Slash commands and personas

A persona is a label you attach to a focused area of the catalog. The Cortex persona owns research skills. The Specter persona owns lead acquisition. The Sentinel persona owns engineering. The split exists because a flat catalog of 57 entries is hard to navigate; bundling them under seven heads of department is easier to reason about for both the user and the discovery model.

The runtime treats personas as a shortcut. When a user types a slash command, the regex extracts the persona id and the discovery step is given the persona's skill subset rather than the whole catalog. The discovery step still picks exactly one skill, runs it through the same engine, and reports through the same activity feed. Nothing about a persona changes how a skill executes.

This matters when you are reading code or activity logs. A row in the activity feed always names the skill that ran, not the persona that routed to it. A profiler trace always shows runSkillMission, never a persona function. If you are looking for the behavior of /cortex, look at the ten skills the Cortex persona points to and the persona prompt that frames the discovery call.

The regex anchors at the start of the message and requires whitespace after the slash term. /cortex map matches; cortex/map does not; /cortex alone with no body also does not match because there is nothing to dispatch. The case-insensitive flag accepts /Cortex from users who autocapitalize.

With agentId set, the chat handler injects a persona prompt at the top of the system message that defines the persona's mission and lists every skill in the persona's subset. The discovery tool sees only those skills as candidates. The model still picks one skill based on the user message, and the engine still runs that skill through the standard execution lifecycle.

Without a slash command, the discovery tool sees the entire catalog. The model has more candidates but no narrowing prompt, so the classification rule is purely semantic match against each skill's description. Both paths land in the same engine call.

Cortex owns the research surface. The persona prompt frames every turn as intelligence gathering: name the target, list the questions worth answering, suggest the artifacts to produce. The skills under Cortex bias toward web search, news ingestion, competitive teardowns, and memory writes that subsequent personas can read.

Specter owns prospecting and outbound. The persona prompt frames each turn as filling the pipeline: name the ICP, surface candidate leads, draft the outreach. The skills under Specter call lead databases, scrape signals, and generate the email sequences that feed Striker.

Striker owns deal momentum. The persona prompt sits the model inside a sales cycle: warm up the call, handle objections, follow through after the meeting. The skills are tightly scoped because Striker's job is consistency, not creativity. Each one runs short with predictable output.

Pulse owns the content engine. The persona prompt frames each turn around brand voice, distribution channels, and the format pick. Skills under Pulse cover long-form blog outlines, social posts, multi-platform repurposing, and the humanizer pass that strips the model-generated cadence.

Sentinel owns the software side. The persona prompt routes work into an isolated per-conversation sandbox. Skills under Sentinel build apps, deploy services, run tests, refactor code, and audit security. The execution surface for Sentinel skips most of the regular tool catalog and leans on code_task, deploy_task, and the bundled engineering CLI inside the sandbox.

Amplify owns paid acquisition. The persona prompt frames campaign setup, audience targeting, creative briefs, and bidding strategy. Skills lean heavily on the company profile and brand voice context so creative is consistent with everything Pulse produces organically.

Counsel owns contract work. The persona prompt frames each turn around the standard contract questions: parties, scope, term, fees, IP, liability, termination. Skills include a clause library lookup, drafting from a template, redline review with citations, and a plain-English explainer pass.