James Brady

I lead the engineering team @elicitorg.

Join us, we're nice and build cool stuff:

[James Brady @james_elicit]

@shauryr @0x_rody 👋

[James Brady @james_elicit]

@DamonDanieli @0x_rody 👋

[James Brady @james_elicit]

Full post: https://t.co/WIjfl3XF6i

[James Brady @james_elicit]

Two years ago we published our stance on AI in engineering interviews. Core question: does the candidate understand the code?

That question still holds—but it's no longer where we gather our most interesting signal.

[James Brady @james_elicit]

Worst pattern: yolo --dangerously-skip-permissions, accept every diff without reviewing. The code kind-of-works but tells us nothing we couldn't learn from the model itself.

Surprise finding: plan-first thinking is way rarer than expected. Most candidates jump straight to "make the model write code."

[James Brady @james_elicit]

Same pattern as alarm fatigue in hospitals—when monitors beep constantly, nurses learn to ignore them. The safety system becomes the danger.

Auto mode could be the right trade-off: let the model assess risk rather than prompting on everything. Cautiously hopeful!

[James Brady @james_elicit]

Anthropic just announced auto mode for Claude Code permissions—I am unreasonably excited about this!

For the last few weeks, Claude Code's permission system has been quietly training us to be **less** safe.

[James Brady @james_elicit]

What the team actually did:
• Attempt to preempt the permission model—custom allow/deny lists, hooks to block destructive commands, mining past transcripts for things we'd been approving on autopilot
• Switched to pi ($1K in API tokens, still better than the prompt loop)
• Moved work to remote agents (Devin, Niteshift) partly just to avoid the babysitting
• Probably ran with --dangerously-skip-permissions without telling me, tbh

The safety mechanism was producing the opposite of safety.

[James Brady @james_elicit]

@ntkris Getting increasingly hard to find (although obvs not impossible) to find things which don't fit the latter two categories – at least in white-collar settinga

[James Brady @james_elicit]

We're automating @elicitorg. When everyone goes on holiday in December, the company keeps running. Bugs triaged. Features shipped. Metrics dutifully updated by robots who don't know it's Christmas.

Step one: figure out what to automate. I set up a Notion database, asked people to decompose their work into triggers, steps, outputs.

Turns out this was the wrong question…

[Elicit @elicitorg]

The Elicit API is now available in preview for Pro and Teams users. You can search 138M+ papers and generate Research Reports from your code, scripts, or AI tools.

Get your API key at https://t.co/Q3r5PggChI and check out https://t.co/yMxHqkK0Ob

[James Brady @james_elicit]

@VivaLaPanda @voooooogel *First* token was already bork for me!

[James Brady @james_elicit]

@jerhadf I sent message IDs etc. to Brittney T

[James Brady @james_elicit]

Is anybody else getting absolutely bonkers hallucinations from Claude!?

I just tried to check a couple of things off my todo list 😅

[James Brady @james_elicit]

@arm1st1ce Full transcript https://t.co/QWsRL40Cj9

[James Brady @james_elicit]

@ntkris We get a lot of value from spending time with customers to see what's *actually* slowing them down – pretty hard to shadow yourself though!

… or is it easy? I don't know at this poitn

[James Brady @james_elicit]

We're hiring for other people who don't want to do boring stuff!

DM me or check out https://t.co/Q9I5XDyqu2

[James Brady @james_elicit]

We're running a retreat next week to automate as many of these rote tasks as we can. Demos on Friday. More on what we learn soon.

Starting lesson: don't ask people to map their jobs. Ask what they'd love to never do again.

What's the most tedious part of your job?