Skills overview
Markdown playbooks that teach the agent how you want a job done, loaded only when relevant.
A skill is a Markdown file that gives the agent step-by-step guidance for a specific kind of work. Excelente ships eleven, you can write your own, and the agent can propose one from work you have already done together.
Skills follow the Open Agent Skills format, so a skill written for Excelente is readable by other agent harnesses that support it.
Why they exist#
A general model asked to "underwrite this deal" makes reasonable choices that are not your choices. It picks a cap rate convention, decides whether to net reserves above or below NOI, and formats a schedule its own way. A skill fixes those decisions once, so the agent arrives already knowing how your shop does it.
Skills carry method. Workbook memory carries facts about one file. See Workbook memory.
Progressive disclosure#
A skill has a body and optional reference files. Only short summaries of your installed skills reach the system prompt. The agent loads a skill's body when the work calls for it, and pulls individual reference files after that if it needs them.
That layering is what makes a large skill library affordable. Twenty installed skills cost roughly twenty short descriptions per turn, not twenty full playbooks.
load_skill and read_skill_resource results are exempt from compaction, so a loaded skill survives
a long build.
Turning skills on#
Three ways, all equivalent:
- Capabilities → Skills, and tick the checkbox.
- The + button in the composer, under Skills.
- Type
/and pick the skill by name, which enables it for the conversation.
The / route is the fastest mid-conversation. Type /rent, pick rent-roll-standardizer, carry on.
How the agent finds one itself#
The agent has a find_skill tool and will reach for a skill you have enabled without being told.
Matching is weighted: a skill's when-to-use counts three times as much as its description, and
its description twice as much as its name.
This is why a good when-to-use matters more than a good name. Write it as the situations a reader
would recognize, not as a restatement of the title. See
Write your own skill.
Core skills you never see#
Three bundled skills are hidden from the Skills panel, the + menu, and the / menu. The agent
reaches them on its own:
| Skill | Role |
|---|---|
cre-modeling-conventions |
A.CRE house style: layout, color coding, number formats, formula discipline. Injected in full into every system prompt, so it shapes every build whether or not you asked |
office-js-patterns |
The mental model and footguns for writing run_excel_script code |
verify-model-outputs |
The Reviewer sub-agent's playbook for checking a finished build |
If your agent color-codes inputs blue and formulas black without being asked, that is
cre-modeling-conventions.
Inspecting a skill#
In the Skills panel, Show expands a skill to reveal its description, when-to-use, author,
version, and source, plus a file tree. Click SKILL.md or any reference file to read its contents in
the panel.
Worth doing before you trust a skill with real work. A skill is a prompt, and reading it tells you exactly what the agent has been told.
Installing your own#
Upload skill (.zip / .skill / .md) in the Skills panel, or drop a file onto it. Uploaded skills can be uninstalled with the ✕ on their row. Bundled skills cannot be removed, only switched off.
Details in Write your own skill.
What ships#
See Bundled skills for the eleven, what each covers, and their licensing.