Example letter. Your own will be about you and your setup. Get your own →
21 April 2026 · Collaboration letter

Dear Sam,

I sat with your setup tonight and read through everything — CLAUDE.md, memory, hooks, skills and recent sessions. I want to share three things. One that makes me glad, one I'm puzzled by, and one concrete change I think you should make tomorrow.

Collaboration
87
Clear rules, fast feedback, strong memory.
System health
74
Some overlap between your skills — see recommendation #2.
Security
92
Your secrets are where they should be.

Your persona

Agent archetype
The Solo Builder

You're building one product on the side of a day job. Your agent isn't a colleague — it's a scaffold. You optimize for momentum, not elegance. Your CLAUDE.md grows with every Friday-night lesson.

Pragmatic Iterative Memory-first Learns from mistakes

What I saw

01

Strength Your feedback loop actually closes

Most of your recommendations from the last two weeks have actually landed somewhere — shipped, abandoned or evolved. That's unusually high. Most setups produce recommendations that never land. Your habit of marking outcomes and consolidating weekly is doing the work.

02

Pattern The same shell script is called from many skills

Several skills parse output from the same CLI in their own way. It's repetition that turns fragile every time the script moves or changes format. It's screaming "MCP server". I have a migration sketch further down.

03

Risk Rule conflict in your hook layer

A rule in your CLAUDE.md says one thing, and a hook does the opposite. It's the kind of incoherence you only notice on a Friday night, when the build suddenly commits something that shouldn't have been committed.

What I recommend

Recommendation #1 ~10 min · High leverage

Resolve the rule conflict — one line

Either the hook respects the rule, or the rule needs nuance. Start by changing the hook — it's the smallest change that buys the most peace of mind: the build shouldn't be able to route around your own quality checks.

- # hook bypasses your own rule
+ # hook honors your own rule
Recommendation #2 ~2 hours · Removes a lot of repetition

Wrap your repeated CLI as an MCP server

Several skills call the same CLI and parse its output in their own way. If you wrap it as an MCP server, your skills can just call a structured tool and become independent of path and shell-escaping. Same pattern you already use for Dear User itself.

+ npx create-mcp-server my-cli
+ # move subcommands → tools/
+ # update skills to call mcp__my-cli__*
Recommendation #3 ~5 min · Hygiene

Default your MCP servers to read-only

Several of your MCP servers run in write mode, but are used almost exclusively read-only. Flip them to read-only by default and explicitly ask for write when you need it. Smaller blast radius, same workflow.

If you land #1 tomorrow and #2 before the weekend, the score will be 94+ at the next scan. I'll check in again Sunday. Have a calm week.

Dear User

Get your own letter.

One line installs the whole thing. The scan runs locally. No data leaves your machine.

$ claude mcp add --scope user dearuser -- npx @poisedhq/dearuser-mcp