Instructions & skills
Change what the agent knows, how it reasons, and when it reaches for a playbook.
Instructions
agent/instructions/ is concatenated in filename order into the system prompt.
The numeric prefixes are the ordering mechanism, so leave gaps when you add
files.
Two kinds are supported:
You are a research assistant. Search before you answer…Anything that is true for every user and every turn.
Behaviour problems are usually instruction problems
When the agent does something structurally wrong (announcing a plan it never
follows, skipping a research step, citing inconsistently) reach for the
instructions before you reach for the code. Most of these turn out to be a
missing sentence in 00-core.md, not a bug in the timeline renderer.
Adding one
Create agent/instructions/30-house-style.md:
## House style
Answer in British English. Lead with the conclusion, then the evidence.
Never present a single source as corroboration.Restart bun run dev afterwards. eve recompiles the agent on change, but
instruction files are read at compile time, not per turn.
Skills
Skills are Markdown playbooks the agent loads on demand. It sees only the
name and description until it calls load_skill, at which point the body enters
context. That's what lets a 2,000-word deep-research procedure exist without
paying for it on every "what's the weather" turn.
---
description: Profile a company's direct competitors with sourced claims.
---
## When to use this
The user names a company and asks who it competes with…
## Procedure
1. Establish the company's actual market from its own site…The frontmatter description is the only thing the agent sees before loading,
so it carries all the routing weight. Write it as a trigger condition ("the user
names a company and asks…"), not as a topic label ("competitors").
If a skill never fires, the description is almost always the cause. Confirm
with an eval that asserts load_skill was called for a representative prompt
before you rewrite the body.
Per-project instructions
Projects layer their own instructions on top of the global ones, edited in the
UI rather than in the repo. They're stored on the project table and prepended
for any chat started inside that project. That is the right place for
"this-client-only" rules that shouldn't ship in agent/.