Deloitte Salesforce Practice Toolkit

A shared knowledge base and agent framework for Salesforce implementations, powered by Claude Code.

What This Is

An npm-installable toolkit that provides AI-powered Salesforce implementation assistance:

• 5 Specialized Agents — architect, developer, QA, code reviewer, communications
• 21 Skills — firm standards, cloud-specific patterns, quality checklists, review workflows
• Templates — solution designs (standard, Agentforce, Data Cloud), task specs, runbooks, project scaffolds
• Learning Loop — automated capture, admin-gated promotion of lessons learned
• Checklists — security review, go-live readiness

Requirements

• Node.js 18+
• Claude Code CLI installed and authenticated

Installation

Recommended (all platforms)

npm install -g @deloitte-digital-il/sf-toolkit
dsf install

Clone-and-run

If you have cloned this repo and want to install without going through npm, use the platform-native script:

# macOS / Linux
./install.sh

# Windows (PowerShell 5.1 or 7)
powershell -ExecutionPolicy Bypass -Scope Process -File .\install.ps1

The -ExecutionPolicy Bypass -Scope Process flag affects only the current PowerShell process; it does not change machine policy. No admin elevation required.

This installs:

• 5 agents to ~/.claude/agents/ (available in all Claude Code sessions)
• 6 slash commands to ~/.claude/commands/
• Skills symlinked to ~/.claude/skills/deloitte-sf-toolkit/
• Templates, learnings, and checklists symlinked for access
• ~/.claude/CLAUDE.md updated with toolkit reference

After installing, restart Claude Code (or start a new session).

CLI Commands

dsf install      # Install/update agents, skills, and commands
dsf uninstall    # Remove all toolkit components from Claude Code
dsf update       # Check for and install the latest version from npm
dsf status       # Show installation status, versions, and component health
dsf todos        # List TODO items that need your firm's standards
dsf migrate      # Migrate a v2.0.x project to v3.0.0 (backup + idempotent)
dsf help         # Show help

Migrating a v2.0.x project to v3.0.0

Run dsf migrate from the project root. The runner snapshots every declared path to .dsf/migrations/<timestamp>/backup/ before any mutation, then applies the v2_to_v3 transform. The migration is idempotent (safe to re-run) and reversible (dsf migrate --rollback <timestamp>).

For the per-transform breakdown, the backup tree shape, the --rollback and --list-backups flags, and troubleshooting, see docs/migrations/v2-to-v3.md.

Updating

dsf update       # Downloads latest version from npm
dsf install      # Applies the update (copies agents/commands, refreshes symlinks)

Uninstalling

dsf uninstall                                      # Remove toolkit from Claude Code
npm uninstall -g @deloitte-digital-il/sf-toolkit   # Remove npm package

Quick Start

claude /dsf:scaffold                              # 1. create a project
cd <project> && claude                            # 2. start a session in it
/dsf:discovery "case management with SLA tracking" # 3. define requirements (prints SPEC-NNN)
/dsf:design <SPEC-id from step 3>                 # 4. design + decompose (prints TASK-NNN per task)
/dsf:develop <TASK-id from step 4>                # 5. build + review
/dsf:qa <TASK-id>                                 # 6. test
/dsf:release                                      # 7. pre-deployment checks

For Agentforce projects, /dsf:design runs the canonical sf agent generate agent-spec DX flow when Agentforce ∈ clouds_in_scope. /dsf:develop runs sf agent create to materialize the bundle; /dsf:qa runs sf agent test create + sf agent test run.

Other commands: /dsf:review, /dsf:document, /dsf:status, /dsf:diagnose, /dsf:estimate, /dsf:eval. Run dsf guide for the full workflow.

Agents

| Agent | Model | Purpose | |-------|-------|---------| | sf-architect | Opus | Solution design, architecture decisions, ADRs, task specs | | sf-developer | Sonnet | Code generation from task specs (Apex, LWC, Flows) | | sf-qa | Sonnet | Adversarial testing, defect reporting, test planning | | sf-reviewer | Opus | Code review, standards enforcement, design conformance | | sf-comms | Sonnet | Client emails, status updates, meeting summaries |

Agents auto-delegate based on context, or you can invoke them explicitly. They communicate through structured files in the project repo (docs/architecture/, docs/qa/, docs/reviews/, docs/learnings/inbox/).

Skills

Skills provide domain-specific knowledge that agents read before performing work.

| Category | Skills | |----------|--------| | Foundation | firm-standards, architecture-decisions, project-scaffolding, session-management | | Development | apex-development, lwc-development, flow-development, integration-development, soql, connected-apps | | DevOps | deployment | | Clouds | Agentforce triad — developing-agentforce/ (AScript reference, agent-spec YAML, action authoring), testing-agentforce/ (test specs, preview, async test runs, CI), observing-agentforce/ (STDM, traces, Observe → Reproduce → Improve loop) — plus data-cloud, marketing-cloud | | Quality | code-review, testing, debugging, documentation, diagrams, permissions-audit, architecture-review, review-and-revise, iterative-feedback |

Each skill has a TLDR for quick relevance assessment and a references/ directory with distilled knowledge docs (anti-patterns, patterns, checklists).

Project Structure

├── CLAUDE.md                  # Master config — standards, skill routing
├── claude-progress.md         # Cross-session progress tracking (per project)
├── .claude/
│   ├── agents/                # Agent definitions (5 agents)
│   └── commands/              # Slash commands (scaffold, document, etc.)
├── skills/                    # Domain knowledge (SKILL.md + references/)
│   ├── foundation/            # Standards, architecture, scaffolding
│   ├── development/           # Apex, LWC, Flow, integration, SOQL
│   ├── devops/                # Deployment
│   ├── clouds/                # Agentforce, Data Cloud, Marketing Cloud
│   └── quality/               # Review, testing, debugging, docs
├── templates/                 # Document and project templates
│   ├── solution-design/       # Standard, Agentforce, Data Cloud
│   ├── task-spec/             # Standard, Agentforce
│   ├── runbook/               # Standard runbook
│   └── project-scaffold/      # SFDX project files
├── mcp-configs/               # Per-agent MCP server configurations
├── learnings/                 # Lessons learned system
│   ├── promoted/              # Admin-approved lessons (read by agents)
│   └── review-queue/          # Entries awaiting admin review
├── checklists/                # Go-live, security review
├── evals/                     # Agent evaluation framework
│   ├── tasks/                 # Eval scenarios per agent
│   ├── graders/               # Grading rubrics
│   └── results/               # Eval run results
├── bin/                       # CLI (dsf command)
└── package.json               # npm package definition

Salesforce DX MCP Integration

Agents interact with Salesforce orgs through the Salesforce DX MCP server (@salesforce/mcp).

Prerequisites:

1. Install the Salesforce CLI: https://developer.salesforce.com/tools/salesforcecli
2. Authenticate your org: sf org login web --alias my-org
3. Run /dsf:scaffold — it generates .mcp.json with the correct DX MCP configuration

What agents can do via MCP:

• Retrieve and inspect metadata (retrieve_metadata)
• Query org data (run_soql_query)
• Run Apex tests (run_apex_test) and agent tests (run_agent_test)
• Deploy to sandboxes (deploy_metadata) — never directly to production
• Run static code analysis (run_code_analyzer)

Configuration:

• Project-level .mcp.json is the canonical runtime config
• Role-specific toolset examples are in mcp-configs/ (not active at runtime)

Customization

After installing, run dsf todos to see placeholders that need your firm's actual standards:

• Trigger framework interface code
• Service/Selector examples
• Error handling patterns
• TestDataFactory
• API version numbers

Edit CLAUDE.md and skills/foundation/firm-standards/SKILL.md to customize.

Adding a New Skill

1. Create a directory in skills/[category]/ (kebab-case)
2. Add SKILL.md with YAML frontmatter (name + description required)
3. Include a > **TLDR:** line after frontmatter for quick relevance assessment
4. Add references/ for detailed knowledge docs
5. Update the skill routing table in CLAUDE.md

Reviewing Learnings

Agents write learning entries to docs/learnings/inbox/ during their work. Review and promote them:

claude /review-learnings

License

UNLICENSED — Deloitte Confidential