Home / Docs / SfSkills documentation
SfSkills documentation
Who this is for: everyone, as a map. Find your section, then follow one link.
This page is an index only. It links and classifies, it does not teach. The sections run in the order a reader usually needs them: get it working, understand what it is, see the build loop in action, look things up, install it properly, run it day to day, and finally change it.
| Label | Means |
|---|---|
Consumer | You are USING the library on a Salesforce project: searching skills, running agents, wiring the MCP server into your editor. You never edit this repo. |
Contributor | You are CHANGING the library: adding or revising a skill, an agent, a template or the tooling. You run the sync and validation gates. |
Both | Useful either way. |
1. Getting started
| Doc | Audience | What it answers |
|---|---|---|
| ../README.md | Both | What is this and why would I want it? |
| getting-started.md | Consumer | Install to first useful answer, for the three real entry points: Claude Code checkout, MCP server, plain export to another tool. |
| worked-example-trigger-consolidation.md | Consumer | One complete Salesforce task, start to finish, with the real command output. |
2. Concepts
| Doc | Audience | What it answers |
|---|---|---|
| architecture.md | Both | How skills, agents, commands, templates, decision trees, registry, index, evals and the MCP server fit together, and which of the three retrieval mechanisms each accuracy figure describes. |
| agent-invocation-modes.md | Consumer | The ways to invoke an agent, and why MCP is the canonical channel for production use. |
| glossary.md | Both | This repo's own vocabulary: mechanism 1/2/3, gloss, roster, tier, coverage gate, envelope. |
| ../evals/measurement/README-model-routing.md | Both | How the shipped routing path is benchmarked, and the retraction of the "79.2% → 92.2% Hit@1" headline. Read before citing any routing number. |
3. The build loop
Asked to make a Salesforce change (a field, a flow, a rule, an integration, a whole process) rather than answer a question? This is the path. The "Two Ways In" section at the top of ../CLAUDE.md says the same to an AI session.
| Doc | Audience | What it answers |
|---|---|---|
| build-loop.md | Consumer | Start here. The requirement-to-build loop told through the five worked examples, smallest first, with every number sourced. |
| ../examples/builds/README.md | Both | The five scenarios side by side, how each column was counted, every org refusal and where it lives now, and what validate-only could not see. |
| ../standards/build-orchestration.md | Both | The contract: stages, gates, the plan file, the sizing rule (§ 3.1), steps, acceptance tests, roles. The authority when anything else disagrees. |
| ../commands/build-from-requirements.md | Consumer | The exact command sequence, gate by gate. |
The worked examples, each a folder with its own README.md:
| Example | Size, from its README | Read it for |
|---|---|---|
| opp-amount-lock | ask tier: one-line ask, 13 questions, 1 step, two gate decisions | The loop at its smallest; one rendered RUN.md read at both gates |
| tier2-webhook | feature tier: integration, 45 questions, 5 steps, 13 validate-only runs | Why compile-only Apex evidence is not enough |
| case-onboarding | project: 97 clarifications, 5 milestones, 22 steps, 13 gates | The largest build; the org as the last reviewer |
| northwind-sales | project: 16 steps, 4 milestones, 9 gates, 16 runs reached the org, 14 org-taught facts | The broadest build: config, access, approvals, Apex, LWC, reporting |
| cold-start-lead-source | no tier: the loop was never found | What a session that skipped the loop missed |
| cold-start-case-escalation-email | feature tier: 19 questions, 1 step | The same experiment after the fix: found, sized and driven to done |
4. Library reference
| Doc | Audience | What it answers |
|---|---|---|
| SKILLS.md | Both | The full skill catalog, by domain. Generated. |
| agents.md | Both | Every agent package with its status and role. Generated. |
| ../agents/_shared/RUNTIME_VS_BUILD.md | Consumer | Which agents do Salesforce work and which maintain the library. |
| ../agents/_shared/SKILL_MAP.md | Consumer | Which agent cites which skills. |
| ../standards/decision-trees/README.md | Consumer | Routing before technology choice, across seven trees: Flow vs Apex, flow pattern, Agentforce capability, async tier, integration pattern, sharing mechanism, performance tuning. |
| ../templates/README.md | Consumer | The canonical Apex, LWC, Flow and Agentforce building blocks that skills point at. |
5. Installing
| Doc | Audience | What it answers |
|---|---|---|
| installing.md | Both | The canonical setup reference for a clone: one bootstrap command, every flag, what a clone does and does not contain, embeddings cost, MCP install paths, and the maintainer runbook for cutting a release. |
| installing-the-plugin.md | Consumer | Install the library as a Claude Code plugin from the marketplace, and what the plugin costs at session start. |
| installing-single-agents.md | Consumer | Ship one agent into another project without dropping its skill and probe dependencies. |
| ../mcp/sfskills-mcp/docs/CONNECT.md | Consumer | MCP client config for Claude Code, Claude Desktop, Cursor, Windsurf, Zed, VS Code, Cline, Continue, Codex CLI, Gemini CLI, Goose. |
| ../mcp/sfskills-mcp/README.md | Consumer | The MCP tool schemas, annotations and design notes. |
6. Operating
| Doc | Audience | What it answers |
|---|---|---|
| consumer-responsibilities.md | Consumer | What a consuming tool MUST do when it runs a run-time agent: persist reports, honour the JSON envelope. |
| multi-ai-parity.md | Consumer | Which export targets are first-class and what each one loses. |
| troubleshooting.md | Both | Symptom to cause to fix, for the failure modes a fresh clone actually hits. |
| faq.md | Both | Do I need an org? Why is search slow? Why do the CLI and MCP disagree? |
| MIGRATION.md | Both | Which agents were retired in the Wave 3 consolidation, what replaced them, and the state of each redirect. |
| validation/README.md | Both | How the library verifies itself against a live org: three re-runnable harnesses. The harnesses ship; their reports do not. |
| ../SECURITY.md | Both | Threat model, secret handling, and how to report a vulnerability. |
| ../CHANGELOG.md | Both | What changed, when. |
7. Contributing
| Doc | Audience | What it answers |
|---|---|---|
| ../CONTRIBUTING.md | Contributor | Add a skill, fix a skill, report a gap, flag stale content. |
| ../AGENT_RULES.md | Contributor | The full repo-wide workflow rules. |
| ../CLAUDE.md | Contributor | The rules an AI assistant follows inside this repo. |
| ../AGENTS.md | Contributor | The agent-facing entry point (the AGENTS.md convention). |
| ../standards/skill-authoring-style.md | Contributor | How a skill package is written: voice, section shapes, Questions to Ask Before Configuring. |
| ../standards/validation-gates.md | Contributor | Every gate validate_repo.py enforces, with file and line citations. Generated. |
| ../standards/official-salesforce-sources.md | Contributor | The official Salesforce documentation a skill's claims must be grounded in. |
| ../agents/_shared/AGENT_CONTRACT.md | Contributor | The 8-section shape every AGENT.md must have. |
| ../evals/README.md | Contributor | Golden P0 output-quality cases for the flagship skills. |
| ../commands/onboard-source.md and source-integrations/README.md | Contributor | Onboarding an external skill source, and the licence and chain-of-title record each one leaves. |
Generated artifacts: never hand-edit
Regenerate with the command in the third column. Only some of these are gated.
| Doc | What it is | Regenerate with | Drift gated? |
|---|---|---|---|
| SKILLS.md | The skill catalog | scripts/skill_sync.py --all (via scripts/generate_docs.py) | Yes. validate_repo.py recomputes it through pipelines/sync_engine.py and errors on any difference. |
| agents.md | The agent roster | scripts/generate_agent_roster.py | Only by hand: --check exists, and no workflow or hook runs it. |
| queue-progress.md | Backlog dashboard: status counts, drift, next pick | scripts/generate_queue_dashboard.py | Yes, when BACKLOG.yaml produces a dashboard. |
docs/reports/duplicate-candidates.md | Near-duplicate skill pairs | scripts/audit_duplicates.py | No. docs/reports/ is gitignored, so it exists only after you run the script. |
Outside docs/, registry/, vector_index/chunks.jsonl, vector_index/manifest.json and standards/validation-gates.md are written by the sync engine and drift-gated by validate_repo.py. The plugin artifacts under .claude/ (routers, rosters, agent loaders) are generated by scripts/build_plugin.py and checked in CI by scripts/build_plugin.py --check (.github/workflows/validate.yml).
The queue itself is hand-authored: ../BACKLOG.yaml holds the rows, and ../MASTER_QUEUE.md is the workflow contract for claiming one. queue-progress.md is derived from the first.
What this library is not
- Not a deployment tool. Nothing here pushes metadata to an org, and the build
loop stops at a validate-only command a human runs.
- Not an org scanner on its own. Reading an org needs the MCP server plus your
own authenticated Salesforce CLI session.
- Not org-dependent for skills. Search, agents, templates and decision trees
work with no Salesforce org at all.
Not indexed
docs/product-v2/anddocs/source-integrations/2026-09-01-public-skill-source-audit.mdare historical: dated records of earlier work sessions, kept for provenance, not maintained as documentation.
docs/reports/and everything underdocs/validation/except its README aregitignored run output: per-agent reports, validation runs, dated one-off analyses. They describe the tree as it was when they ran.