Inside Cursor's Official Plugins: Four Workflow Cards That Show How Agents Should Work

Last night's post ("One Rulebook, Seven Moves") covered how Cursor plugins migrate across clients. This time we crack them open: reading several officially-authored plugins as workflow cards—they are effectively a platform vendor's official answer sheet on "how an agent should do its work."

Cursor's plugin marketplace manifest currently lists 113 plugins. Six official ones are most representative: advisor, ralph-loop, orchestrate, thermos, agent-compatibility, and teaching. One by one.

Card One: advisor—A "Second Trial" for the Agent

The main model works as usual; advisor is consulted at three checkpoints: before major decisions, when stuck on an error, and before declaring a task complete. The consultee receives a brief and the full conversation transcript, thinks, and returns a ruling with recommendations—it never edits any file; the main model still decides whether to change anything.

The engineering details are more interesting than the concept. State lives in .cursor/advisor/state.json at the project root; enabling it rewrites this file to rebind to the current session and clears pending flags left by other sessions; a stop hook reminds the system at the end of every round to "consider asking"; it even counts consultation invocations.

Most interesting is the default: the consultee is "the latest Grok, highest reasoning tier" (currently cursor-grok-4.6-xhigh), with options like composer-2.5 listed as alternatives. The platform vendor's default second reviewer for its own main model is a competitor's model—this default is itself an evaluation.

This pattern isn't unique to Cursor: Anthropic's platform documentation also describes an identically-named advisor tool—a fast, cheap execution model consulting a stronger advisor model mid-generation. Two vendors have nearly simultaneously converged on the same shape.

Card Two: ralph-loop—Feeding the Same Prompt Back

The origin: the Ralph Wiggum technique, named by Geoffrey Huntley in July 2025—infinitely looping the same prompt into a coding agent, where each round it can see its own previous round's work. Huntley's first commandment: one loop does one thing.

Cursor's version welds the loop into hooks: a stop hook intercepts each round's end and feeds the original prompt back as a follow-up message, prefixed "[Ralph loop iteration N.]". State is written to .cursor/ralph/scratchpad.md: iteration count, an upper limit (default 0 = unlimited), and a completion_promise—a completion passphrase. The agent must utter that agreed-upon string for the loop to stop.

The point in one sentence: completion is not "feels done"—it is a string written in configuration.

Card Three: orchestrate—Fan Out, Then Fold Back In

Large tasks are split among parallel cloud agents: a planner publishes tasks, workers pick them up, finish, and hand off upward; a final script reconciles the trajectories and wraps up. Under the hood it uses the Cursor SDK to launch parallel cloud agents.

What's worth remembering is the division of labor: what each party does belongs to the model; trajectory alignment belongs to the script. Coordination logic is split half in the model, half in deterministic code—the parts requiring judgment go to the model; the purely mechanical parts never bother it.

Card Four: thermos and agent-compatibility—Auditing Others, and Auditing the Foundation

thermos's self-introduction is combative: "thermonuclear-grade branch review"—deep correctness and security checks, paired with a brutal code-quality scorecard, with parallel sub-agents investigating in parallel. agent-compatibility works the other way: scanning repositories via real CLI test runs, specifically checking startup flows, validation logic, and whether documentation "matches reality"—if the docs say it runs, run it and see.

One audits code written by humans; the other audits the ground the agent itself depends on. Their common trait: neither trusts verbal reports—scorecards and real executed commands are their forensic instruments.

Six Cards Side by Side

Not a single one teaches the model to be "smarter." advisor adds a pair of eyes; ralph adds a loop; orchestrate adds a division-of-labor sheet; thermos adds a scorecard; agent-compatibility adds a ruler of real tests. Judgment goes to the model; process goes to files—state files, hooks, completion passphrases, scorecards: all things outside the model, visible and tangible.

Closing on that advisor default: when the main model gets stuck, whom does it ask? Cursor hard-codes the first answer: the latest Grok, highest reasoning tier. The vendor tests these models every day—this default is more honest than most leaderboards.

Cursor plugin repository: github.com/cursor/plugins (13,205 stars, MIT)