Amplifier Superpowers

Superpowers v2

Major Methodology Refresh

Comprehensive update to our Amplifier-native Superpowers — better enforcement, richer guidance, leaner skill footprint, more resilient recipes.

Active — Live on main
March 2026
Overview

What Changed

A significant update bringing our Amplifier-native Superpowers current with obra's latest methodology thinking. Every layer was touched — skills, modes, agents, and recipes.

📡
Live Upstream Skills
Methodology content pulled directly from obra's repo. No more stale copies.
🎛️
6 Modes Enriched
Every mode updated with latest methodology on enforcement, escalation, and completion discipline.
🤖
Smarter Agents
Status protocols, architecture-level review, and institutional distrust.
🔧
Hardened Recipes
Retry blocks, proper sub-recipe composition, tuned timeouts.
Architecture

Live Skills from obra/superpowers

As Jesse iterates on anti-rationalization, TDD, debugging, brainstorming, and everything else — we get those improvements automatically. We went from maintaining 7 bundle-owned skill copies to 2 Amplifier-specific ones.

Before: 7 bundle-owned copies
  • Skills copied into our bundle repo
  • Drift from upstream as obra iterates
  • Manual sync needed for every update
  • Unclear which version we're running
After: 2 Amplifier-specific skills
  • Methodology skills pulled live from obra's repo
  • Our skills listed first — can override when needed
  • Automatic methodology updates
  • Clear ownership boundary
The Right Architecture
We own the platform integration and enforcement mechanisms. Jesse owns the methodology content. Our bundle references his skills at runtime. When we need Amplifier-specific behavior, we override with our own skill — listed first in the resolution order.
Modes

All 6 Modes Enriched

Every mode now carries the latest thinking on methodology enforcement, escalation signals, and completion discipline.

💡
Brainstorm
  • Architecture guidance: design for isolation, testability, minimal interfaces
  • Scope assessment
🐛
Debug
  • "3+ fixes → question the architecture" escalation
  • Human partner signal recognition
Execute-plan
  • Implementer status protocol: DONE / DONE_WITH_CONCERNS / NEEDS_CONTEXT / BLOCKED
  • Model selection guidance
Verify
  • Red-Green regression test pattern
📝
Write-plan
  • File structure planning
  • Plan size guidance: >15 tasks → split into phases
🔄
Mode Activation
  • Agents now activate modes via tool call on user consent
  • Not just conversational — actual tool invocation
Agents

Smarter Agents

Three agents got meaningful upgrades — better status reporting, architectural awareness, and institutional distrust.

🔨
Implementer
  • Status protocol — structured exit codes (DONE, DONE_WITH_CONCERNS, NEEDS_CONTEXT, BLOCKED) instead of free-form text
  • Architecture guidance — knows when to escalate vs. proceed
🔍
Code Quality Reviewer
  • Architecture-level checks — not just line-by-line
  • YAGNI enforcement at the review layer
🛡️
Spec Reviewer
  • Strengthened institutional distrust framing
  • "The implementer finished suspiciously quickly" — built-in skepticism
Recipes

Recipe Hardening

Reliability improvements across the recipe pipeline — retries, composition fixes, and timeout tuning.

🔁
Retry with Backoff
Exponential backoff on all agent steps. API hiccups no longer kill pipeline runs.
🔗
Sub-Recipe Composition
Full-development-cycle recipe now properly composes the SDD recipe via staged sub-recipe composition. Fixes a quality gap where the full cycle was using self-review instead of the three-agent review pipeline.
⏱️
Timeout Tuning
Plan-writer and load-plan timeouts bumped for large plans. No more premature timeouts on complex projects.
Anti-Rationalization

One Source of Truth

Cross-mode anti-rationalization content — spirit vs. letter, YAGNI, false completion prevention — now lives in a shared context file, @-mentioned from every mode. One place to update, consistent enforcement everywhere.

Before
  • Anti-rationalization text duplicated across modes
  • Updates required editing each mode individually
  • Drift between modes was inevitable
After
  • Shared context file with all anti-rationalization content
  • @-mentioned from mode bodies — always in attention when a mode is active
  • Enabled by a new ecosystem capability: mode bodies now resolve @-mentions
What's Enforced
  • Spirit vs. letter compliance
  • YAGNI — you aren't gonna need it
  • False completion detection
  • Scope creep prevention
Ecosystem

Platform Improvements

Shipped alongside as improvements for everyone — not just Superpowers users.

🧩
Staged Sub-Recipe Composition
Any recipe can now compose staged sub-recipes with approval gates. Gates surface through the parent — no lost prompts.
📎
@-mention in Mode Bodies
Modes can reference shared context files using @-mentions. This is what enabled the shared anti-rationalization pattern.
⚙️
Bundle Source Overrides
Point any bundle to a different version or local path via settings.yaml. Useful for any bundle development or testing.
Sources

Sources & Links

Date: March 2026

Feature status: Active — live on main branch

Repositories:

Contributors:

Get Started

Update, then start.

Update Amplifier to get the latest bundles, then start or restart your session. The improvements are active immediately.

PR #24
Full changelog
obra/superpowers
Upstream methodology
Bundle repo
Our integration
More Amplifier Stories