Gas-Shot and Ignition: Two Ways to Build With AI Without Losing Control
🎯 The problem: Fully autonomous AI builds skip your input. Fully manual ones waste your time.
🔧 Gas-Shot: You approve every phase AND get a scoring loop that blocks bad output from shipping.
🔥 Ignition: You co-design the plan, then one approval fires off a fully autonomous, quality-gated build.
The Control Problem
Autonomous AI skills like One-Shot are fast. You send a prompt, walk away, come back to a finished product scored against a rubric.
But you never get to weigh in. If the AI misunderstands your intent at minute one, it spends the next hour perfecting the wrong thing.
On the other end, Gas gives you an approval gate after every single phase. You see everything, decide everything. But Gas has no quality scoring — the output is only as good as your review eye.
Think of it like this: One-Shot is hiring a contractor who builds the whole house and hands you the keys. Gas is being on-site every day telling the crew what to do next. Neither is wrong — but most people want something in between.
That's what Gas-Shot and Ignition are. Two different answers to the same question: how do I stay involved without babysitting, and still get quality-gated output?
The Family Tree
Both skills are splices — they inherit DNA from two parent skills and combine their strengths.
Gas (the chassis)
Collaborative ideation. 5-agent blind-spot review. Approval gates between every phase. Your vision drives everything.
One-Shot (the engine)
Ultrathink reasoning. Agent teams. Dimensional scoring. Fix-and-loop until every score passes threshold. Zero follow-up needed.
Gas-Shot bolts One-Shot's engine onto Gas's chassis and keeps the driver's seat. Ignition splits the car in two: you steer stage one, then the engine drives stage two on its own.
Gas-Shot: Full Control + Quality Scoring
Gas-Shot is for when you want to approve every phase and want a scoring loop that catches what your eyes miss. You're the director. The scoring engine is your quality department.
How it flows
↓
🛑 APPROVAL GATE
↓
🔍 Phase 1c: Deep Research (APIs, assets, dependencies)
↓
🛑 APPROVAL GATE
↓
👁 Phase 2: Blind-Spot Check (5 specialist agents review the plan)
↓
🛑 APPROVAL GATE
↓
🏗 Phase 3: Build (full implementation, no stubs)
↓
🛑 APPROVAL GATE
↓
🧪 Phase 4: Test + Score (code tests, 10-agent review, dimensional scoring)
↓
🛑 APPROVAL GATE — scores presented here
↓
🛡 Phase 5: Harden (security, performance, resilience)
↓
🛑 APPROVAL GATE
↓
✅ Phase 6–8: Verify, Visual Audit, Polish
↓
📊 Final Score — if below 0.85/0.92, loop back and fix
↓
📦 Delivery
Every red octagon is a hard stop. The skill literally will not move forward until you say "go." Silence is not consent.
What Gas-Shot adds over plain Gas
- Ultrathink — extended reasoning (deep analysis before every action, not just "read the file and edit it")
- Phase 1c (Research) — a dedicated deep-research step before the blind-spot review, with its own approval gate
- Dimensional scoring — after testing, every dimension (correctness, security, polish, etc.) gets a number. If any dimension scores below 0.85 or the overall composite drops below 0.92, the skill loops and fixes before you ever see the output
- Visual audit — if the deliverable has a UI, the skill screenshots it and checks visual quality
- Agent teams — uses the TeamCreate protocol (structured team hierarchy) instead of ad-hoc agent spawning
Core insight: Gas-Shot's scoring loop runs within gates, not across them. If the score is low, it fixes and re-scores before presenting results at the next approval gate. You only see output that's already passed quality checks.
Ignition: Plan Together, Build Autonomously
Ignition asks a different question: what if you only need control over the plan, not the execution?
Stage 1 is collaborative. You and the AI brainstorm, research, and have five specialist agents poke holes in your plan. Then you hit one button — the Ignition Point — and stage 2 fires autonomously through build, test, harden, document, verify, visual audit, polish, and scoring. No more gates. No more approvals. Just output.
The two-stage architecture
📋 Phase 1: Ideation (collaborative, no files)
↓
🛑 APPROVAL GATE
↓
🔍 Phase 1c: Deep Research
↓
🛑 APPROVAL GATE
↓
👁 Phase 2: Blind-Spot Check (5 agents)
↓
🛑 IGNITION POINT — last gate. After this, hands off.
=== STAGE 2: AI DRIVES ===
🏗 Phase 3: Build
↓
🧪 Phase 4: Test + 10-Agent Review
↓
🛡 Phase 5: Harden
↓
📝 Phase 6: Document
↓
✅ Phase 7: Verify
↓
📸 Phase 8: Visual Audit (if UI)
↓
✨ Phase 9: Polish
↓
📊 Score — if below threshold, autonomous fix loop
↓
📦 Delivery
Stage 2 has nine phases, no gates, and a built-in scoring loop. If the composite score falls below 0.92, the skill autonomously loops back through weak phases, fixes them, and re-scores. You don't intervene. You don't even see the loop happening.
Think of it like this: Ignition is briefing an architect on exactly what house you want — reviewing floor plans, debating materials, signing off on the blueprint — then letting the construction crew build it without you standing on-site every day. The crew has its own quality inspector (the scoring engine) who won't let anything ship that doesn't pass.
Side-by-Side Comparison
Here's the full breakdown of what you get with each approach.
| Dimension | Gas-Shot | Ignition |
|---|---|---|
| User control | High — approval gate at every phase | Planning only — 1 gate at Ignition Point |
| Approval gates | 6+ gates throughout | 3 gates (all in Stage 1) |
| Total phases | 8 | 9 (Stage 2 adds Document as separate phase) |
| Quality scoring | Yes — loops within gates | Yes — loops autonomously in Stage 2 |
| Blind-spot review | 5 agents, gated | 5 agents, gated (Stage 1) |
| Ultrathink | Always on | Always on |
| Agent teams | TeamCreate protocol | TeamCreate protocol |
| Visual audit | Yes (if UI) | Yes (if UI) |
| Execution style | User-steered + quality loop | Two-stage hybrid |
| After build starts | You still approve every phase | Fully autonomous |
Same engine, different cockpits. Here are the receipts you walk away with on the same prompt across all four modes — representative numbers from internal builds, not benchmarks.
When to Use Which
Gas-Shot
You want to see and approve the build at every stage. The project is high-stakes or unfamiliar territory. You want to course-correct mid-build. You care about the journey, not just the destination.
Ignition
You trust your plan but don't want to babysit execution. The build is well-defined after planning. You'd rather review one finished output than six intermediate checkpoints. You want speed without sacrificing quality.
Neither is "better." They're different answers to the same question: where do you want the boundary between your control and the AI's autonomy?
The Full Spectrum
Gas-Shot and Ignition sit between two extremes. Here's the full control spectrum of Godmode execution skills.
↓
🔧 Gas-Shot — max control + scoring loop
↓
🔥 Ignition — controlled planning, autonomous execution
↓
🎯 One-Shot — fully autonomous, quality-gated
Moving down the stack, you trade user control for speed and autonomy. Moving up, you trade speed for visibility and decision-making power.
Core insight: The scoring engine (dimensional scoring, 0.85/0.92 thresholds, fix-and-loop) is identical across Gas-Shot, Ignition, and One-Shot. The only difference is who decides when to proceed — you, a shared handoff, or the AI alone.
Under the Hood: Shared DNA
Both skills share the same quality infrastructure inherited from One-Shot. Here's what runs identically in both.
The scoring engine
After testing, every build gets scored across eight dimensions. Here's what the rubric looks like for code tasks:
| Dimension | Weight | What it measures |
|---|---|---|
| Context Load | 10% | Did the skill read the files it needed to? |
| Execution Completeness | 20% | Are there any stubs, TODOs, or placeholders? |
| Test Coverage | 20% | Branch coverage percentage |
| Security | 15% | OWASP findings resolved |
| Alternatives Explored | 5% | Did it consider other approaches? |
| Documentation | 10% | Are exports documented? |
| Ripple Integrity | 10% | Does the full test suite still pass? |
| Polish | 10% | Lint issues per lines changed |
Every dimension must hit 0.85. The weighted composite (overall score) must hit 0.92. Below either threshold, the skill loops: fix the weak dimensions, re-score everything, repeat until clean.
The blind-spot review
Before any code gets written, five specialist agents review your plan independently. They're looking for things a single perspective would miss — security gaps, missing edge cases, architectural issues, scope creep.
Both Gas-Shot and Ignition run this in Stage 1 / Phase 2, behind an approval gate. You see every agent's findings and decide what to incorporate before building starts.
How to Use Them
Both are Claude Code slash commands. Type them in your terminal and follow the prompts.
# Gas-Shot: collaborative build with scoring
/gas-shot build a dashboard that shows real-time sales data
# Ignition: plan together, then autonomous build
/ignition build a dashboard that shows real-time sales dataSame prompt, different experience. Gas-Shot will stop and ask you six-plus times during the build. Ignition will stop three times during planning, then build the whole thing without interrupting you.
Build with control and quality
Gas-Shot, Ignition, One-Shot, and 20+ other execution skills. All included.
Get Godmode What are skills? →