Skills
Capabilities your agents reach for.
Package instructions, triggers, and tool chains into reusable Skills. The right one activates itself the moment a conversation calls for it — versioned, measured, and shared across every bot in your org. Start from 50 prebuilt templates.
Write it once, reuse everywhere
A skill is a package, not a copy-paste. Build a refund flow once and every support bot in your org inherits it — update it in one place and the change propagates.
Loads only when it's needed
Skills activate themselves. The match engine scores each turn and injects just the relevant skills — no bloated mega-prompt, no skills firing where they don't belong.
Proves it earns its keep
Every activation is measured — success rate, confidence, how often it fires. Keep the skills that work, retire the ones that don't, and roll back any regression.
How it works
The right skill, the moment the conversation needs it.
Every turn, the match engine scores your skills against the live message, activates the top performers, and injects their instructions into the prompt — for that turn only.
Skill activation
message → match → activate → injectTrigger match
Score & rank
Activate top-N
Inject inline
Define
Bundle instructions into a reusable package
A skill is more than a prompt. It packages system instructions, natural-language trigger conditions, an optional tool chain, and effectiveness metrics into one reusable unit. Scope it org-wide so every bot inherits it, or pin it to a single bot.
How it works
- System instructions injected on match — not a static system prompt
- Org-level (every bot) or bot-specific scoping
- Four types: prompt template, tool chain, code snippet, workflow
Trigger
Describe when it should fire, in plain language
Each skill carries natural-language trigger conditions — "refund", "talk to a human", "too expensive". Keywords are extracted automatically for fast lookup, and the reviewer engine can suggest new triggers as it learns from real conversations.
How it works
- Plain-English trigger phrases, no regex required
- Auto-extracted keywords for fast pre-filtering
- Reviewer engine proposes new triggers from live traffic
Match
Lexical + semantic scoring, every turn
On every turn the match engine scores the user's message against each skill — 40% from keyword overlap, 60% from trigger-condition hits. Turn on semantic matching and cosine similarity blends in, so paraphrases the keywords never saw still fire the right skill.
How it works
- score = 0.4 · keywords + 0.6 · trigger hits
- Optional semantic embedding blend for paraphrase coverage
- Identical scoring in the Python engine and the dashboard preview
Activate
Top skills injected into the live prompt
Skills above the threshold are sorted by score, capped at your limit, and injected into the agent's system prompt under an ACTIVATED SKILLS block — for that turn only. No bloated mega-prompt, no skills firing when they shouldn't.
How it works
- Score-ranked, capped per turn — only what's relevant loads
- Injected inline alongside KB chunks and guardrails
- Every activation traced with its match reason and score
Measure
Effectiveness metrics + versioning
Every match writes a usage event — outcome, confidence, and score — aggregated into per-skill counters. See which skills earn their keep, which mis-fire, and roll back any skill to a prior version in one click when a change regresses.
How it works
- Per-skill success/failure, confidence, and activation counts
- Full version history with one-click rollback
- Draft → Active → Disabled → Archived lifecycle
Anatomy of a skill
one package · six partsRefund Policy
prompt template · active
runtime-readyTriggers
"refund", "money back", "cancel order"
Instructions
Ask for the order #, check the window…
Tool chain
lookup_order → check_eligibility
Metrics
412 fires · 94% success
Scope
Org-level · all support bots
Version
v3 · active · rollback ready
Four kinds of skill
From a single instruction to a multi-step procedure.
Prompt Template
Runtime-readyA reusable instruction pattern injected into the system prompt the moment its trigger matches. The workhorse skill type — fully live in the engine.
Tool Chain
Schema-readyAn ordered sequence of tool calls with parameter mapping — outputs of one step feed the inputs of the next, with optional conditions between them.
Code Snippet
Schema-readyA programmatic capability executed in a single turn — for deterministic logic the LLM shouldn't be guessing at.
Workflow
Schema-readyA multi-step procedure with conditional branches — the heaviest skill type, for end-to-end procedures that span several turns.
Prompt templates run live in the engine today. Tool chains, code snippets, and workflows are modelled end-to-end in the schema and surface in the UI — ready as the runtime lights each one up.
Under the hood
Matching you can read, a lifecycle you control.
How matching is scored
Deterministic, traceable, and identical across runtimes.
score = 0.4 · keywords + 0.6 · trigger hits
activate when score ≥ 0.25 · rank desc · cap at max_skills
Keyword overlap
Case-insensitive substring match against auto-extracted trigger keywords — the fast first pass.
Trigger hits
Quoted fragments and phrase splits from the skill's natural-language conditions, weighted highest.
Semantic blend
Optional: cosine similarity between the message and each skill's embedding catches paraphrases.
Skill lifecycle
Stage every change before it reaches production.
Draft
Build and preview
Active
Live on matching turns
Disabled
Paused, kept intact
Archived
Retired, still versioned
Six sources
Skills you build, and skills that build themselves.
Author them by hand, install from the catalog, or let the platform promote them automatically from patterns it finds in real conversations.
Prebuilt
Installed from the prebuilt catalog and forked into your own editable copy.
Manual
Authored by hand in the dashboard for capabilities unique to your business.
From rules
Promoted automatically when a cluster of learned rules earns its own skill.
Architect
Generated by the Workspace Architect while it bootstraps a new bot.
Reviewer-suggested
Proposed by the reviewer engine after it spots a recurring pattern in traffic.
Imported
Brought in from another bot, another org, or the hub.
Prebuilt catalog
50 ready-to-install templates, 15 industries.
Install any template to fork your own editable copy — triggers, instructions, and tags included. Customize it, version it, and measure it like any skill you wrote yourself.
Customer Experience
10 templatesCustomer Support · Conversation Design · Knowledge & Escalation
Sales & Commerce
8 templatesSales · E-commerce
Compliance & Risk
6 templatesCompliance · Finance & Insurance
Industry Solutions
9 templatesHealthcare · HR & Internal · Real Estate & Education
Technical & Orchestration
8 templatesIT & Technical · Multi-Agent
Content, Data & Utilities
9 templatesContent & Marketing · Data Collection · Utilities
Answers
What teams usually ask
How is a skill different from a prompt?
A prompt is text. A skill is a package — instructions plus the trigger conditions that decide when those instructions load, an optional tool chain, scope, and effectiveness metrics. Skills activate themselves; prompts are always-on.
What decides which skill fires?
Every turn the match engine scores the user message against each skill: 40% from keyword overlap and 60% from trigger-condition hits. Anything above the threshold activates, ranked by score and capped per turn. Semantic matching can blend in embedding similarity so paraphrases fire too.
Won't loading skills bloat the prompt?
No. Only skills that match the current turn are injected, ranked by score and capped at your limit. An ACTIVATED SKILLS block carries just the relevant ones — the rest stay dormant.
Can I undo a bad change?
Yes. Every save creates a version. Roll back to any prior version in one click, and use the Draft / Active / Disabled / Archived lifecycle to stage changes before they reach production.
Deep dive
Why agents deserve a skill library, not a longer prompt
Context that arrives just in time
Stuffing every rule into one system prompt dilutes the model and burns tokens on instructions that don't apply. Skills invert that: each capability declares when it's relevant, and only the matching ones load. The agent reads the refund procedure precisely when someone asks for a refund — and nothing extra the rest of the time.
Matching that's the same everywhere
The Python engine that serves live channels and the dashboard preview run the identical scoring formula. What you see fire in the playground is exactly what fires in production — no “works on my machine,” no drift between test and live, every activation traced with its score and the reason it matched.
Capabilities you can measure
A prompt edit is a guess. A skill is a unit you can score. Every match records an outcome and a confidence, aggregated into per-skill counters, so you can answer “is the de-escalation skill actually helping?” with data — and roll back the version that made it worse.
A library that grows on its own
Skills don't only come from you. When a cluster of learned rules proves itself, it gets promoted into a skill. When the reviewer engine spots a recurring pattern, it suggests one. Your library compounds as your bots run — and the 50-template catalog gives you a running start.
Give your agents a toolkit
Build skills your bots reach for on their own.
Install a template in a click, or author your first skill in minutes. Set the triggers, watch it activate on the right turns, and measure it improving release over release.