The system that builds itself.
Interview, generate, critique, converge — deliver.
A well-structured bundle needs agents, behaviors, modes, recipes, context files, and skills — all wired together correctly. The thin bundle pattern and context sink discipline are easy to violate.
Without surveying existing bundles, authors rebuild capabilities that already exist. Duplicate agents proliferate. Reusable behaviors go undiscovered.
Hand-built bundles ship with structural errors, philosophical violations, and untested assumptions. There's no systematic way to verify convergence.
Every bundle is a composition problem. Getting the architecture right requires expertise in Amplifier's patterns, and even experts make mistakes without systematic review.
BundleWizard is an adversarial convergence pipeline. Describe what you need — or point it at an existing bundle — and it interviews you, designs the composition, generates artifacts through iterative critique loops, and delivers a ready-to-use bundle.
Modes are the steering wheel. Recipes are cruise control.
Adaptive-depth interview. Detects experience level, routes to create/improve/rebuild/design. Scouts the ecosystem for similar bundles.
Designs the full bundle composition — output tier, file structure, agent definitions, behavior wiring, delegation targets.
Breaks the spec into an ordered task list with file-by-file dependencies. For improvements: ordered renovation tasks from audit findings.
The convergence loop: Generate → Critique → Refine → Evaluate. Iterates up to 10 times until three-level convergence is met.
Independent verification collects evidence the bundle works. Packager stamps version, writes provenance, initializes git, and delivers.
The critic is spawned with context_depth="none" — it evaluates artifacts as-is, not as-intended. The refiner only touches what the critic flagged. Scope creep in refinement destroys convergence.
Write artifacts
Adversarial review
Targeted fixes
Three-level scoring
Pass/fail gates. Bundle loads, agent refs resolve, URIs valid, no duplicate context, modes discoverable. If any gate fails, stop.
Scored rubric ≥0.85. Thin bundle pattern, context sink discipline, agent description quality, composition hygiene. 25% each.
Domain-specific ≥0.80. Does it actually do what it claims? Expert judgment with triangulation against spec requirements.
Interviews & routes to create, improve, rebuild, or design path
Design the composition, break into ordered tasks
Writes bundle artifacts from spec & plan
Adversarial review + targeted fixes
Three-level convergence scoring after every iteration
Version stamp, provenance, git init, delivery
Ecosystem Scout surveys existing bundles for reuse. Bundle Auditor runs three-level audits on existing bundles. Orchestrator Advisor applies a 12-point litmus test when orchestrator-adjacent language is detected.
Interactive — navigate modes manually with 3 human approval gates. Autonomous — say "yolo" and the full pipeline runs without checkpoints. Machine quality gates remain enforced in both.
After exploration, the full pipeline can run autonomously via the bundle-autonomous-post-explore recipe. If the run hits trouble — failed verification, convergence stall — it offers a takeover point via STATE.yaml rather than forcing one.
The bundle-batch-generation recipe iterates over a target list with STATE.yaml tracking. Generate an entire portfolio of bundles from a single specification set.
Amplifier itself can spawn BundleWizard when it detects a missing capability. It performs the explore phase directly, launches autonomous continuation, and reports: "I needed X capability, so I built it."
Detects older bundlewizard: provenance metadata, migrates to the canonical generated_by schema, and runs targeted improvements. Point it at any older bundle — it auto-detects and offers to upgrade.
A single contributor — michaeljabbour — built a complete bundle factory with adversarial convergence loops, 11 specialized agents, 7 modes, 6 recipes, and comprehensive test coverage. In 11 days. March 16–27, 2026.
BundleWizard is the closing of the loop. Amplifier can now generate, critique, and deliver its own extensions — without human intervention in the quality pipeline.
When Amplifier detects a capability gap during autonomous operation, it spawns BundleWizard, builds the missing bundle, and continues. The system grows itself.
Three-level convergence replaces human review with machine rigor. Structural gates catch errors. Philosophical rubrics enforce patterns. Functional evaluation verifies intent. The bar is higher than most hand-reviews.
The convergence formula is non-negotiable: L1 == PASS AND L2 ≥ 0.85 AND L3 ≥ 0.80. "Simple" bundles are where the bugs hide. There are no shortcuts.
Primary source: michaeljabbour/amplifier-bundle-bundlewizard repository, cloned and analyzed directly.
Category: Developer Tools
Accent rationale: #00BCD4 (cyan) — tool/factory convention
All statistics verified via direct repository analysis. No estimated or projected numbers.