Install the productivity plugin so Claude remembers your TASKS.md, your team, and your jargon

A plugin in the Claude Code / Cowork sense is a bundle of skills, agents, slash commands, and MCP server entries you install as one unit. The productivity plugin, published in the Anthropic marketplace, packages four skills aimed at the problem most teams hit on day three of using Claude every day: Claude doesn't remember what you're working on, who is on your team, or what your project codenames mean. The four skills are start to bootstrap the system, memory-management for a two-tier memory of facts about your work, task-management for a shared TASKS.md, and update to sync from your project tracker. Installed together they turn Claude from a stateless assistant into a colleague that reads the same backlog you do.

Why pick this as Thursday's plugin. Last week's plugin pick was engineering — ten skills mapped to the working engineer's day. This is its companion. Engineering covers what Claude can do for the work; productivity covers what Claude needs to know about the work. The split matters because most people install task-style skills first and then find they can't use them well: Claude treats every new prompt as the start of a new world, so half the prompt becomes re-orientation. The fix isn't a longer prompt every morning; it's a memory system that loads itself, plus a sync skill that keeps it in line with the project tracker you already use. The plugin is the pre-built version of that loop.

start. The bootstrapping skill. Run it once after install and it does three things: it reads any task list you already have (a local TASKS.md, plus a quick scan of ClickUp, Monday, or MS365 if those connectors are installed), it asks for the shorthand you use day-to-day — nicknames, acronyms, project codenames — and it writes the dashboard view that lives at the top of your CLAUDE.md. The point of having a dedicated start skill (instead of "just ask Claude to set it up") is that the prompt is the same every time and the output format is the same every time. You can re-run it after onboarding a new project and the existing memory is preserved.

memory-management. The heart of the plugin. It enforces a two-tier model: CLAUDE.md at the project root holds the small set of facts you need every prompt (current sprint, key dates, who-is-who), and a memory/ directory holds the longer-form knowledge (per-project notes, decision history, glossaries). Claude reads CLAUDE.md on every session; it pulls from memory/ only when a description-matched file is relevant. The skill provides the writes — add a fact, fix a stale one, prune the index — so the memory drifts forward instead of becoming a graveyard of outdated notes nobody trusts.

task-management. The simplest of the four. It owns a single TASKS.md file, plain Markdown, and gives Claude a stable vocabulary for adding, completing, deferring, and reprioritising entries. The reason this is a skill at all (instead of "ask Claude to edit a text file") is the conventions: every task line has the same shape, completed items move to an archive section instead of disappearing, and stale items get a timestamp so the picker can flag them. Plain text is portable; the skill keeps it consistent across sessions, projects, and people.

update. The sync skill. Once a week — or once a day if your tracker is busy — run it and Claude pulls new tasks from Linear, Asana, Jira, ClickUp, or Monday (whichever connector you have), triages stale items, and fills memory gaps it noticed during the week ("who is 'Priya from the platform team'? add to memory/"). The output is a diff against TASKS.md and CLAUDE.md, not a silent rewrite — you approve the changes before they land.

The trap to avoid is over-stuffing CLAUDE.md. Working memory has to fit in your prompt budget; if every fact lives at the top, Claude reads them all every time and you've reinvented the problem you were trying to solve. Push the long-form into memory/ and trust the description-match to load it when it's relevant. If a project's section in CLAUDE.md grew past a screenful, that's the signal to split it.

See the try-it block for the install and the one command that bootstraps the system.