Introduction

In conversations with colleagues and stakeholders, working out “will this phrasing land with this person, or step on a landmine?” from scratch every time is quietly exhausting. There’s a limit to how many people I can keep in my head, and I forget anyway.

So I built supernemawashi, a Claude Code plugin that solves this by managing profiles as data.

The workflow: collect facts from Slack / Gmail / Calendar / GitHub, classify them with psychology frameworks, generate situation-specific DO / DON’T rules, and call them up when I’m writing a reply. Everything runs locally — nothing is sent anywhere.

Why I built it

Saying “most of an engineer’s job is talking to people” is blunt, but it’s true. I spend more time asking, reporting, arguing, and persuading than I do writing code. And for all of it, I’m working out the optimal phrasing for each person’s type, every single time. That wears you down.

Concretely, situations like these:

• I phrased something “too bluntly” for someone and stepped on a landmine. I tell myself I’ll be careful next time, then forget.
• One person is convinced by numbers; another responds better to a story. Do it backwards and you get ignored.
• One person acts when you push in a meeting; another shuts down when you push in a meeting.

“Remembering it from experience” breaks down as the number of people grows. With someone I don’t have committed to memory, I start from zero every time — and even with people I think I remember, I forget things.

How it solves the problem

The idea itself is simple:

1. Record facts about the person (who said what, when) as data
2. Classify the facts with psychology frameworks (defense mechanisms / conflict modes / ego states / motivators / biases / attachment style)
3. From the classification, generate situation-specific DO / DON’T rules
4. When writing a reply, pull up that person’s profile plus the rules for the relevant situation

In other words, pre-compute “how do I talk to this person” and keep it in a file — so I don’t re-derive it on every reply.

Profiles live in ~/.local/share/supernemawashi/profiles/<person-name>/, entirely local. Even when it reads Slack / Gmail over MCP, the collected data never leaves the machine.

Skills

The operational skills are all verb-first under a nemawashi- prefix:

A basic workflow looks like this:

[1] /nemawashi-collect "John"
        ↓ collect facts from MCP sources
[2] /nemawashi-analyze "John"
        ↓ classify per framework, generate DO/DON'T
[3] /nemawashi-reply "How should I reply to John about the deadline?"
        ↓ load John's profile + only the relevant frameworks
[4] 2–3 drafts come back

Install is via the marketplace:

/plugin marketplace add Kohei-Wada/supernemawashi
/plugin install supernemawashi@supernemawashi

Restart the Claude Code session and the using-supernemawashi entry-point skill is injected via a SessionStart hook — from there it runs in natural language.

Gotchas

• The discomfort of “turning people into data”: profiles are for people you talk to, not surveillance targets. This isn’t built for profiling random coworkers behind their backs (and the MCP sources only run through your own account).
• MCP rate limits: hammering Slack or Gmail hits limits. The --all bulk operations are throttled to batches of 5.
• Stale profiles: a profile not analyzed for a month is less trustworthy. nemawashi-check surfaces staleness, and nemawashi-update --all re-analyzes everything at once.
• “Collect everyone” is heavy: collect is MCP-bound and slow, while analyze is local-file-only and fast — so the two are split into separate skills, letting you re-run just the analysis.

Results

• For someone I have a profile on, the time spent figuring out “how do I phrase this” dropped from ~5 minutes to ~30 seconds.
• For “someone I haven’t talked to in a while” or “someone I rarely talk to,” facts auto-collect from past Slack / Gmail, so the first-contact cost went down.
• I built a profile of myself and ran analyze on it — my own coping patterns became visible, which turned out to be a bit therapeutic (an unexpected side effect).

Wrap-up

• The cost of figuring out “how to talk to this person” from scratch every time can be cut down by turning it into data
• supernemawashi wires fact collection / psychological analysis / situation-rule generation / reply drafting into a single plugin
• Profile data never leaves your machine; MCP sources are read only through your own account
• It installs as a Claude Code plugin

The repo is here: supernemawashi (GitHub)