ericdodds

Mostly a writer these days. I work in technical marketing @Vercel.

[ericdodds @ericdodds]

No matter the process, you can tell when someone isn’t doing the work.

But it isn’t as simple as “AI slop or not.” There’s a spectrum of fully generated to assisted to “artisanal, homemade sentences.” If it’s generated and untouched, you notice quickly. Assisted pieces usually feel incongruent if the author phones it in on edits.

There are some interesting ideas around evals for writing, which is highly subjective, but I have no idea how we could create universally applicable benchmarks (current tools probably pattern match and use synthetic examples).

[Rhys @RhysSullivan]

when you're reading a blog post, are you able to tell when it's ai generated? do you care?

i generally can tell within the first few paragraphs and it puts me off of reading it / overall what the content is about

[ericdodds @ericdodds]

Burning $100 in Astra tokens through the Davinci Resolve MCP is worth it just to see how capable this setup is for video editing. I was blown away.

I'm using OpenCode, AI Gateway, and Warp for the MCP interface.

[ericdodds @ericdodds]

"Go through this session (including pre-compaction) and print your response to every prompt where I corrected you, pushed back on your output, or replaced your draft with my own rewrite."

[ericdodds @ericdodds]

@reed_barnes @lgrammel Thanks. We are already working on deep dives for the individual agents that explain the mechanics and stack.

[ericdodds @ericdodds]

@tryleadpilot @vercel 75% fixes, 13% docs, 11% features. Features averaged ~450 lines of code. You can see every merged PR authored by the factory here: https://t.co/yOG0aCai6i

[ericdodds @ericdodds]

In my last job I moved into product because I love working with engineers. They possess a unique combination of creativity and clarity.

Working with @lgrammel on this piece was a true joy. It's an incredible system, built with an incredible amount of thought and care.

[Vercel @vercel]

We built a software factory for AI SDK.

Each step is an agent, and humans merge changes. Four weeks in:

▪️ The factory authors up to 35% of merged PRs
▪️ It closed 70% of issues in July
▪️ Open bugs are down 25%

https://t.co/UNGeeQJdAu

[ericdodds @ericdodds]

@mboyanna @rauchg @adiaddxyz @shardara Cooking on the post!

[ericdodds @ericdodds]

@infinterenders Let the man render

[ericdodds @ericdodds]

@brandon_galang You can learn quickly by asking models good questions, but collaborating with a really good human engineer is categorically different than chatting with Fable.

[ericdodds @ericdodds]

I run a content agent in a local harness, and if you don't pay attention to how your terminal client handles shell sessions, things can get wonky fast.

Coding harnesses are great at sorting things out when you back everything with git, but I agree worktrees are a great way keep project workflows isolated.

[brandon galang @brandon_galang]

day 9 of overthinking your agentic engineering setup

unless you want to be a solo vibecoder forever, learn to use worktrees to create distinct lanes for your agents to work within

before I start any new feature work, I setup a new worktree (your agent can do this itself too). nothing else changes

I never experience agents stepping on each other's work or mixing PRs that are ready to ship with work that's still in progress

[ericdodds @ericdodds]

Not too long ago I thought to myself, "It seems crazy that there isn't a cross-harness plugin spec." Just a few weeks later I was tapped in to work on the Agent Plugins launch with @hefnerdotpro.

This was a fun one, and a great picture of what it's like to work @vercel: shortly after you feel friction, product ships to remove it.

[Vercel @vercel]

Introducing Agent Plugins, an open standard for extending agents.

Supports Agent Skills and MCP, with more to come.

Built in collaboration with: @awsdevelopers, @code, @cursor_ai, @github, and @openaidevs.

https://t.co/duFCQqigZD

[ericdodds @ericdodds]

Justin is a talented writer. Every talented writer I work with who uses AI as part of their process:

1. Understands its power and
2. Is constantly frustrated and underwhelmed by it

AI for code can feel like magic, but you can't write tests for a great blog post in the same way you can for a feature. More importantly, in writing, words are the final product, where code is a means to a product end, and that's one of the reasons writers feel these sharp edges more acutely.

[sisyphus bar and grill @barsisyphus]

it has become legitimately difficult to use Claude b/c of the cringe. the writing style is so uncomfortable, so sycophantic, so unreasonably confident that I am actually struggling to keep going

not to mention the obvious holes in the product experience

make it joyful again!

[ericdodds @ericdodds]

AI is capable of some wildly complex tasks, but I often find myself most delighted in the simple things it can do, like cleaning up my desktop.

[ericdodds @ericdodds]

"The most important skill in prompting is expertise in the domain you’re prompting for."

@sjgoedecke is right, and this is why even though many perceive AI as "good at writing," it doesn't make people good writers. https://t.co/4mymfjdgRn

[ericdodds @ericdodds]

@barrald @itunpredictable The training on words was reductive from the beginning. Now that the generative novelty has worn off, it’s clear the expectations for writing quality were far too grandiose. But good for the people who made money while people still believed AI would make cold outbound better.

[ericdodds @ericdodds]

@shardara @codewithdavid @rauchg @adiaddxyz Blog post initiated

[ericdodds @ericdodds]

@joshgonsalves_ @vercel There are two guides and a handful of templates in our knowledge base. The content agent is a good example of eve implementing every primitive: https://t.co/f2boKVandY

[ericdodds @ericdodds]

@brandon_galang For real. How are we even supposed to do our jobs?

[ericdodds @ericdodds]

@KostasPardalis True to your philosophical heritage, you cut to the heart of the matter: what is slop?

[ericdodds @ericdodds]

No matter the process, you can tell when someone isn’t doing the work.

But it isn’t as simple as “AI slop or not.” There’s a spectrum of fully generated to assisted to “artisanal, homemade sentences.” If it’s generated and untouched, you notice quickly. Assisted pieces usually feel incongruent if the author phones it in on edits.

There are some interesting ideas around evals for writing, which is highly subjective, but I have no idea how we could create universally applicable benchmarks (current tools probably pattern match and use synthetic examples).

[Rhys @RhysSullivan]

when you're reading a blog post, are you able to tell when it's ai generated? do you care?

i generally can tell within the first few paragraphs and it puts me off of reading it / overall what the content is about

[ericdodds @ericdodds]

@KostasPardalis That’s the question. I see a lot of hand-wringing about AI slop in writing and some of it is obvious: long-tail SEO content and BDR outreach (these were always bad just amplified and more uniform), but vibes seems to be the primary eval.