Launching this week

Context engineering for
engineers who ship.

Most agent failures aren't model failures.
They're context failures.

~/your-repo
$npx skills add johnnichev/nv-context -g -y
See the case study →
440→67
selectools CLAUDE.md
−85% lines
58/60
leverage score
up from 49/60
15.8K
tokens saved per session
nichevlabs
Production Proof

Real repos. Real numbers. No demos.

Three production codebases. Same skill applied. Here's what changed.

selectools Most detailed
Python SDK · 4,612 tests
Maturity L3→L5-L6
Leverage 49→58 /60
CLAUDE.md 440 → 67 (−85%)
Token budget −53%
Read the full report
nichevlabs
Multi-product SaaS · orchestrated by selectools
Maturity L4→L6
Leverage 17→49 /60
SESSION.md 805 → 59 (−93%)
Tokens saved / session 15,800
Bugs found 81
Read the full report
sheriff
Python + TypeScript
Maturity L4→L5
Leverage 36→42 /60
Strategy Incremental polish
Already-strong setup Coverage gate
Read the full report
The Problem

Bad context doesn't just not help.
It actively hurts.

Most agent failures aren't model failures.
Five things the research keeps confirming:

3%
Auto-generated configs reduce agent success ETH Zurich
19%
Slower performance for experienced devs with bad AI context METR controlled study
40%+
Of AI project failures stem from poor context IntuitionLabs
40%
Of context window outperforms 90% (focus beats volume) Dex Horthy
300
Focused tokens beats 113K unfocused tokens FlowHunt
See all 200+ sources

We applied all of this to nv:context.

The 8 Laws

Eight rules,
one philosophy.

Distilled from 200+ sources and three production repos.

01

Less is more

Every line competes with the actual task for attention.

02

Landmines, not maps

Document what agents can't discover by reading code.

03

Commands beat prose

Snippets with full flags beat paragraphs of explanation.

04

Context is finite

150-200 instructions max. Beyond that, attention degrades.

05

Progressive disclosure

Load context layer by need. Not everything, not at once.

06

Hooks for determinism

Critical rules need 100% compliance. Use hooks, not hope.

07

Negative instructions backfire

Say "MUST do X", not "don't do Y". Negation is fragile.

08

Compact proactively

60% safe. 70% precision drops. 85% hallucinations begin.

What Gets Generated

Nine files. Every AI tool.
One workflow. Zero magic.

One command. Configs for every AI coding tool you already use.

Works with Claude Code Cursor Copilot Aider Codeium Continue Windsurf Zed Replit Agent Bolt.new v0 Sweep Cody Tabnine Void Cline Roo Cline Kodu Refact Warp AI Codefuse Amazon Q JetBrains AI Goose Pythagora
~/your-repo
your-repo/
├── AGENTS.md←Universal · read by 25+ tools
├── CLAUDE.md←@imports root
├── tests/
│ └── CLAUDE.md←Test scope
├── src/
│ └── CLAUDE.md←Source scope
├── HANDOFF.md←Session handoff
├── .claudeignore←Token budget
├── .claude/
│ └── settings.local.json←Hooks
└── .github/workflows/
└── learn-from-reviews.yml←Compounding

Each file has a job. No duplication.

Hierarchy of Leverage

Six layers.
One score.

Every repo gets graded on six dimensions. Here’s a sample diagnostic.

Verification
8/10
CLAUDE.md quality
8/10
Hooks
2/10← gap
Skills
8/10
Subagent patterns
0/10← gap
Session management
8/10
Overall 34 /60
Top recommendation Set up hooks → +14 points → 48/60 Highest-leverage move on this repo.
The Research

Not pop AI.
The actual research.

Primary sources only: academic studies, lab papers, and frontline practitioners.

200+ sources
8 distilled laws
471 KB distilled findings

Sources include

AI Labs & Companies
Anthropic Claude model & Agent SDK Google DeepMind Agent design patterns OpenAI Function calling & tool use Manus Autonomous agent architecture JetBrains IDE-integrated AI research GitHub Copilot deployment data
Academic
ETH Zürich Agent evaluation study METR Controlled developer study FlowHunt Context-window benchmarks
Practitioners
Boris Cherny Claude Code co-creator Dex Horthy Context engineering research
Browse the full research library

Stop fighting your AI agent.
Start engineering its context.

~/your-repo
$npx skills add johnnichev/nv-context -g -y

Or install all four nv: skills:

~/your-repo
$npx skills add johnnichev/nv-context -g -y $npx skills add johnnichev/nv-dev -g -y $npx skills add johnnichev/nv-ops -g -y $npx skills add johnnichev/nv-design -g -y