Build

Skills & Prompt Library

Two small libraries of reusable text: skills are focused playbooks composed into an agent's system prompt at run time; prompts are complete starter system prompts. Both exist so good instructions get written once and reused everywhere.

Skill Library

Skills at /skills follow the Anthropic-style skill.md pattern: a skill is a focused, reusable playbook — "how to write a structured code review", "how to cite sources from a knowledge base" — that you attach to any agent or swarm node. At run time every selected skill is composed into the agent's effective system prompt, so the same agent can be specialized for different jobs just by swapping skills.

  • My skills — skills you write yourself in the editor, or generate with AI from a plain-English description and then edit.
  • Sample skills — a curated set you can copy into your own library with one click and adapt.
  • Attaching — open an agent in the Agent Builder (or a swarm node inspector) and select skills from the picker. Deleting a skill is safe: agents that referenced it simply drop it.

Prompt Library

The Prompt Library at /prompts is a searchable collection of complete system prompts, filtered by category — support, sales, engineering, research, data, writing, creative, operations, productivity, and education. Each entry has a preview so you can read the full prompt before using it, and you can save your own prompts to the library for reuse across agents.

Prompt Compare

/prompt-compare answers the most practical question in prompt engineering: given this exact prompt, how do different models behave? It runs the same prompt against two or three models side by side, streaming the outputs next to each other with per-panel latency, token counts, and real cost. Preset experiments are included — constraint-following, JSON-only output, and similar discriminating tasks — each with a note on what to watch for.

Rule of thumb: if the reusable thing is a complete identity ("you are a support agent for…"), it belongs in the Prompt Library. If it is a technique an agent should apply on top of its identity ("when reviewing code, always…"), it is a skill.

Fields on a skill

FieldRequiredNotes
NameYesHow you find it in the picker
DescriptionNoWhat it is for
bodyYesThe instruction text itself. This is what gets prepended to the agent's prompt.
TagsNoFor filtering a long library

Attached skills are resolved at run time and prepended as a "Skills available to you" block ahead of the agent's own system prompt. Both saved agents and individual swarm nodes can attach them.

Why it works this way

A skill is instructions, not code — which is exactly why it is cheap to add. If a capability only needs the model to know a procedure ("how we format a change request"), it is a skill. If it needs to reach something outside the model, it is a tool, and tools need a handler and a permission gate.

Writing one that works

Skill body — change request format
When asked to write a change request, always produce exactly these sections:

TITLE      one line, imperative
RISK       low | medium | high, with one sentence of justification
ROLLBACK   the exact steps to undo this, or "none — irreversible"
BLAST      what breaks if this goes wrong

Never omit ROLLBACK. If a change genuinely cannot be undone, say so
explicitly rather than leaving the section out.
  • Be procedural. A skill is at its best describing a repeatable format or sequence, not general advice.
  • Keep it short. Every attached skill is prompt tokens on every turn.
  • One skill, one job. Two loosely-related skills beat one that covers both, because you can attach them independently.
PreviousSwarm CanvasNext Developer workspace