Put the constraint in, not the politeness
“Do not add dependencies” changes the output. “Please be careful” does not. If the model cannot fail your instruction, it is decoration.
Opening X27
13 items · 6 essentials · 7 advanced · Run with Builder
Skills, agents and workflows you install into Claude Code and run. Not reviews of software you already know about — specific things that do one job, built here or by members, with the install command on each.
Read. Learn. Then build.
Browse every tool and workflow. Academy includes all learning routes; running library tools starts with Builder ($25 a month).
Sign in →The copy pass is the expensive part. Most site generators skip it and it shows.
Read-only · run with BuilderView plans ↗once, then any item by slug
These commands illustrate the intended installation flow. The X27 CLI and member-key verification are not connected in this preview.
npm i -g @x27/clix27 loginnpx x27 add repo-briefItems land in .claude/skills · nothing runs until you invoke it
10 servers worth wiring in
MCP is how a model reaches something outside the chat. Each server widens what an agent can touch — which is the point, and the risk.
| Server | What it does | Why it matters |
|---|---|---|
| Filesystem | Read and write inside directories you whitelist. | Scope it narrowly. An agent with your whole home folder is a bad afternoon waiting to happen. |
| GitHub | Issues, pull requests, code search, repo metadata. | Lets an agent open a PR instead of describing one. |
| Postgres | SQL access, read-only by default. | Read-only by default is the correct default. Think hard before changing it. |
| Playwright | Drive a browser — navigate, click, type, screenshot. | How an agent verifies its own frontend work instead of asserting it looks fine. |
| Git | Local commits, branches, diffs, blame, log. | Blame and log let an agent answer “why is this here” instead of guessing. |
| Sentry | Query error issues and events from your projects. | Hand it a production error and let it trace back to the commit. |
| Fetch | Retrieve HTTP content as markdown. | Reading current docs beats relying on what the model remembers. |
| Memory | Persist facts across sessions in a local knowledge graph. | Useful and over-used. Store decisions, not transcripts. |
| Linear | Tickets, projects, cycles, sprints. | Closes the loop from ticket to PR without copying text between tabs. |
| Slack | Channels, DMs, search, posting. | The easiest one to regret. Never give write access to a channel you would not want surprised. |
habits that outlive the tools
“Do not add dependencies” changes the output. “Please be careful” does not. If the model cannot fail your instruction, it is decoration.
Point at the file that demonstrates your conventions. Models generalise better from one real example than three sentences about your preferences.
Paste the whole trace. Your summary encodes your assumption about the cause, and if that is wrong you have steered it into your own blind spot.
Context is full of wrong approaches it keeps reaching for. Start a new one with what you learned rather than pushing a third time.
Confident, well-formatted, plausible, wrong. If you are nodding along rather than checking, stop and run it.
A failing test turns “is this right” into a question the machine can answer, and stops it declaring victory on something that was never broken.
Every server you wire in widens what an agent can touch. Add write access one capability at a time, never to something you cannot undo.
Searching, summarising and classifying rarely need a frontier model. Reserve the expensive one for where being wrong is costly.
Draft · catalogue, versions and install counts are placeholder content