For executives
Situation
Teams ship with LLMs. The speed is there — developers produce code faster than ever.
Complication
Without imposed discipline, AI produces plausible but unverified code. Test coverage is declarative, reviews are superficial, and technical debt accumulates silently.
« The only way to go fast is to go well. » — Martin, R. C., Clean Architecture, 2017.
Question
How do you frame AI so that delivery speed translates to delivered value?
Answer — The 3 SKRAFT levers
1. Enforceable discipline
Each pipeline phase applies the CQS principle: the executor agent writes, the reviewer verifies. No agent validates its own work. Invariants are checked automatically before every phase transition.
2. Empirical quality metrics
SKRAFT measures Mutation Score, not declarative coverage. A test that detects no mutation is a test that protects nothing.
« Software delivery performance predicts organizational performance and profitability. » — Forsgren, N., Humble, J. & Kim, G., Accelerate, 2018.
3. Speed without hidden debt
The Walking Skeleton validates end-to-end architecture before writing any business logic. Outside-In TDD ensures every line of code is guided by an acceptance test.
« High performers spend less time on unplanned work and rework, freeing capacity for new value. » — Forsgren, N., Humble, J. & Kim, G., Accelerate, 2018.
Measurable ROI — DORA metrics
| Metric | Without pipeline | With SKRAFT |
|---|---|---|
| Change Failure Rate | High — bugs found in production | Reduced — validated by Mutation Score before merge |
| MTTR | Slow — manual diagnosis | Fast — traceability from issue → test → code |
| Deployment Frequency | Slowed by fear of breaking things | Accelerated — every commit is verified |
« If you only quantify one thing, quantify the cost of delay. » — Reinertsen, D. G., The Principles of Product Development Flow, 2009.
Current limits and roadmap
The DORA metrics table above is a qualitative projection — not a benchmark measured on SKRAFT itself. The causal chain is grounded in established literature (Forsgren et al., Wiegers), but no controlled study has yet measured SKRAFT’s impact on a production team. Treat those figures as direction, not guarantee.
Active guardrails (G1, G4, G5, G7, G8)
Five guardrails are operational today. They enforce dispatch order, artifact structure, reviewer verdicts, and direct state-file writes at the runtime level — before the tool executes. These are mechanical, not advisory.
Planned guardrails (G2, G3) and infrastructure (US6, US13)
Two guardrails are not yet active:
- G2 (skill injection) and G3 (skill audit) — a sub-agent currently starts without a guaranteed skill set. The expected constraints may not be injected, which is a real blind spot for hallucination of method.
- Boundary-to-boundary tests (US6) — guardrails are unit-tested but not tested end-to-end; a silent regression remains possible.
- Recovery/rollback (US13) — if state is corrupted, the pipeline may stall with no clean exit path.
Next measurable ROI lever
The largest unactivated cost lever is the out-of-LLM verdict schema: today the model formats reviewer verdicts, which represents a non-trivial token tax on reviewer output. When implemented, this lever will produce a quantifiable reduction in reviewer spend.
What SKRAFT requires
- 2–3 days of training for the team on methodology and agents.
- An identified sponsor who drives adoption and arbitrates resistance.
- Respect for invariants — the pipeline guardrails are non-negotiable (see §5.3).