A skill is a chunk of domain knowledge — a methodology, a checklist, a runbook, a composition recipe — written in plain markdown and made available to one or more agents on your dashboard. When you ask a question, Delphi decides which of your installed skills are relevant for that request and uses them as additional context for that single turn. The base behavior stays clean; the skill provides the specialist context.
Skills follow the SKILL.md convention popularized by Anthropic and Vercel. Each skill is a markdown file with a short header declaring its identity, plus an optional bundle of supplementary files (reference docs, example outputs, templates) that ship alongside the main body.
Why skills instead of bigger system prompts
A general-purpose agent has to be small enough to load on every turn and broad enough to cover everything. A skill body can be much larger — a multi-page runbook, a decision framework with worked examples, a complete composition playbook — because it only loads when the model actually needs it. The same agent can have access to a dozen skills and only pay the cost for the ones it touches.
Skills also make it possible to share methodology across organizations. The same C-suite cockpit composition recipe that ships built-in can be customized per dashboard, or replaced with a tenant-wide variant, without forking the underlying agent.
Where skills come from
Skills come from three sources, in order of precedence:
- Built-in — shipped with the platform and updated on every release. They’re the same on every command center.
- Tenant-wide — installed by a tenant admin from the admin area’s Skills tab. Visible on every dashboard within the tenant. Use these when a methodology should apply uniformly across all your command centers.
- Dashboard-scoped — installed by a dashboard editor and visible only on that command center. Use these for one-off methodologies that only matter on a single dashboard.
If two sources ship a skill with the same name, the higher-precedence one wins. Built-in beats tenant-wide; tenant-wide beats dashboard-scoped. Built-ins are versioned with each release, so they’re always the authoritative answer for their name.
Installing a skill
Two flows, depending on scope:
- Dashboard-scoped — open the Agents tab on a command center, switch to the Skills sub-tab, and either paste a SKILL.md body directly or install from a public registry of skill catalogs. Editor scope or above.
- Tenant-wide — open the admin area’s Skills tab. Same installer, but the skill becomes visible on every dashboard in the tenant. Admin or owner only.
Uninstall is a single click from the same surface. Built-in skills can’t be uninstalled, but you can hide them per agent (see below).
Activating a skill on an agent
Skills are off by default. Each agent carries a skill allow-list that controls which skills Delphi will consider for that agent’s turns:
- None — the agent has no skill access. Use this when you want an agent to behave on its base prompt alone.
- All — the agent can reach for any skill installed at the tenant or dashboard level. Delphi picks per turn based on the user’s message and the skill’s description. This is the right default for a general-purpose agent.
- A curated list — the agent can only reach for the skills you’ve explicitly enabled. Useful for tightly-scoped specialists where you want the model to lean into one specific methodology and ignore the rest.
The shipped Delphi generalist defaults to All so any skill you install becomes available to chat immediately. The visualization specialist ships with no skills — it doesn’t need them. Custom specialists carry whatever allow-list fits the work.
What a SKILL.md looks like
A minimal skill is a short header plus a markdown body:
---
name: my-skill
description: One paragraph explaining what this skill teaches, who it's for, and when the model should reach for it. The description is what Delphi uses to decide whether to use the skill for a given turn — make it specific.
---
# My Skill
The main body. Write it like a runbook, a methodology guide, or a domain
briefing for someone who already knows the platform but doesn't yet know
the specific work.
## When to use
- Trigger patterns
- Audience cues
- Example phrasings
## How it works
The body can be long. Multi-section runbooks, decision trees with worked
examples, full composition playbooks — all fine. Only the skills Delphi
chooses for a given turn pay token cost.A skill can also ship supplementary files — additional markdown alongside the SKILL.md, such as reference docs or templates the body refers to. Those are included in the install and made available to the agent whenever the skill activates.
What’s in the built-in catalog today
- Common Operating Picture — the foundational composition methodology behind every multi-layer command center. Defines the five-layer abstraction, four operator vectors, and the configural-display posture that the executive-style dashboards use.
- C-Suite Cockpit — commercial-enterprise dialect of Common Operating Picture. CEO / CFO / COO / Board vocabulary, Value-Based Management as headline KPI, capital-allocation and strategic-vector framing.
- Joint Warfighting — public-sector / defense dialect. Domains-aware (land, sea, air, space, cyber), commander’s-intent framing, operational vs. strategic split.
- Dashboard Composition — the methodology Delphi follows while populating an empty dashboard from chat.
- Governed Actions — when and how to propose actions, the approval workflow, the autonomy-tier ceiling, and how to draft a proposal the approver can act on.
- PII Handling — the canonical rules for redacting personal data in chat output. Aggregate-only HR data, name with role and department only (never paired with addresses, salaries, or government IDs), how the upstream connector filters interact with chat-side redaction.
The built-in set will grow over time as new domain dialects ship with the platform.
When skills aren’t the right tool
Skills give the model methodology and vocabulary. They don’t add tools, change capabilities, or give the agent access to data it didn’t already have. If you need a new tool, attach an MCP server. If you need a new agent personality with a tighter boundary, define a child agent. If you need to change the agent’s autonomy or budget, edit the agent directly on the Agents tab.
Use a skill when the model has the tools and access it needs but lacks the methodology or vocabulary to use them well.