One person, hundred-person-team engineering judgment.
7 principles extracted from production-grade systems (Claude Code v2.1.88) and generalized into a thinking framework that applies to any product domain — SaaS, CLI/DevTools, Agent orchestration, AIGC, e-commerce, content platforms, and beyond.
"We don't just incubate unicorns — we build the engine that continuously fissions new ones." — Nebutra Manifesto
A Claude Code SKILL that upgrades engineering decision-making. Not a tutorial. Not a checklist. A thinking operating system.
The gap between "can write code" and "can ship a unicorn-grade product" isn't technical — it's cognitive. This SKILL closes that gap by encoding the thinking patterns behind $1B products into 7 actionable principles.
| # | Principle | One-liner |
|---|---|---|
| 1 | Minimum Intervention | Lightest fix first — escalate only when cheaper options fail |
| 2 | Boundary Is Product | Every line you draw shapes user experience |
| 3 | Lifecycle, Not Function | Design how things are born, live, and die — before writing code |
| 4 | Earn Trust Progressively | Never demand all trust upfront |
| 5 | Constraint As Fuel | Limitations are design starting points, not excuses |
| 6 | Policy In Code, Not Wiki | Rules in wikis get broken; rules in compilers don't |
| 7 | Soul Before Scale | Product personality is foundation, not decoration |
# Add as a git submodule to your project
git submodule add https://github.com/Nebutra/claude-code-harness.git .claude/skills/claude-code-harness
# Or copy the SKILL.md into your skills directory
cp claude-code-harness/SKILL.md ~/.claude/skills/claude-code-harness/SKILL.mdEach principle has a deep-dive reference document:
- 01 — Minimum Intervention
- 02 — Boundary Is Product
- 03 — Lifecycle, Not Function
- 04 — Earn Trust Progressively
- 05 — Constraint As Fuel
- 06 — Policy In Code, Not Wiki
- 07 — Soul Before Scale
Each reference covers: what the principle says, why it matters, how Claude Code implements it, how to migrate it to YOUR product, anti-patterns, and self-check questions.
Supplementary deep-dives (for specialized domains):
- Cache-Aware Architecture — Prompt cache engineering, prefix sharing, cache-first design
- Multi-Agent Coordination — 3-layer hierarchy, permission propagation, communication patterns
- Streaming and Real-Time — AsyncGenerator, backpressure, progressive rendering
Blueprints (step-by-step decision checklists):
- Agent Product Blueprint — 3 Agent invariants, 7 design decisions, phased build path, production pitfalls
- Self-Assessment Checklist — 29-question audit across all principles + Agent-specific checks
| You're doing... | Focus on... |
|---|---|
| Building from 0 to 1 | Soul, Constraint, Boundary |
| Designing API / architecture | Boundary, Lifecycle, Policy |
| Performance / cost optimization | Minimum Intervention, Constraint |
| Permission / security design | Trust, Boundary, Policy |
| Building Agent / AI products | All, especially Intervention + Lifecycle |
| Identifying tech debt | All — use self-check questions as audit |
| Pre-launch review | Full pass through all 7 |
Tested with real prompts across 4 rounds, 6 subagents, 2 tasks (Agent architecture design + production codebase tech debt audit). Compared: no SKILL vs v1 SKILL vs v2 SKILL.
| Metric | No SKILL | v2 SKILL |
|---|---|---|
| Tokens | 23K | 33K (+43%) |
| Boundary reasoning | Lists component responsibilities | Justifies each boundary by semantic difference; rebuts 3 common wrong approaches |
| Error handling | Decision tree (classify → route) | 5-level intervention ladder (retry → replan → human → fail) |
| Type safety | Standard interfaces | Opaque BudgetToken forces compile-time budget check |
| Crash recovery | Checkpoint mentioned | Every state answers "what if process crashes here?" + two-phase commit |
| Tech recommendations | Specific (BullMQ, Postgres) | Specific + rationale per choice |
| Anti-patterns section | None | 5 anti-patterns with explanation of why each is tempting |
| Metric | No SKILL | v2 SKILL |
|---|---|---|
| Tokens / tool calls | 116K / 113 | 87K / 52 (-25% / -54%) |
| Unique findings | Zod version split, phantom schemas | S2S HMAC mismatch, mock payment path, hardcoded plan limit |
| Fix specificity | Directional recommendations | Inline code fixes, "fix today" vs "next sprint" triage |
| Insight quality | Describes what exists | "Half-completed good intentions are more dangerous than never starting — they create the illusion of safety" |
- Opaque type patterns that make wrong usage a compile error
- Intervention cost ladders for error handling and resource management
- Watchdog processes that don't trust the executor to self-terminate
- "Phantom reliability" diagnosis (code that looks production-ready but runs in-process memory)
- Bugs found through boundary/lifecycle lens that standard audit misses (HMAC auth actually broken)
- +43% token cost on architecture tasks — worth it for high-stakes decisions, overkill for CRUD
- Best on Sonnet/Opus — may degrade on smaller models
- Not a domain expert — improves architectural judgment, doesn't replace business knowledge
- Diminishing returns on simple tasks — writing a utility function doesn't need 7 principles
| Use | Skip |
|---|---|
| 0-to-1 architecture design | Simple bug fixes |
| Tech debt audit | Writing unit tests |
| Agent/AI product design | CRUD endpoints |
| System review before launch | Daily coding tasks |
| Upgrading demo to production | Documentation updates |
Nebutra Manifesto (philosophy) → "Infinite fission, not infinite expansion"
Unicorn Harness (cognition) → How to think ← this repo
Nebutra Sailor (infrastructure) → 53 modules built on these principles
Sleptons (community) → OPC ecosystem powered by this thinking
Migration examples from new domains are welcome. Each principle's reference file has a "Migrate to your product" section — if your domain isn't covered, submit a PR.
MIT — thinking frameworks should spread without friction.