A focused, source-grounded knowledge base for working with Salesforce Agentforce Agent Script, packaged as a Claude Skill.
When loaded into Claude, this skill gives the assistant accurate, current knowledge of Agent Script syntax, idiomatic patterns, common antipatterns, and the surrounding tooling (Agentforce Builder, Agentforce DX, Agentforce Vibes) — instead of relying on whatever happened to be in its training data about a young, fast-moving DSL.
Agent Script is Salesforce's declarative language for building hybrid-reasoning AI agents on the Agentforce 360 Platform. It went into Public Beta in November 2025 and has been moving fast since: what was called a topic until April 2026 is now a subagent, the recipe library grows weekly, and the tooling (Agentforce DX CLI, Agentforce Vibes) ships on a continuous cadence.
Out of the box, large language models — Claude included — have unreliable knowledge of Agent Script. Common failure modes I've seen in the field:
- Mixing
@topic.X(legacy) and@subagent.X(current) without flagging the rename. - Inventing a
then:action-chaining property without checking the spec. - Treating the slot-fill
...token as a variable default when it's only valid in action inputs. - Recommending the JSON-map
with { key: value }syntax instead of the canonicalwith key = valueform.
Each of those costs hours of debugging time inside the Builder. This skill
closes the gap by distilling the official Salesforce developer guide, the
Salesforce Developers Blog Agent Script Decoded series, and the
open-source salesforce/agentscript repository into a structured reference
that Claude consults whenever a conversation turns to Agentforce or Agent
Script.
skill/
├── SKILL.md # entry point — loads when skill triggers
└── references/
├── syntax-reference.md # complete language reference
├── patterns.md # idiomatic patterns
├── antipatterns.md # common mistakes and how to avoid them
└── builder-and-dx.md # Agentforce Builder, DX CLI, VS Code, Vibes
The skill follows the Anthropic progressive disclosure convention: the SKILL.md is small (~300 lines) and always loads when the skill triggers; reference files (~200–550 lines each) are pulled into context only on demand.
In scope. Agent Script language (blocks, expressions, hooks, variables,
transitions, action chaining), Hybrid Reasoning concepts at the script level,
Agentforce Builder authoring (Canvas / Script view, AI Assist), Agentforce DX
(CLI, VS Code extension), Agentforce Vibes integration, the topic →
subagent migration.
Deliberately out of scope (different skills, different concerns): Atlas Reasoning Engine internals, Data Cloud grounding, multi-org deployment strategy, Apex / Flow / Prompt-Template authoring, model selection.
- Download the latest
salesforce-agent-script.skillfrom the Releases page. - In Claude.ai, open Settings → Capabilities → Skills.
- Click Upload skill and select the
.skillfile.
The skill activates automatically when your conversation mentions Agentforce,
Agent Script, subagents, transitions, @utils.*, @variables.*, or related
concepts. You don't need to invoke it explicitly.
Skills can be loaded by reference in agent SDK contexts. See the Anthropic skills documentation for the current loading pattern in your environment.
Claude Skills use a three-tier loading model:
| Tier | Always in context? | Contains |
|---|---|---|
| 1. Metadata (name + description) | Yes | When the skill should trigger |
| 2. SKILL.md body | When triggered | Mental model, terminology, decision rubrics, quick reference, pointers |
| 3. Reference files | On demand | Full syntax, patterns, antipatterns, tooling |
The metadata description is the most important piece — it's what Claude matches against the user's intent to decide whether to consult the skill. This skill's description is engineered to trigger on Agentforce, Agent Script, and related vocabulary even when the user doesn't say "Agent Script" explicitly (e.g., "help me debug my subagent transition" or "how do I do slot filling").
The skill is a distillation of publicly available Salesforce documentation and source code. Primary sources:
- Agent Script developer guide
- Agent Script Reference index
- Agent Script Patterns
- Agent Script Recipes (sample apps)
salesforce/agentscript— open-source language tooling- Salesforce Developers Blog — Agent Script Decoded series
When the skill states a fact that I couldn't confirm against a canonical
source, the content is marked [Unverified]. This convention signals to
Claude that the claim should be re-checked against live documentation before
being relied upon — and signals to human reviewers where the gaps are.
Agent Script is moving fast. The content in this repository reflects the
state of Salesforce documentation on the date of each release (see
CHANGELOG.md). Between releases, expect drift on:
- New utilities in
@utils.* - New action target schemes (
flow://,apex://,promptTemplate://are stable; others may appear) sf agent ...CLI flags (the Salesforce CLI ships weekly)- Beta-to-GA promotion of individual features
The skill itself instructs Claude to verify against the live developer guide whenever syntax looks unfamiliar or when the Builder rejects something the skill claims should work. Pull requests with updated syntax, new recipes, and freshly observed antipatterns are welcome.
Planned additions, in rough priority order:
-
references/recipes.md— annotated walk-throughs of the most useful entries from theagent-script-recipesrepository. -
references/migration.md—topic→subagentmigration playbook for existing 2025-era Agentforce implementations. -
references/testing.md— deeper coverage ofsf agent test run, Testing Center assertions, and CI/CD integration. -
references/multi-org.md— patterns for sharing agents across Salesforce orgs (split out from the multi-org DevOps body of work).
Contributions are welcome — especially:
- Syntax corrections. If a fact in the skill doesn't match current Salesforce documentation, open an issue or pull request with the source link.
- New patterns that recur in real Agentforce work and aren't yet covered.
- Antipatterns observed in production. Anonymize as needed.
- Translations of the SKILL.md (English-only for triggering reliability, but reference files in other languages are an option).
For larger changes, please open an issue first to discuss scope before investing the time.
# Clone
git clone https://github.com/<your-username>/salesforce-agent-script-skill.git
cd salesforce-agent-script-skill
# Edit the markdown files
$EDITOR skill/SKILL.md skill/references/*.md
# Validate the skill structure
python3 -m scripts.quick_validate skill/
# Package into a .skill file
python3 -m scripts.package_skill skill/quick_validate and package_skill are part of Anthropic's
skill-creator tooling. Use whichever
local copy you have available.
Released under the MIT License. See LICENSE.
You are free to use, modify, redistribute, and adapt this skill for any purpose — commercial or otherwise — provided the copyright notice is preserved.
This is an unofficial, community-maintained project. It is not affiliated with, endorsed by, or sponsored by Salesforce, Inc. or Anthropic, PBC.
- Salesforce, Agentforce, Atlas Reasoning Engine, and related marks are trademarks of Salesforce, Inc.
- Claude and the Anthropic skill format are products of Anthropic, PBC.
The skill content is a distillation of publicly available Salesforce documentation, provided for educational and productivity purposes. Always verify critical syntax against the live Salesforce developer guide before shipping production agents.
Christopher Ramm DCX CTO Germany, Capgemini · Salesforce Certified Technical Architect · Salesforce MVP (Class of 2025)