🔥
The #1 Claude Code Problem: Token Burn
Context limits, re-investigation loops, zero memory between sessions, no cost routing — these are the failure modes engineers hit daily. obra/superpowers gives you 14 structured workflows. It does not give you memory, context management, cost routing, or codebase onboarding. Every session restart costs 5,000–15,000 tokens to re-establish context. Every repeated problem costs 8,000–15,000 tokens to re-investigate. This is the gap that was engineered closed.
99%
Token Reduction
on Repeat Problems
Chronicle 3-layer memory vs obra's zero memory
0
LLM Tokens
on Mechanical Tasks
Vector Tier 0 → shell tool, no API call
80%
Savings on
Codebase Exploration
Pathfinder 5-phase vs random file reads
100%
Update
Survivability
9 skills pinned via SessionStart hook

Token Burn by Scenario — The Brutal Comparison

Real engineering scenarios. Actual token costs modeled from source file word counts (words × 1.3 = tokens). obra numbers use their available skills only.

6 Real Engineering Scenarios — obra vs Steroids (tokens consumed)

Lower is better for routine tasks. Steroids costs more upfront on first encounter — pays back massively on every subsequent encounter.

obra/superpowers
Claude Code on Steroids
Scenario 1: Session Bootstrap
obra (using-superpowers)1,023 tok
Steroids (ascend)1,484 tok
Delta+461 tok
Steroids loads 13 domain triggers + 6 workflow chains. The overhead pays back immediately on first real task.
Scenario 2: Debug — 1st Encounter
obra (hunter only)1,955 tok
Steroids (oracle+chronicle+hunter)4,246 tok
Delta+2,291 tok
First encounter costs more — oracle classifies, chronicle is queried (cache miss), hunter executes. Investment.
Scenario 3: Debug — 2nd Encounter ⚡
obra (full re-investigation)8,000–15,000 tok
Steroids (chronicle cache hit)~50 tok
Savingsup to 99%
Chronicle retrieves the stored solution in a 3-layer token-efficient search. obra has zero memory — re-investigates from zero every time.
Scenario 4: Context Limit Hit
obra (restart + re-establish)5,000–15,000 tok
Steroids (horizon compresses)0 restart cost
Savings100%
obra has no context management skill. At limit: crash, restart, manually reconstruct. Horizon tracks budget and compresses before the wall hits.
Scenario 5: Mechanical Task
obra (LLM call always)500–2,000 tok
Steroids (Vector Tier 0)0 tok
Savings100%
Rename variable, format file, add import — Vector routes these to shell tools. Zero API call. obra routes everything through the LLM.
Scenario 6: New Codebase Exploration
obra (random file reads)30,000–60,000 tok
Steroids (pathfinder 5-phase)8,000–12,000 tok
Savings60–80%
Pathfinder maps entry points, dependency graph, and architectural traps systematically. obra = explore and hope.

Simulated 5-Day Engineering Sprint — Cumulative Token Burn

8 tasks/day across all 6 scenario types. obra re-investigates every repeated problem. Steroids caches, routes, compresses.

Cumulative Token Burn — 5-Day Sprint

Day-by-day total. obra diverges after day 1 as repeat problems compound. Steroids flattens due to chronicle cache hits.

obra/superpowers
Claude Code on Steroids

Assumes: 2 debug repeats/day, 1 mechanical/day, 1 new codebase/week, context limit 2×/week. Midpoint estimates used for ranges.

Per-Scenario Token Savings (single occurrence)

Where the savings actually come from. Chronicle and Pathfinder dominate. Vector is a perfect zero.

Tokens Saved vs obra

Skill Content Depth — Word Count Analysis

Actual word counts from source SKILL.md files. More content = more encoded decision-making = less LLM reasoning required.

Evolved Skills — Content Depth Comparison (words)

12 skills evolved from obra originals. GadaaLabs additions (API verification, Phase 0, domain triggers) drove the largest gains.

obra original
Steroids version

10 New Skills — Token Load (not in obra)

New skills that don't exist in obra at all. Each adds structured knowledge the LLM would otherwise derive from scratch each session.

Tokens loaded per skill

System Coverage — 6-Dimension Radar

Composite scoring across the dimensions that matter for serious engineering work.

Feature Coverage Radar

obra excels at structured workflow coverage. Steroids adds intelligence, memory, and domain layers.

obra/superpowers
Claude Code on Steroids

Skill Count Breakdown

obra: 14 skills. Steroids: 24. 9 overridden with GadaaLabs engineering additions. 10 entirely new.

obra skills (14 total, none in Steroids)
Brand-new skills (10)
Evolved + pinned (9)
Evolved, upstream (5)

GadaaLabs Engineering Additions — What Was Missing

6 specific gaps identified in both obra and the official expansion. Directed and built by GadaaLabs.

Engineering Gap Coverage — Binary Presence Check

Does the capability exist? obra is a zero across all 6 gaps. The official expansion missed them too.

obra/superpowers
Claude Code on Steroids

Complete Skill Inventory — All 24 Skills

Full origin tracking, GadaaLabs additions flagged, override status shown.

Codename Role Origin GadaaLabs Addition Pinned
ascendSession bootstrapEvolved (obra: using-superpowers)13-domain trigger table + 6 workflow chains✓ Pinned
blueprintPlanningEvolved (obra: writing-plans)Phase 0 API Documentation Discovery✓ Pinned
forgeTDDEvolved (obra: test-driven-development)API Pre-Verification before test authorship✓ Pinned
commanderParallel dispatchEvolved (obra: dispatching-parallel-agents)COMMANDER vs PHANTOM 6-row decision table✓ Pinned
vectorModel cost routingNew (no obra equivalent)Tier 0 ($0) → Tier 3 (Opus) routing logic✓ Pinned
chronicleSelf-learning memoryNew (no obra equivalent)ReasoningBank + 3-layer token-efficient search✓ Pinned
pathfinderCodebase onboardingNew (no obra equivalent)5-phase systematic exploration framework✓ Pinned
phantomSequential executionEvolved (obra: executing-plans + subagent-driven-development)✓ Pinned
legionSwarm coordinationNew (no obra equivalent)4 topology templates: hierarchical, mesh, ring, star✓ Pinned
oracleTask classifierNew (no obra equivalent)Complexity tier + model + skill chain assignment
horizonContext managementNew (no obra equivalent)Token budget tracking + compression + clean handoff
gradientML engineeringNew (no obra equivalent)Pipelines, training, serving, MLOps, drift detection
nexusAI/RAG engineeringNew (no obra equivalent)RAG, agents, prompt engineering, LLM eval
ironcoreEmbedded systemsNew (no obra equivalent)ISRs, RTOS, state machines, HAL, timing
prismUI/UX engineeringNew (no obra equivalent)67 styles, 25 chart types, WCAG 2.1 AA, Core Web Vitals
hunterSystematic debuggingEvolved (obra: systematic-debugging)
sparkBrainstormingEvolved (obra: brainstorming)
mirrorReceive code reviewEvolved (obra: receiving-code-review)
criticRequest code reviewEvolved (obra: requesting-code-review)
exodusFinish branch / shipEvolved (obra: finishing-a-development-branch)
sentinelVerificationEvolved (obra: verification-before-completion)
scaffoldWriting skillsEvolved (obra: writing-skills)
anchorGit worktreesEvolved (obra: using-git-worktrees)
archivistThis skill guideEvolved (obra: using-superpowers)