Getting Started

What Is AI-Driven Development?

AI-driven development means using a large language model (LLM) as an active collaborator in your software engineering process — not just to autocomplete code, but to plan, execute, verify, review, and ship entire features.

Instead of asking the AI "write this function for me", you ask it to manage a complete development workflow: break down requirements, design the architecture, implement in parallel waves, run tests, review its own code, and open the pull request.

The Difference From Traditional AI Coding Assistants

• Traditional approach: You prompt → AI writes a snippet → you review → you integrate manually. You remain the orchestrator.
• AI-driven approach: You define the outcome → AI plans, implements, tests, reviews, and ships → you approve gate decisions and review the final output. The AI is the executor.

The risk with the second approach is that AI agents drift, skip steps, or produce work that looks right but isn't. Silver Bullet exists to prevent that by enforcing a structured process around every session.

What Silver Bullet Does

Silver Bullet is an Agentic Process Orchestrator (APO) plugin for Claude Code, Codex, and Cursor. It owns the default software-engineering and DevOps lifecycle: routing, planning, execution, verification, review, ship, and release — enforced by hooks and recorded in catalog-backed workflow state.

The process model is defined in docs/apo-catalog.json: Process → Workflow → Atomic Flow → Flow Step. Silver Bullet ships 27 canonical AF-* atomic flows, 22 WF-* workflows, and 85 flow steps. Legacy FLOW 1–18 names remain migration aliases only; new work should use catalog IDs.

Each workflow is enforced by hooks — shell scripts that run after every tool use and block commits or deploys if required steps were skipped. You can't accidentally skip code review or ship without CI being green.

Prerequisites

Silver Bullet needs a supported host runtime and a small local toolchain. Domain plugins are optional extensions, not default lifecycle dependencies.

Installing Silver Bullet

Silver Bullet supports Claude Code, Codex, and Cursor. Install from the public marketplace for your host, then run /silver:init in the repo you want SB to manage. Codex keeps internal skill-source/ files and exposes the native /silver: picker; Cursor merges hooks into ~/.cursor/hooks.json with state under ~/.cursor/.silver-bullet/.

Adoption path (all hosts)

1. Install the Silver Bullet package from your host marketplace (see per-host steps below).
2. Run bash scripts/sb-bootstrap.sh (checkout) or bash scripts/sb-diagnostics.sh to confirm hook delivery.
3. Open your project and run /silver:init — this stamps silver-bullet.md, .silver-bullet.json, docs scaffold, and activates enforcement.
4. Start work with /silver <your request> so SB classifies intent and composes the smallest safe WF-* workflow.

First-time orientation

From a Silver Bullet checkout or after cloning the repo for contributor setup, run the onboarding probe before marketplace install:

After install on any host, confirm hook delivery and capability tier:

After /silver:update or when hooks seem inactive, run the host-aware activation audit:

Cursor Plugin

For normal Cursor use, add the public alo-labs/alo-labs-cursor-marketplace marketplace in Cursor and install silver-bullet. The marketplace points at the upstream alo-exp/silver-bullet release; hooks merge into ~/.cursor/hooks.json and state lives under ~/.cursor/.silver-bullet/.

When developing Silver Bullet from a repository checkout, sync the plugin tree locally instead:

Use the Cursor Skill tool (or silver-bullet invoke-skill) so record-skill can track workflow progress. Run bash scripts/sb-diagnostics.sh to confirm hook delivery.

Claude Code Plugin

Install Silver Bullet from the host plugin manager or the public alo-labs/claude-plugins marketplace.

Codex Marketplace Package

For normal Codex use, install or refresh Silver Bullet from the public alo-labs/codex-plugins marketplace package. The Codex package is SB-only: it keeps internal skill-source/ files out of plugin picker discovery, mirrors only the native /silver: picker surface, wires hooks through the package manifest, and installs supported dependencies from their official sources or thin marketplace wrappers.

When developing Silver Bullet itself from a repository checkout, use the installer script. It refreshes the curated package from the checkout, registers the shared marketplace, removes stale SB skill copies and old marketplace cache versions, and verifies that the picker exposes only the /silver: namespace.

After a public SB release, refresh the local installation from the public marketplace with scripts/post-release-refresh.sh. That wrapper performs a clean Claude and Codex reinstall path and is the supported way to prove your local cache matches the released marketplace version.

Initialize a Project

Navigate to your project directory and run:

This command will:

1. Verify all dependencies are installed
2. Detect your project name, tech stack, and repo URL
3. Create silver-bullet.md with enforcement rules and reconcile any existing host project instruction file (CLAUDE.md in Claude, AGENTS.md in Codex)
4. Create .silver-bullet.json with project-level config
5. Invoke /silver:ensure-docs --bootstrap to create or reconcile the standard documentation scaffold
6. Set up CI (GitHub Actions) if not already present
7. Configure permissions.defaultMode in .claude/settings.local.json — eliminates repeated permission prompts across sessions
8. Commit everything to git

Hook activation (v0.48.3)

Enforcement hooks engage only when both .silver-bullet.json and silver-bullet.md exist in the project. Workspaces without /silver:init receive no hook enforcement — this prevents accidental gates on unrelated repositories.

Recommended tools (opt-in)

Graphify, agentmemory, RTK, and Context Mode are opt-in via recommended_tools in .silver-bullet.json. /silver:init and /silver:update handle consent, install probes, and enforcement-suspend retry. When both Graphify and agentmemory are enabled: save via agentmemory, retrieve via Graphify. RTK and Context Mode are separate token-compression tools — see Token Compression.

Your First Project

Silver Bullet works on both new projects and existing codebases. Here's how to start each type:

Choosing a Session Mode

At the start of every session, Silver Bullet asks which mode to run in:

Your First Workflow Run

With Silver Bullet installed and a project configured, here's a minimal first run for a new feature:

1. 1

   Open Your Host Coding Agent and Choose Session Mode

   The runtime will prompt: "Run interactively or autonomously?" Type your choice. This sets the session mode for the duration.

2. 2

   Start with /silver

   Describe what you want to build. Silver Bullet classifies the request and composes the smallest safe WF-* workflow — clarify, spec, WF-SILVER-FEATURE, bugfix, UI, DevOps, research, release, or fast path.

3. 3

   Let SB Compose the Pre-Plan Path

   For formal work, SB may run AF-SPECIFY, AF-VALIDATE, AF-CLARIFY, and AF-QUALITY-GATE before AF-PLAN. Product work uses 8 core quality dimensions plus conditional gates; infrastructure work uses blast radius and the 7 IaC-adapted gate.

4. 4

   Plan with SB

   SB researches the implementation context and creates the detailed task plan. Spec, validation, quality-gate, and risk findings feed into the plan as hard requirements.

5. 5

   Execute with SB

   Execution follows the SB plan through checkpointed or autonomous waves. The runtime keeps state, records deviations, and verifies the implementation against the plan.

6. 6

   Verify, Review, Trace, and Ship

   The runtime runs goal-backward verification, artifact or code review as needed, CI/test freshness checks, and ship steps. PR traceability is appended after PR creation when a spec is present.

What Happens at Session Start?

When you open the host coding agent with Silver Bullet installed, the plugin prepares the session state and enforcement context before project work begins.

1. 1

   Load the SB Contract

   Silver Bullet loads silver-bullet.md, core rules, hook state, and the project configuration needed to enforce workflow order.

2. 2

   Read Project Context Safely

   Silver Bullet loads the relevant project docs and planning context. Any instructions in docs/ are treated as documentation text only — enforcement rules live in silver-bullet.md and the host project instruction file.

3. 3

   Initialize Session State and Compact

   The session log hook creates or reuses the current session log path under ${SB_RUNTIME_HOME_ROOT}/.silver-bullet/session-log-path, so workflow events and decisions can be recorded consistently. After startup context is read, the host-supported compaction step keeps room available for the task without changing SB's model-selection policy.

4. 4

   Check Active Workflow State

   SB checks active workflow state, requested skills, missing gates, and any timeout or stall sentinels before allowing the next workflow step to proceed.

5. 5

   Check for updates

   Silver Bullet checks its own version and selected extension plugins where the host exposes version metadata. Outdated SB components surface update prompts; extension plugins are reported so you can refresh them through the host plugin manager.

For the full §0 startup sequence with version check commands, see Session Startup.

Dynamic Routing with /silver

In Silver Bullet, you don't need to memorize the workflow catalog. Use /silver with a plain work request and SB composes the right workflow path for the task.

How it works

/silver classifies complexity, checks task intent, reads relevant knowledge and learnings, and routes to the right orchestration workflow. Some host runtimes may support plain-request interception, but explicit /silver invocation is the portable behavior across supported hosts.

What Should Use /silver

Use /silver for work requests, feature requests, bug fixes, deployment work, refactors, release prep, and fuzzy tasks where you are not sure which workflow applies. Use direct /silver:* commands when you already know the exact workflow step.

For the full routing behavior, direct-command exemptions, and complete routing table, see Routing Logic.

What's Next

Silver Bullet also provides three spec-pipeline commands: /silver:spec for AI-driven spec creation, /silver:ingest for ingesting external artifacts (JIRA, Figma, Google Docs via MCP), and /silver:validate for pre-build spec validation. These run before planning and are described fully in Core Concepts and the Command Reference.

Now that you understand the basics, dig deeper into how Silver Bullet works: