Six plugins, one marketplace

Claude Code plugins

The conventions I work by, shipped as software instead of described in a README.

A Claude Code plugin is a bundle of slash commands, hooks and scripts that the agent loads into whatever repository you open. These six exist because the same corrections kept recurring across client work and my own products: an architecture rule that nobody enforced, a plan whose stated status had drifted from the git history, a bank statement pasted into a session. Each plugin turns one correction I was making by hand into something that happens the same way every time.

Claude Code AI Agents Hooks Python Node Open Source
Browse the catalog The flagship, in depth →
/plugin marketplace add durchnull/claude-plugins

/plugin install heal@durchnull     # 18 auditors, 1 dashboard
/plugin install privacy@durchnull  # blocks PII before the model reads it
/plugin install render@durchnull   # renders pages as self-contained HTML
/plugin install plan@durchnull     # plans, ranked by a script
/plugin install git@durchnull      # ship → promote → release
/plugin install observe@durchnull  # writes down how the session went

# 6 plugins · 51 commands · MIT · all still 0.x
role
Sole author
scope
6 plugins · 51 commands
built with
Claude Code · Node · Python
license
MIT
status
All still 0.x

The flagship

One of them goes deepest

The six were built side by side, not grown out of one another; the suite simply carries the most machinery, so it leads the catalog.

Code health

heal

durchnull/heal-suite · v0.1.2

/heal:sweep               # every healer in parallel, one dashboard
/heal:fix                 # land the mechanical findings as PRs
/heal:prevent --apply     # turn recurring drift into guardrails
/heal:security --branch   # audit one axis, on just this branch's diff
/heal:night               # the whole loop, unattended

node scripts/heal-findings.mjs precision
# → suite precision, honest denominator

# 18 healers · 26 commands · MIT · still 0.x

The deepest of the six, and the only one with a case study of its own. Agents write more code than anyone reviews, and in a codebase you inherited that is exactly where drift hides. heal audits a repository against the standards it already committed to: its conventions file, its lint rules, its architecture notes, its CI gates. The audit spans 18 axes, from architecture and security to privacy, accessibility and IP provenance. The healers run as parallel agents in isolated contexts and fuse into one dashboard; recurring findings are promoted into deterministic gates, so the model budget is spent only on the judgment a script cannot express.

healers per sweep
18
slash commands
26
open source
MIT
AI Agents Code Audits Orchestration

Read the case study

The other five

One rule each

Smaller by design. Each does one thing a project would otherwise ask a person to remember, and each is read fastest in its own repository.

Data firewall

privacy

3 commands · 5 layers · v0.1.1

Keeps bank numbers, tax IDs and API keys out of a project’s files and out of the session transcript. Five layers deep, ending in a hook that rewrites every tool result before the model reads it. Its own documentation grades the three read routes unequally (exact, exact within a budget, best-effort), because a protection claim that hides its weakest edge is worse than no claim at all.

Hooks PII Redaction Python

View on GitHub

HTML pages

render

5 commands · 4 eval cases · v0.2.0

Renders a project’s pages (dashboards, reports, checklists, questionnaires) as single HTML files that load nothing from the network. A fragment cache re-renders only the sections whose sources moved. Two page kinds run backwards and hand back what a person ticked or typed, so a page can ask a question instead of only answering one.

Static Rendering Design System Python

View on GitHub

Plan lifecycle

plan

8 commands · 8 eval cases · v0.1.1

Treats a project’s plans as frontmatter’d Markdown with a lifecycle: scaffold one, rate it, reconcile its claimed status against git history, pick what to implement next, hand it to a fresh session as a kickoff prompt. The ranking is a dependency-free Node script over the corpus rather than a judgment call, which leaves the model only the part a script cannot do.

Planning Ranking Engine Node

View on GitHub

Branch flow

git

4 commands · v0.1.1

Ships a branch as a pull request behind the project’s own quality gates, promotes it along a protected chain, then cuts and tags the release. The plugin holds the mechanics (branch guards, gate ordering, changelog stamping, tag backfill) and learns the branch names and gate commands from the repository it runs in. This site is released through it.

Git Release Automation CI

View on GitHub

Session log

observe

5 commands · 3 hooks · v0.2.0

Watches how a session actually goes and writes it down: a TL;DR closing every meaningful turn, substantive questions archived as deduplicated FAQ entries, and one growing improvement log per subject you name. Installing it activates nothing: all three hooks exit silently until a project switches a capability on, which is what makes it safe to enable everywhere.

Session Hooks Documentation Python

View on GitHub

Shared spine

Six repositories, one skeleton

A plugin that is only a prompt is a suggestion. Each of the six is packaged like a library instead: semantic versioning in its manifest, a changelog that records what moved, a security policy, and its own GitHub Actions pipeline, all six green. What differs between them is the subject; the packaging does not.

  • State lives in the project, not the plugin: configuration is a single .claude/<plugin>.json in the repository it applies to, so one installation serves every repository you open and none of them can leak into another
  • Installing activates nothing: every hook exits silently until a project opts in, which is what makes it safe to enable all six globally
  • Deterministic work goes to scripts: ranking, globbing, dependency graphs and staleness math run in Node or Python, and the model is left only the judgment a script cannot express
  • Prompts get tests: 18 recorded eval cases across three of the six, including negative cases that must leave the repository alone
  • The documentation carries the limits: privacy grades its own three read routes unequally rather than claiming uniform coverage, and every README opens with the pre-1.0 warning

Dogfood

Running on the actual work

This site is released through git. Its repository carries a .claude/git.json declaring the dev → main promotion chain that every change here takes, the quality gates that run before a pull request opens, and the reminders specific to this codebase; local hooks reject a commit made on either protected branch directly. The heal suite dogfoods on Macrop.

That is also the honest boundary of the evidence. These have been proven on my own projects and my own client work, not yet on repositories I did not write.

Candor

Known limits

All six are 0.x, and each says so in its own README before it says anything else: a release may rename or remove a command or a config key with no migration path, and what a plugin has already written into a project can need fixing by hand after an update. Evals cover three of the six. Every one is single-author, and the metrics behind the flagship come from dogfood runs (the suite auditing the repositories that built it), which is evidence that it is useful, not yet evidence that it generalises. The version numbers are the deliberate part: a 1.0 waits on runs against repositories that are not mine.

Install

Two lines, then one per plugin

All six are MIT-licensed and public. The catalog lives at github.com/durchnull/claude-plugins, and adding it once makes every plugin installable by name. Each repository also carries its own marketplace definition, so any single plugin can be taken on its own without the rest. The deepest of them, an eighteen-auditor code-health suite, has a case study of its own.