AI and LLMs in DevRel | DevRel Almanac

For fuller treatment, see Section 16: DevRel in the AI Era: a dedicated section covering the dual-audience thesis, LLM-mediated discovery, llms.txt, MCP as DevRel surface, Agent Experience (AX), GEO/AEO, vibe coding, metrics, perspectives and debates, critical takes, and early signals of what works. This page remains as a higher-level trend summary.

From late 2022 through 2026, AI changed how developers discover products, learn technologies, evaluate options, consume documentation, and write code. DevRel teams have had to adjust their public surfaces and measures to account for those behaviours.

The shift in three observations

LLMs are now a developer-discovery channel. Many developers research products through ChatGPT, Claude, or Perplexity before they search Google. Products described inaccurately or incompletely in those answers may not enter the developer’s evaluation.
AI tools consume documentation. Copilot, Cursor, Claude Code, and other coding tools use documentation while generating code. Missing imports, ambiguous versioning, and incomplete examples can therefore reach a developer through generated output.
Developers increasingly delegate implementation to agents. Developers review, redirect, and approve code generated by AI tools, so DevRel material must support both the person choosing an approach and the agent carrying it out.

Implications for DevRel work

Documentation must serve AI agents

Practical implications:

Structured Markdown. Clear heading hierarchy, consistent formatting.
Self-contained code examples. All imports, all setup, runnable as-is.
Explicit versioning. AI agents need to know what’s current.
Avoid burying key info in images. Without alt text, AI doesn’t see it.
Canonical URLs. AI agents store links; stable URLs survive.
llms.txt or equivalent. A growing convention: companies publish a top-level llms.txt indexing AI-readable canonical pages, modelled loosely on robots.txt. Anthropic, Mintlify, and others have implemented and promoted this pattern.
Documentation site maps optimised for AI ingestion as well as for human navigation.

AI tools retrieve sample apps

A public sample-app repository can provide:

Public code that may enter future training datasets.
Reference material that current AI assistants can retrieve.
A tested implementation that assistants can cite or adapt.

Many AI-era DevRel teams now design sample apps for AI ingestion. They publish more examples with a smaller scope, clearer code style, and exhaustive README files.

Model Context Protocol (MCP)

Launched by Anthropic in November 2024, MCP is an open standard for connecting AI systems to external applications and data sources.

Key data points:

From ~100 servers in November 2024 to 17,000+ by early 2026.
Cross-vendor support: OpenAI, Google, and Anthropic all support MCP.
For developer products, an MCP server is a new DevRel surface: exposing the product’s capabilities directly to AI agents.

Teams publishing an MCP server need to decide which operations it exposes, how permissions work, and how an agent can recover from errors.

Discovery surfaces have multiplied

A DevRel team in 2022 had to optimise for Google, Twitter, conferences, and a few content sites. A DevRel team in 2026 must additionally optimise for:

ChatGPT, Claude, Perplexity (LLM-mediated search).
AI-powered IDE assistants (Cursor, Copilot, Claude Code).
Coding agents (Devin, OpenAI’s agents, Claude Code agents).
AI-driven research workflows.

Different surfaces favour different content. Long-form authoritative writing increases the likelihood of being cited by AI assistants; short social posts that worked on X have minimal AI-citation value.

Sentiment in AI assistants matters

Teams can audit how ChatGPT, Claude, and other assistants describe their product, including technical errors, omitted constraints, and the sources cited.

Mechanisms by which AI assistants form views:

Training data. Long-form authoritative content (your docs, your blog, third-party reviews).
Real-time retrieval. What AI search surfaces for your product.
User feedback loops. Some assistants reflect what users have asked recently.

If the AI is wrong about your product, the fix usually involves clearer authoritative content the AI can ingest, plus engagement with the platform if available.

Implications for content strategy

Content type	Pre-2023 emphasis	2024 to 2026 emphasis
Reference docs	Human readability	Human readability + AI ingestion
Quickstarts	TTFHW for human	TTFHW for human + AI completing tasks
Code samples	Educational	Educational + AI training data
Blog posts	SEO	SEO + AI citation likelihood
Conference talks	In-room impact + replay	In-room impact + replay + AI-searchable transcripts
Newsletter	Subscriber relationship	Same
Sample apps	Demonstrative	Demonstrative + AI agent reference

Implications for developer education

The role of developer educators is shifting:

Less “how to use the SDK basics” (AI assistants now handle these explanations).
More “how to architect a working system” (AI is weaker at strategic decisions).
More “how to evaluate trade-offs” (AI hallucinates with confidence; the educator’s job is to teach the developer to detect this).
More content about working with AI tools to build with your product: meta-content about the new development workflow.

Implications for community

Real-time community help is partly replaced by AI assistants. Many questions that used to flood Discord channels now go to ChatGPT first.
Communities are becoming places for harder questions that AI couldn’t answer.
Expert community members handle harder questions. Developers still need accountable people for ambiguous bugs, disputed recommendations, and product decisions.

The smaller communities can be more focused, and expert contributions receive more attention.

Implications for advocacy

Advocates’ content has new audiences. Every video, blog, and tweet is now potential training data and citation source.
Authorship remains useful evidence. Named experts give readers a public record against which to judge a claim, while clear source attribution lets them inspect material surfaced by retrieval systems.
Real-time presence matters less. A great podcast appearance is more durable than a tweet thread.
Open-source repositories provide retrievable evidence. AI agents can inspect GitHub code, documentation, and issue discussions.

Tools and agents in the DevRel workflow itself

DevRel teams use AI tools internally:

Content generation assistance. Draft posts, edit, summarise.
Community sentiment analysis at scale. Tools like Common Room layer AI on top of conversation data.
Personalised outreach. Identifying and engaging individual community members.
Doc gap detection. Asking AI to identify questions docs don’t answer.
Translation. Producing localised content efficiently.

AI-engineering as a DevRel category

Beyond AI’s effect on DevRel practice, AI has produced a substantial new category of developer products requiring DevRel:

Foundation-model providers (OpenAI, Anthropic, Google, Cohere, Mistral, etc.).
Open-source model hubs (Hugging Face).
AI orchestration tools (LangChain, LlamaIndex, Haystack).
Vector databases (Pinecone, Weaviate, Chroma, Qdrant).
AI infrastructure (Modal, Together AI, Replicate, RunPod).
Coding agents and tools (Cursor, Aider, Continue, GitHub Copilot, Claude Code).

These companies built DevRel functions over 2023 to 2026 from scratch. The category employs a substantial fraction of the laid-off DevRel professionals from the 2022 to 2024 contraction and has effectively become a new sub-field with its own conferences (AI Engineer Summit), publications (Latent Space), and senior practitioners (Logan Kilpatrick, Romain Huet, swyx, Harrison Chase, Jerry Liu).

See AI and ML companies.

What gets harder

Attribution. AI-mediated discovery makes attribution to specific DevRel activities harder.
Voice. AI-assisted content tends to converge on average prose; distinctive voice is harder to maintain at speed.
Original research. Long-form posts compete with AI-generated competitive content; the bar is higher.
Real-time community help. Routine questions may go to AI first, leaving community teams with a higher share of ambiguous bugs, disputed recommendations, and product-specific decisions.

What gets easier

Content production volume. A team can produce more drafts faster.
Personalisation. Tailoring content to specific developer segments.
Multi-language reach. Translation and localisation are cheaper.
Analytics. Sentiment, gap analysis, and pattern detection are more tractable.

The agentic-experience (AX) question

From 2024 to 2026, some companies began treating agent experience—how agents interact with their APIs, documentation, and tools—as a distinct part of developer experience.

This shows up in:

API design that is easy for agents to use (predictable, well-typed, well-documented).
Error messages designed for both human and AI consumption.
Authentication flows that work well for agentic access.
Rate limits and quotas designed with agent patterns in mind.

For DevRel teams at AI-adjacent infrastructure companies, AX is becoming as deliberate a function as DX.