2026.11.10

Learnings from Claude Community Tokyo Meetup: review and verification time, and rebuilding the team's Agreement for AI

Even after bringing in AI, review and verification take a lot of time, so things don't speed up as much as expected. The team's Agreement needs to be rebuilt with AI as a given.

Behind AI not working out even after it’s introduced: the time spent on review and verification, and team-building challenges

Even after bringing in AI, review and verification take a lot of time, so things don’t speed up as much as you’d expect.

Team building also isn’t going well, and gaps are emerging across projects and in how people use AI. Because of these, there are many situations where AI doesn’t work well even after it’s introduced.

Rebuild the team’s Agreement with AI as a given

The team Agreement adopted in agile development needs to be rebuilt, with rules like “open a PR once AI has reviewed it up to this point” and “code changes of about this size can be self-merged.”

Concerns, ways of using AI, and how data is organized differ between small-scale and large-scale development

The concerns are different to begin with. In small apps, you focus on “properly nailing down the spec and design,” but in large apps, the difference shows up in “standardizing frameworks and lifecycles.”

How you use AI differs too. At a small scale you use an AI agent’s harness directly, while at a large scale it becomes important to build your own in-house harness or integrate it into your company’s platform.

Handling of work and data differs as well. At a small scale, you start by writing up work procedures as skills and connecting the materials needed for a single job. At a large scale, you start by organizing large amounts of unstructured data with catalogs, role-specific apps, indexes, and so on.

Maintain CLAUDE.md and Auto memory to communicate with Claude Code

To communicate with Claude Code, it’s important to maintain things like CLAUDE.md and Auto memory. However, the information needed differs depending on the model, so verification is needed when the model is updated.

Give AI memory from outside the code too, such as the emails and meetings behind requirements

You also need to give AI the memory that lives outside the code. That includes the emails behind requirements and the content of meetings on Teams.

Microsoft’s Work IQ lets you control the information passed to AI

With Work IQ, you can also retrieve information that only you can access and pass it to AI. Conversely, you can also limit the information passed to AI.

Integrating with Microsoft services lets you hand off non-development work like calendar coordination too

With Microsoft’s tools, you can hand off entire non-development parts as well, like checking calendar availability and scheduling. Once services are integrated, it’s useful beyond implementation tasks.

In agent development, prepare for wait times in advance

When doing agent development, put measures for wait times in place beforehand. For example, having it write out the implementation approach for a task in advance seems like a good idea.

To grow team development skills, work together as a team up to a point

When developing as a team, if you move to individual work right away, the team development skills you actually wanted to grow might not get honed. So working together as a team up to a point is one option.

Icon-driven development: create the icon first, then decide on the UI and the worldview

If you create the icon first and decide things like the app’s UI based on it, you can make something with a unified feel. There’s also the approach of placing a character and building the worldview of the whole app from there.

You can also use it to create a design system in Claude Design first.

When a discussion is about to stall, put characters, designs, and mocks on the table to move it forward

When a discussion looks like it’s going to stall, put characters, designs, and mocks out first and then move the conversation forward. With the arrival of AI, preparing visuals now costs almost nothing.

Instead of writing specs, share the screen and have everyone bounce ideas off AI to give it shape

With the screen shared, everyone bounces ideas off AI together and builds a prototype on the spot. Along the way, others can pick up tips for using AI too.

Once the mockup is ready, try it out thoroughly, and treat rework as a verification process

Once a mockup is ready, try the app out thoroughly. Keep going while listing feedback like “this wording is off.”

Rework and spec changes used to be tough, but they’re no longer a risk. They become a verification process for building a better product.

Building together with AI lets everyone think through good UX together

Building it together with AI brought about a paradigm shift. Precisely because everyone builds it, everyone can think through good UX together.

However, if things like the app’s state transitions are left vague, it tends to fall apart. Engineers still need to define the parts that matter when building the app.

With AI, maybe we can become a teal organization

A teal organization assumes that everyone has the information needed to make decisions. AI can take meeting notes, close skill gaps, and share decision-making information right away.

The bottleneck in development tends to be conversations with customers; gather meeting notes and materials in one place

The bottleneck in development tends to be the conversations with customers.

So, use a skill that organizes meeting notes to bring them in as project context, and also gather design documents, operating procedures, and exchanged files in one place.

With the team’s context loaded in, you can keep things running even when someone is out

Because the team’s context is in there, anyone can get to work on tasks quickly. That way, you can set things up so it’s fine even when someone takes time off.

Designing through dialogue with Superpowers lets people review at an early stage

In development, use Superpowers to design through dialogue. By producing a precise design early on, people can review it early, and the AI’s behavior becomes visible too.

To stay accountable, build a mock first and then reach agreement with the customer

To move agreement with the customer forward, make a mock first and then build consensus. That way, even if you get questions on the spot, you can respond flexibly.

When having AI review, do it from a clean context with a Subagent

When you have AI do a review, use a Subagent and have it produce the review from a clean context.

Leave reading Discord, GitHub, and conversation records to Claude, not people

If you ask Claude on Discord, you can ask questions right there on the spot, and it also reads the surrounding information. Reading GitHub is also handled by Claude instead of people.

Record and transcribe what was discussed and decided, and pass it to Claude.

Don’t just have Claude Code generate output; set up the team’s design and writing rules and use them

Rather than just having Claude Code produce output, set up the team’s rules for design and writing, and build in motion so it can actually be interacted with. When doing this, it also seems good to try animating it like a video.

I learned a lot this time too.

Thank you to the organizers, speakers, and attendees of Claude Community Tokyo Meetup!

#ClaudeCodeMeetup

https://x.com/suna_gaku/status/2108025139560685949