everything...

Stats

Actions

Available In

agent-methodology

Name: agent-methodology
Author: engsimsoft

By engsimsoft

Universal Claude Code methodology + memory + lifecycle for long-running projects (100+ TZ, 500+ sessions, 1-2 year horizon). Anthropic-prescribed stack: disciplined CLAUDE.md, Memory Bank, ADR enforcement, Serena LSP, deterministic hooks. No vector RAG, no cloud memory by default — auditable markdown + structural hook protection.

npx claudepluginhub engsimsoft/universal-agent-starter --plugin agent-methodology

Popularity

Stars

Med: 0·Avg: 284

Installs

Med: 0·Avg: 1

What's Inside

Slash Commands9

/adr

Create a new ADR (Architecture Decision Record) from a one-line input. Owner says "use FastAPI instead of Flask because async support" — this command generates docs/agent/decisions/ADR-NNNN-<slug>.md in MADR format. INDEX.md auto-regenerates via post-tool-edit hook.

/close-series

Formally close a series after DoD is met. Verifies all TZ closed, full test suite passes, manual smoke owner OK, then archives series to specs/_archive/.

/close-tz

Finalize a TZ — verify validation, run docs-sync, process FINDINGS, archive folder, create one commit. Use this at end of every TZ.

/handoff-now

Manually trigger pre-compaction handoff before context degrades. Use when sensing context drift, read-twice signal, or session running long. Saves checkpoint to .claude-data/checkpoint.md.

/init-project

Bootstrap a new project with the agent-methodology structure. Creates docs/agent/, specs/, CLAUDE.md, .claude-data/, invariants.md from plugin templates. Run this once per fresh project.

Agents3

changelog-writer

/changelog-writer

Generate a concise CHANGELOG.md entry from git diff + commit messages. Compresses a TZ's worth of changes into a structured entry without narrative bloat. Use when running /close-tz or docs-sync skill.

code-auditor

/code-auditor

Audit a module or feature in the codebase. Returns a 200-word summary of structure, dependencies, smells, and ADR compliance — without polluting main conversation context. Use for "how is X currently implemented?" or "find all call sites of Y" or "is the auth module aligned with ADR-0023?". Read-only — no edits, no commits.

doc-writer

/doc-writer

Generate SPEC.md and ROADMAP.md for a new TZ from owner's outline. Owner provides a one-paragraph description of what they want, agent produces structured SPEC + ROADMAP per methodology format. Use when starting any non-trivial feature TZ.

Skills7

archive-cycle

/archive-cycle

Use this skill at the end of any TZ closure or series closure to archive completed work and keep specs/ + docs/agent/ from bloating over a 1-2 year project. Trigger automatically when closing series, archiving 30+-day-old TZs, or when changelog exceeds 1000 lines. Critical for projects with 100+ TZ over years.

bugfix

/bugfix

Use this skill ALWAYS whenever there is a reproducible bug, failing test, runtime error, regression, or any defect report. Trigger this skill EVEN if the user just says "X is broken" or "fix Y" without describing the process. Use it for ANY defect work — never just patch and close.

docs-sync

/docs-sync

Use this skill ALWAYS at the end of ANY code-changing task to synchronize project memory with reality. Trigger this skill after closing any TZ, after fixing any bug, after completing any refactor, after adding any feature, after creating any ADR — even if the user did not explicitly request docs update. This skill is MANDATORY before declaring a task complete.

feature-delivery

/feature-delivery

Use this skill ALWAYS whenever adding new functionality, new endpoint, new component, new agent, or extending existing behavior in a non-trivial way. Trigger this skill at the start of ANY feature work even if the user just describes "add X" or "implement Y" without explicitly mentioning a process.

official-docs-first

/official-docs-first

Use this skill ALWAYS whenever the task touches an external technology (SDK, API, library, framework, database, MCP server, AI provider) OR any knowledge after the model's cutoff date. Trigger this skill EVEN for technologies you think you know — your training data may be outdated by months. This skill is MANDATORY before writing any code that uses external dependencies.

Hooks1

Event Hooks

Bash

File writes

10 hooks across 8 events

MCP Servers3

Stats

Version0.1.0

LanguageShell

Stars0

MaintenanceGood

LicenseMIT

Last CommitMay 18, 2026

AddedMay 18, 2026

Actions

View on GitHub View README Plugin Marketplace JSON

Own this plugin?

Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).

Available In

universal-agent-starter

Safety Signals

Caution

Executes bash commands

Hook triggers when Bash tool is used

Modifies files

Hook triggers on file write and edit operations

README

Universal Agent Starter

Claude Code plugin marketplace bundling a methodology + memory + lifecycle stack for long-running projects (100+ TZ, 500+ sessions, 1-2 year horizon).

Built from real lessons of a failed 50K-line project (Simply v1) and verified against 2025-2026 best practices: Anthropic's effective-context-engineering, Claude Code Plugins/Hooks/Skills, Cline Memory Bank pattern, Serena LSP. No vector RAG, no cloud-graph memory by default — auditable markdown + deterministic hook protection.

Why this exists

Long-running AI-agent projects fail not from missing features but from context degradation over time. By session 200, the agent forgets architectural decisions, re-reads the same files, accepts stale HANDOFF as truth, and proposes opposite of what was decided in March. Without deterministic protection, this is inevitable.

This plugin enforces 16 specific failure modes through deterministic hooks that don't rely on agent discipline.

Install

In any new (or existing) project:

/plugin marketplace add github:engsimsoft/universal-agent-starter
/plugin install agent-methodology@universal-agent-starter

After install, bootstrap the project structure (docs/agent/, specs/, CLAUDE.md, ...) by running:

/init-project

Note on the Setup hook: Per Claude Code docs, the Setup hook event fires only on claude --init-only, or --init/--maintenance in -p (print) mode. It does not fire on a normal /plugin install or session start, so if you want auto-bootstrap from CI, run claude --init-only once after install. For interactive use, /init-project is the canonical bootstrap command.

What it provides

9 hook bindings (deterministic protection — see plugins/agent-methodology/hooks/):

Setup — bootstraps docs/agent/, specs/, CLAUDE.md when triggered via claude --init-only (CI-friendly)
SessionStart — loads real git diff + recent ADRs + checkpoint, defends FM-03 (hypothesis before facts)
PreCompact — saves invariants + scope + next-step before auto-compaction loses them, defends FM-08
PostCompact — cleans up checkpoint-history/ files older than 7 days
UserPromptSubmit — keyword-driven skill activation, decision-phrase ADR injection, tech-detection R2 reminder, defends FM-06/FM-09/FM-15
Stop — R4 question-mark check, blocks turn if decision made without ADR reference, scope-creep detector, defends FM-04/FM-05/FM-09 (with strict_mode escape hatch)
PreToolUse(Bash:alembic *) — blocks prod migrations without MIGRATION_APPROVED=YES, stops uvicorn before alembic
PreToolUse(Bash:curl *) — blocks paid-API curls without justification (Anthropic / OpenAI / xAI / Moonshot / Jina / Voyage / Cohere / Gemini / Groq / Together / DeepSeek / Mistral / Perplexity / Fireworks)
PreToolUse(Edit|Write|MultiEdit) — enforces CLAUDE.md ≤80 lines, blocks edits to invariant paths
PostToolUse(Edit|Write|MultiEdit) — auto runs mypy/ruff, regenerates ADR INDEX, surfaces lint errors

Plus a lifecycle-cron.sh script (not a hook event itself) invoked once per day from SessionStart for ТЗ archival, changelog quartering, and series-limit warnings.

7 skills (auto-activated by description, manually invokable as /agent-methodology:<name>): bugfix, refactor, feature-delivery, research-only, official-docs-first, docs-sync, archive-cycle.

9 slash commands: /init-project, /start-series, /start-tz, /close-tz, /close-series, /handoff-now, /status, /pre-flight, /adr.

3 subagents (isolated context): code-auditor, doc-writer, changelog-writer.

3 MCP servers bundled:

Serena (LSP-based code navigation — replaces vector RAG for primary loop, per Anthropic engineers' published guidance)
Context7 (current SDK docs — recommended for R2 official-docs-first; falls back to WebFetch when unavailable)
GitHub (issues, PRs, code search)

Prerequisites

User must have these installed BEFORE plugin works:

Claude Code ≥ v2.1.105
uv for Serena (curl -LsSf https://astral.sh/uv/install.sh | sh)
Node.js 18+ for Context7 (npx)
Docker for GitHub MCP

After plugin install, plugin will prompt for:

context7_api_key — optional; free tier OK, rate-limit higher with key
github_token — optional; PAT with repo scope (Classic) or fine-grained PAT with Contents/Issues/Pull-requests/Metadata read access
strict_mode — boolean, default true. When false, Stop and PreToolUse hooks downgrade hard blocks to reminders (escape hatch for emergency work)

How it differs from older methodologies

This stack is built on what 2025-2026 evidence shows actually works:

View full README on GitHub

agent-methodology

Popularity

What's Inside

Confidence

Setup

Configuration

README

Universal Agent Starter

Why this exists

Install

What it provides

Prerequisites

How it differs from older methodologies

Similar Plugins

context7-plugin

startup-business-analyst

octo

claude-buddy

creative-writing

dotnet-skills

Universal Agent Starter

Why this exists

Install

What it provides

Prerequisites

How it differs from older methodologies

Similar Plugins

context7-plugin

startup-business-analyst

octo

claude-buddy

creative-writing

dotnet-skills

Popularity

Health & Quality