Charlie Graham

5x Founder now @ & Previously Founder/CEO Hawku, Shop It To Me and more. Living in the AI-Rena and having fun!

[Charlie Graham @imcharliegraham]

Companies are scrambling to build AI brains. Most are choosing where it lives by accident, before they understand how hard it will be to move. https://t.co/QJ4sT4Y7x5

[Charlie Graham @imcharliegraham]

@benln Yes please!

[Charlie Graham @imcharliegraham]

@gregisenberg Agreed! I just wrote about this last week!

https://t.co/hvfl2seqpB

[Charlie Graham @imcharliegraham]

@clairevo I kind of stopped on this one when I didn't know who it was and it was asking for my email access. I'm sure this one is probably fine this time but I wonder one day if someone will do a huge privacy breach rug pull.

[Charlie Graham @imcharliegraham]

The dark horse in the model company race is SpaceXAI/grok.

The combination of hardware expertise with their data centers and the cursor acquisition is super interesting.

Their models will be very smart with intuitive harnesses and super fast.

[Charlie Graham @imcharliegraham]

@kiwicopple This is pretty funny.

[Charlie Graham @imcharliegraham]

Apologies if I am not as responsive today.

My Claude Code weekly quota resets tonight and I realized I still have some tokens available.

[Charlie Graham @imcharliegraham]

@Spotonparts Thanks, @Spotonparts! I hope all is well!

[Charlie Graham @imcharliegraham]

Yesterday I fired an AI.

Before it left, I made it do an exit interview.

It totally worked, and it should be in everyone’s AI toolkit.

I was working on a project where an AI had gone down a bad path. I kept correcting it, but it kept trying to rescue its original approach instead of rethinking the task.

At some point, it had too much bad context. I “fired” it by starting a new chat session.

But I did not want to throw away everything that happened in the first session.

While working with that AI, I had clarified requirements, scope, constraints, and decisions that were not in my original prompt. A fresh AI would not know any of that.

So before leaving, I had the old AI write a handoff document for its replacement.

The handoff includes the requirements, constraints, decisions already made, what it learned, and the open questions. I also tell it not to include implementation details unless they capture an important learning.

Then I paste that handoff into a new session.

The new AI gets the useful context, but none of the attachment to the bad approach.

I have been doing this frequently now and it has worked so well that I sometimes start a “test” session on purpose. One AI helps me turn a fuzzy idea into better requirements. Then a fresh AI does the real work.

I've included an exit-interview prompt in the comments. Feel free to use it!

[Charlie Graham @imcharliegraham]

Prompt:

Write a self-contained handoff document in Markdown for a new AI session taking over this task.
Include:

* The full requirements and success criteria
* Relevant context and constraints
* Decisions already made, including why when known
* What you learned while working on the task
* What is complete, incomplete, blocked, or uncertain
* Open questions the next AI needs to resolve
* The best next step

Do not include a detailed implementation plan, code walkthrough, or speculative solution unless it captures an important learning the next AI needs.

*The handoff should give a new AI enough context to restart the work cleanly without seeing this conversation.

[Charlie Graham @imcharliegraham]

@joshelman I believe a lot of it is to put the winter holidays in the middle of the school year between 2 semesters which seems reasonable.

Also I'd rather have them outside in May than August.

[Charlie Graham @imcharliegraham]

@benln ➕1 yes please!

[Charlie Graham @imcharliegraham]

@tristanbob Are these real mtg cards or your own creation? Looks beautiful btw

[Charlie Graham @imcharliegraham]

@JJEnglert It makes things so much more fun right? 😄

[Charlie Graham @imcharliegraham]

Seeing more open source lately? It's because we are entering a golden age of open source applications. https://t.co/itbgblq16R

[Charlie Graham @imcharliegraham]

@rileybrown How is this different than say having Hermes agent-channels and having agents associated to them in slack/discord?

[Charlie Graham @imcharliegraham]

@david__booth @scottbelsky It just means you will get the average advice of what you ask for. Like average of best practices. Which in many ways is really good.

But it’s not particularly unique or special. If you need that the models are not as good.

[Charlie Graham @imcharliegraham]

Getting so close to 1000

https://t.co/gz6WSzKOtw

[Charlie Graham @imcharliegraham]

Used 88% my weekly ChatGPT /Codex quota in 4 hours. AMA! 😁

[Charlie Graham @imcharliegraham]

Your best AI-native employees are currently incentivized NOT to share their best AI workflows.

If someone finds a workflow that makes them 30% more productive, keeping it private has real upside.

They can work less and be just as productive as they were before.

Or they can work just as hard and look like a rock star compared to their peers.

Sharing their skills, automations, and workflows removes the personal advantage that makes them special.

If a company really wants AI successes to spread across the organization, it needs to change the incentives.

A few ways I have seen this work (and that you can do in your company):

* Leaders should consistently and loudly recognize people who share useful workflows that their peers adopt. Share those successes in company-wide meetings.

* AI workflow creation and sharing should be part of performance reviews and 1:1s. Not just “what AI workflows have you created?” but “which workflows did you create that are now used across the department?”

* Make sharing AI wins an explicit company value. Create a Slack channel where people share what is working. Keep reminding people that helping the team get better at AI is valuable work, not an extra thing they do on top of their job.

By default, the people getting the best results from AI have a pretty rational reason to keep it to themselves.

The company has to make sharing that advantage more valuable than protecting it.

[Charlie Graham @imcharliegraham]

@petergyang @meetgranola @TrungTPhan Just don’t delegate it to “someday” as it won’t get done and then it’ll be too late