plungeai-skills-plugins

Declare and understand PlungeAI's capability-injection fields on a type: harness task or plungeai_run_mission — skills, experts, persona, backgrounds, plugins, and MCP servers — including eager-vs-deferred budgets and degrade-to-warning failure semantics. Use when the user wants an agent to follow a house style or method, adopt an identity/voice, get standing company context, or wire in a domain toolkit or MCP server. For authoring the mission itself use plungeai-missions; for saving a skill from this chat use plungeai-memory's plungeai_learn.

Download zip (opens in a new tab) · View raw SKILL.md (opens in a new tab)

Six capability kinds shape a type: harness task (or plungeai_run_mission) beyond its goal: skills (methods — HOW), experts and personas (identity — WHO), backgrounds (ambient truth), plugins (bundled domain toolkits), and mcp (connected servers). All six are resolved by one enrichment pass at run start.

Prerequisites

• A self-service ozk_ API key from Dashboard → One API → Keys (https://dashboard.plungeai.com (opens in a new tab)), or an MCP client connected to https://mcp.plungeai.com/v1.
• Discover every id before declaring it — a missing id never fails the run, it degrades to a silent warning: plungeai_list_agents {kind: "skills"|"experts"| "personas"|"connectors", search: "<topic>"} (REST: GET /v1/discovery/search?kind=…&q=…). Never guess.

Declaring capabilities

- type: harness
  goal: "Build a valuation snapshot for {input}"
  mission: |
    You are a buy-side analyst.
  persona: analyst
  experts: [securities-law]
  skills: [dcf-modeling-method]
  backgrounds: [acme-corp]
  plugins: [finance-bundle]
  mcp: [daloopa]

Injection order into the system prompt is fixed: BACKGROUND (verbatim, first) → persona text → EXPERT sections → skill bodies under ADDITIONAL INSTRUCTIONS → free-text instructions: (always appended, last word).

The eager budget — the thing that trips people up

All eager identity/instruction text shares one 24 KB budget: at most 3 backgrounds, 1 persona, 3 experts, 5 skills load eagerly; the rest defer to an ## AVAILABLE ON DEMAND index the agent can pull mid-run with load_skill. Order your lists by importance — must-follow items go first or into instructions: (which always lands, never defers). A plugin's bundled skills are always indexed, never eagerly injected, regardless of count — see references/plugins.md.

Which to declare

Do not unbundle a plugin into individual skills: entries to force eager injection — you lose its script index and blow the shared budget. A plugin's .mcp.json / manifest mcpServers documents the connectors its skills were written for; it is NOT auto-connected — declare that server's id yourself in mcp: alongside the plugin (references/plugins.md).

Gotchas

• Arrays REPLACE, never union, across mission-card merge (last wins: card → workflow root → task) — a task-level experts: [x] replaces the card's whole list.
• Only the FIRST entry of personas: [x] (plural) is used — you never get a blend.
• Missing/bad ids degrade to a warning, never a hard failure — a run that "lost its voice" or ignored a skill almost always has one in its warnings (plungeai-results-traces).
• A plugin declaration never fails the run even if it can't be expanded — a bad id degrades to a warning and the plugin's skills/scripts are simply absent.

Related skills

• plungeai-missions — the harness task these fields attach to, and the merge order (card → workflow root → task).
• plungeai-memory — plungeai_learn distills a chat session into a new skill.
• plungeai-results-traces — where a degraded-capability warning shows up.
• plungeai-tools-connectors — MCP server ids for the mcp: field.

Reference

• references/skills.md — declaring/creating skills, the eager budget, debugging.
• references/plugins.md — plugin expansion (skill index, MCP servers, script index), failure semantics.
• references/experts-personas.md — backgrounds, persona, experts: injection order, budgets, merge behavior.

Reference pages