Ty Dunn's Context Engineering for AI Coding

Table of content

Ty Dunn is the co-founder and CEO of Continue, the leading open-source AI code assistant with 23,000+ GitHub stars and 11,000+ Discord community members. His mission: amplify developers, not automate them. He coined “context engineering” as the discipline that separates productive AI coding from frustrating prompting.

Twitter | GitHub | Blog | LinkedIn

Background

BS in Cognitive Science (Computer Science) from University of Michigan, 2015-2019
Built dialogue management systems while studying language and computation
First product manager at Rasa, grew to group PM
Rasa’s open-source conversational AI framework reached millions of downloads and 17,000 GitHub stars
Founded Continue in June 2023 with Nate Sesti
Y Combinator Summer 2023

Context Engineering

Dunn identifies a core problem: AI coding assistants fail because of bad context, not bad models.

From his Context Engineering blog post:

“Bad context is worse than no context.”

When irrelevant or outdated information reaches the model, you get context poisoning. The AI becomes overconfident in incorrect information and propagates errors through your codebase.

Problem	Cause
Context crisis	Knowledge lives in Zoom calls and hallway conversations
Scattered docs	Confluence, Jira, Slack, email have no unified access
Wrong structure	Documentation optimized for human reading, not AI consumption
Silent degradation	Information rots without visible breakage like code

The Three Data Gaps

Dunn argues that source code repositories miss critical information for AI assistance:

Process data - The step-by-step flow developers take to complete tasks
Context data - What information informed each decision point
Reasoning data - Natural language explaining the “why”

His solution: collect data from developer-LLM interactions. When developers accept or reject suggestions, edit completions, or iterate on prompts, they generate implicit feedback signals.

Continue’s Approach

Continue installs as a VS Code or JetBrains extension, connecting to any LLM provider:

{
  "models": [
    {
      "title": "Claude 3.5 Sonnet",
      "provider": "anthropic",
      "model": "claude-3-5-sonnet-20241022"
    }
  ],
  "rules": [
    {
      "slug": "typescript-conventions",
      "rule": "Use strict TypeScript. Prefer interfaces over types."
    }
  ]
}

Key differentiators:

Rules encode standards that apply automatically
Context providers connect to Linear, Jira, incident systems
Development data collection measures AI effectiveness
Open-source core with Apache 2.0 license

Rules Files

Continue supports .continue/rules/ directories for project-specific context. From Dunn’s async context engineering article:

# verify-changes.md
When modifying .js or .ts files, always verify:
1. Import paths resolve correctly
2. Type definitions match usage
3. Tests pass after changes

These rules trigger automatically based on file type, catching errors like hallucinated import paths before commits.

Task Decomposition

Dunn describes a progression from supervised to autonomous AI assistance:

Stage	Trust Level	Example
Supervised	Low	Accept/reject each suggestion
Assisted	Medium	AI drafts, human reviews
Autonomous	High	AI works independently, human verifies

The pattern that emerges: repetitive tasks become obvious candidates for automation. Trust builds gradually through repetition.

Async Context Engineering

For developers working in fragmented sessions (Dunn mentions parenting as a common interrupt source), the workflow becomes:

Set context once - Load architecture docs, rules, examples
Let AI work - 15-minute autonomous bursts
Verify on return - Automated quality gates catch errors

The rules act as persistent guardrails that don’t require re-prompting each session.

Key Takeaways

Principle	Implementation
Engineer context, don’t prompt	Encode standards as rules, not ad-hoc instructions
Bad context poisons results	Filter irrelevant information aggressively
Collect development data	Track LLM interactions to measure effectiveness
Build sustainable processes	Treat context engineering as ongoing, not one-time
Trust builds through repetition	Start supervised, move to autonomous

Links

Next: Jesse Vincent’s Superpowers Framework