alexmskills
Claude Code plugins that get better at your repo every time you use them.
A real Claude Code session, recorded. The build registers itself — its live bar and learned time-left ride in the status line, where you're already looking.
Get started in 30 seconds · Browse the catalog · Read the docs
Why this marketplace
🧠 It learns, and writes it down.
Skills keep what they learn in your repo, not in a chat that scrolls away. A CLAUDE.md that prunes and archives itself. Roles that accumulate lessons run after run. A coach that stops nagging about habits you've already mastered.
👥 Teams, not a single prompt.
Three orchestrators compose role-specialized agents to fit the task — one to discover, one to decide, one to deliver — and they share one evolving pool of personas.
📡 You can watch it work.
Live progress in the status line and a local dashboard. Every demo below is a recording of the real tools running — two of them inside real Claude Code sessions — not a mock-up.
✅ Measured, not asserted.
Rule-based plugins cite a source for every rule and calibrate against real prompt corpora and real repos before shipping. The script-backed plugins carry test harnesses, and CI validates the whole catalog on every push.
See it in action
📊 Every long job, one glance — progress-channel
A pipeline and its current step, nested. The parent's bar and time-left roll up from the step that's running, so a long stage never reads as a frozen bar. Nest a whole shell pipeline with one line — export PROGRESS_PARENT=$T — and every sub-script, even one that knows nothing about pipelines, files itself underneath.
- No setup in Claude Code — a built-in mod draws the bars above your prompt, and steps aside if you wire the status line instead.
- Learned ETAs from each job's own history, borrowed from similar jobs when the name is new.
- Bars from the tool's own output — Maven, Gradle, Cargo, Go, pytest, Jest, Docker, Ninja, CMake, .NET and more, read in passing by a transparent pipe tap.
- Runnable integration examples for shell, Python, Node, Go and Make in
examples/.
- Stall and orphan detection from each job's own rhythm, not a fixed timeout.
▶ Also: the terminal view — progress watch and a pre-start forecast.
🎯 An expert panel on any problem — brainstorm-panel
A real session. Asked to pick this repo's next demo workload, it seats a director, a skeptic and a practitioner — each with a stated job — and debates to a decision.
Hand it a problem, a draft, or a half-formed idea. It picks the experts the work needs — a skeptic is always seated — chooses how they should argue, and runs generate → critique → refine until they converge. The panel learns which seats earn their place in your repo.
🔁 Discover → decide → deliver — the Role System
flowchart LR
S["🔍 research-sweep<br/><b>discover</b><br/><sub>parallel scouts + adversarial verify</sub>"]
P["🎯 brainstorm-panel<br/><b>decide</b><br/><sub>debate until it converges</sub>"]
C["🚚 dev-crew<br/><b>deliver</b><br/><sub>gated relay: architect → dev → qa</sub>"]
R[("👥 roles<br/>one shared, evolving<br/>pool of personas")]
S --> P --> C
R -.-> S
R -.-> P
R -.-> C
One skeptic is a verifier in the sweep, a seat on the panel, and the adversarial check in the crew — and everything it learns in one place goes with it to the others. Each orchestrator works on its own; together they chain. How the Role System works →
🧭 A coach that learns how you prompt — prompt-coach
It reads each prompt as you send it and, when one would go better, rewrites it with you — naming the rule and the fix. Mastery is earned by using the technique, not by avoiding the mistake. And rules know their model: advice that a newer Claude already follows by itself switches off on that model.
🗺️ Think in a map, ship a prompt — mindmap-prompt
Some ideas don't arrive in order. Drop them on a canvas, connect them, mark what must not change — then press Compile and the map becomes an organized prompt. The compiler is plain code, not a model: the same map always gives the same prompt.
📝 A CLAUDE.md that keeps itself current — evolving-claude-md
📝 CLAUDE.md audit (23.2 KB total; 49 entries, 56 lines in Decisions & Learnings).
Compaction RECOMMENDED.
🔎 CHANGELOG MIRROR: 28/49 entries (57%) name a version CHANGELOG.md already
documents. A release is not a learning — keep only what the release TAUGHT
(a constraint, a trap, a reversal) …
A real audit of this repo's own CLAUDE.md, shown at session start — yes, it's telling us to tidy up.
Decisions and gotchas land in a log with a format enforced on every edit. The log prunes, graduates stable lessons into conventions, archives the old, and tells you when an entry has gone stale — a path that no longer exists, a version that has moved on.
When the audit says it's time, /evolving-claude-md:compact proposes the whole clean-up as one reviewed edit — lessons graduate into standing rules before the rest is archived, so nothing worth keeping leaves the file.
Mods
Not skills — code that draws inside Claude Code itself. A mod has no instructions for the model to read, costs no tokens, and makes no extra API requests. These two put what you'd otherwise have to ask for right above your prompt.
Both mods in one real session, recorded: your context window on top, your rate limits beneath.
| Mod | | Version | What it draws |
|---|
context-bar | 🧩 | 0.2.2 | Your context window as a stacked bar, a nudge to compact before auto-compact cuts in, and top to see what fills it. |
usage-bar | 🧩 | 0.4.0 | Your 5-hour and 7-day limits against the clock, plus how long the prompt cache stays warm. |
context-bar answers "what is filling my window?" without you asking: one segment per category, with a token legend, refreshed after every turn from a local estimate. Toggle it with /context-bar.
usage-bar keeps your rate limits in view, and changes colour before they bite — green with room to spare, yellow from 70%, red from 90%:
The three states, drawn by the mod's own functions at sample percentages — a live recording can only show the state an account is really in.
/plugin install context-bar@alexmskills
/plugin install usage-bar@alexmskills
Mods install this way only. npx skills add copies skills, and a mod has none.
progress-channel ships a third: its live job bars, drawn in the same band with /progress-bar to toggle. It comes with that plugin rather than as its own install.
Install both and both draw: each adds its rows and then lets the next mod draw its own. Mods need a Claude Code build that loads hooks modules, and are Claude Code only. How mods work →
The catalog
🧠 learns from use · 🔧 workflow · 🔍 research & review · 📚 domain knowledge
| Plugin | | Version | What it does |
|---|
evolving-claude-md | 🧠 | 1.9.0 | A CLAUDE.md that prunes, graduates and archives itself. |
brainstorm-panel | 🧠 | 1.3.3 | A task-fit expert panel that debates until it converges. |
dev-crew | 🧠 | 1.3.1 | A task-fit delivery crew, run as a gated relay. |
roles | 🧠 | 1.4.1 | One evolving pool of personas, shared by every orchestrator. |
prompt-coach | 🧠 | 1.5.1 | Rewrites your prompts toward better habits; mastery is earned. |
conductor | 🧠 | 0.1.1 | The main session's playbook while parallel agent lanes run. |
learn-on-failure | 🧠 | 1.2.1 | Saves a lesson whenever a fix took more than one try. |
memory-hygiene | 🧠 |
Two plugins in progress ship with a -beta suffix: tune-repo-beta and systemic-fix-beta.
Get started in 30 seconds
There are two ways to install. Pick one for each plugin.
Way 1: Claude Code plugin. Installs everything: skills, hooks, agents, slash commands and mods.
/plugin marketplace add alexmond/alexmskills
/plugin install progress-channel@alexmskills
Way 2: npx skills. Works for Claude Code, Cursor, Codex and other agents. Copies the skills only.
npx skills add alexmond/alexmskills --list # see what is there
npx skills add alexmond/alexmskills --skill brainstorm-panel # install one
Way 2 is complete for 11 plugins and incomplete for 11 that rely on hooks, agents or scripts. The two mods need way 1. Which is which →
Then just ask:
| Say this | And you get |
|---|
| "make CLAUDE.md evolve" | evolving-claude-md sets up a self-maintaining decisions log, with hooks that register themselves. |
| "get a team on this and make it better" | brainstorm-panel proposes a panel for your sign-off, then runs the debate. |
| "run the crew on this feature" | dev-crew composes a roster and relays it, stopping at hard gates. |
| "set up the progress status line" | progress-channel wires live bars into your Claude Code status line. |
/plugin install evolving-claude-md@alexmskills
/plugin install memory-hygiene@alexmskills
/plugin install dev-crew@alexmskills
/plugin install brainstorm-panel@alexmskills
/plugin install learn-on-failure@alexmskills
/plugin install roles@alexmskills
/plugin install prompt-coach@alexmskills
/plugin install screenshot-sweep@alexmskills
/plugin install skill-linter@alexmskills
/plugin install implement-issue@alexmskills
/plugin install maven-quality@alexmskills
/plugin install security-audit@alexmskills
/plugin install mindmap-prompt@alexmskills
/plugin install screenshot-tour@alexmskills
/plugin install progress-channel@alexmskills
/plugin install context-bar@alexmskills
/plugin install usage-bar@alexmskills
/plugin install ticket-triage@alexmskills
/plugin install review-agents@alexmskills
/plugin install research-sweep@alexmskills
/plugin install spring-batch@alexmskills
/plugin install conductor@alexmskills
/plugin install tune-repo-beta@alexmskills
/plugin install systemic-fix-beta@alexmskills
Or browse with /plugin (Discover tab). Pull updates with /plugin marketplace update alexmskills.
Try one without installing: claude --plugin-dir ./plugins/dev-crew.
Using Codex? Every plugin also ships a Codex manifest, and the repo carries a second marketplace at .agents/plugins/marketplace.json. 9 plugins port as-is, 6 ship a Codex tool translation, and 9 depend on Claude Code hooks — the two mods among them, which are Claude Code only. Codex compatibility →
Learn more
📖 Full documentation — one page per plugin, the Role System, and Codex compatibility · 📜 Changelog
Versioning. Claude Code versions per plugin, not per skill. Each plugin carries a semantic version in its plugin.json and a matching entry in marketplace.json. Users on a pinned version only receive an update when the number changes. Tightly coupled skills (like maven-quality's) share one plugin; everything else moves independently. Record every bump in CHANGELOG.md.
Beta plugins. In-progress plugins live in the same marketplace with a -beta suffix, so it's obvious at install time. Graduation renames the directory, drops the suffix and bumps the version — prompt-coach graduated this way on 2026-07-28.
| Command | Purpose |
|---|
make validate | Validate the marketplace, every manifest, docs coverage and this README's catalog. Runs in CI. |
make list | Print the catalog. |
make bump PLUGIN=<name> VERSION=<x.y.z> | Bump a plugin's version in both manifests together. |
make graduate PLUGIN=<name>-beta | Graduate a beta plugin. |
make codex | Regenerate the Codex manifests from the Claude-side source. |
claude plugin tag --dry-run plugins/<name> | Check a plugin's manifest agrees with its marketplace entry. |
Contributing a plugin. Add it under plugins/<name>/ with a .claude-plugin/plugin.json and a skills/ and/or agents/ directory, register it in .claude-plugin/marketplace.json, add a docs page under docs/modules/ROOT/pages/ with a nav.adoc entry, and run make validate.
Docs are an Antora component under docs/, published by the alexmond.github.io site build. Demos run against skillsample, a small stage repo, and are recorded with VHS; each .tape source sits next to its GIF.
Provenance & attribution
All plugins in this marketplace are original work by Alex Mondshain, MIT-licensed — authorship is
recorded in each plugin's plugin.json (author) and the repo LICENSE. One third-party
inspiration is credited: the roles plugin's persona set was seeded by a public prompt thread from
@nahidulislam404 on X — the wording here is generalized and rewritten,
but that thread sparked the original set.
License
MIT © Alex Mondshain
