The nv:context Manifesto

April 7, 2026 · View on GitHub

The Problem

Most AI coding agents fail. Not because the models are dumb. They're remarkably capable. They fail because we feed them garbage context.

"Most agent failures are not model failures. They are context failures." Philipp Schmid, Google DeepMind

The evidence is damning:

  • 40%+ of AI project failures stem from poor context management (IntuitionLabs)
  • Auto-generated config files reduce success by 3% and increase costs 20%+ (ETH Zurich)
  • A focused 300-token context outperforms an unfocused 113K-token context (FlowHunt/LongMemEval)
  • Using 40% of context window outperforms using 90% (Dex Horthy)
  • Experienced developers are 19% slower with AI tools when context is poorly managed (METR)

That last one bears repeating: bad context engineering doesn't just not help. It actively makes you slower.

The Shift

Prompt engineering is dead. The "prompt whisperer" era, where clever phrasing was the skill, is over.

What matters now is context engineering: the systematic discipline of designing the information ecosystem around your AI agents so they have exactly what they need and nothing more.

"If prompt engineering was about coming up with a magical sentence, context engineering is about writing the full screenplay for the AI." Addy Osmani, Google Chrome

The best engineers don't write better prompts. They build better systems:

  • Minimal config files that document landmines, not maps
  • Hooks that enforce rules deterministically, not probabilistically
  • Progressive disclosure that loads context on-demand, not all-at-once
  • Session management that prevents context rot
  • Compounding feedback loops that make the system smarter over time

The Hierarchy of Leverage

Not all context engineering is equal. Here's what moves the needle, in order:

1. Verification (tests, linters, CI)     ← highest leverage
2. CLAUDE.md / AGENTS.md quality
3. Hooks (deterministic enforcement)
4. Skills (on-demand context)
5. Subagent patterns (context isolation)
6. Session management (compaction, handoffs)
7. Worktrees (parallel isolation)          ← lowest leverage

Most people optimize from the bottom up. The best engineers start at the top.

The 8 Laws

  1. Less is more. Every line in your config competes with the actual task for attention.
  2. Landmines, not maps. Document what agents can't discover. Delete everything else.
  3. Commands beat prose. One snippet outperforms three paragraphs.
  4. Context is finite. 150-200 instructions max. Budget them wisely.
  5. Progressive disclosure. Layer your context: always → on-reference → on-demand → runtime.
  6. Hooks for determinism. If a rule MUST be followed 100% of the time, use a hook.
  7. Negative instructions backfire. "Don't use X" increases X. Say "MUST use Y."
  8. Compact proactively. Don't wait for 95%. Manage your context like you manage your memory.

The Standard

nv:context is the implementation of these principles. It analyzes your repository, interviews you about your landmines and workflows, and generates the complete context engineering infrastructure (config files, hooks, session management, token budgets) so every AI agent that touches your code works at maximum effectiveness.

It's not a prompt template. It's not a config generator. It's a context engineering system built on 200+ research sources from the people who build and study AI agents for a living.

For engineers who ship.