Salesforce AI Skills Programme
An end-to-end agentic Slackbot suite that automates the Document Development Life Cycle, from multi-source PRD analysis and automated UI validation to generating structured, DTD-compliant Oxygen XML.
Scaling documentation across a fast-moving product org
In enterprise software, the Document Development Life Cycle spans multiple source systems: Quip, Google Docs, live Confluence pages, and engineering discussions in Slack. Keeping documentation accurate at the pace of product development requires a different approach than traditional workflows allow.
The opportunity was to build a system that handles the research, synthesis, and validation phases automatically, so writers could focus on content strategy and quality rather than source-chasing and boilerplate.
Multi-source context
PRD content distributed across Quip, GDocs, Confluence, and Slack, unified by the pipeline.
Speed-to-publish
Closing the gap between product ship cadence and documentation delivery.
Review efficiency
Automating validation steps to reduce editorial cycles and SME round-trips.
An agentic Slackbot pipeline
I designed and shipped a suite of ten production AI Slackbot skills operating as an orchestrated multi-agent pipeline. Running directly in the writer’s natural workspace, Slack, the system manages the end-to-end DDLC through five specialised, autonomous stages.
- Multi-source RAG and MCP sweeps across Quip, GDocs, live Confluence pages, and Slack channels.
- Scope reasoning: flags deviations and detects undocumented dependencies.
- Drafts Slack-ready messages for PMs and devs; triggers Human-in-the-Loop questions for SMEs.
- Generates a Slack Canvas with synthesized context, provenance tracking, and confidence scoring.
- Ingests company style guidelines and target deliverable formats such as Release Notes.
- Enforces voice, tone, and structural constraints for immediate publishing readiness.
- Uses Playwright in a secure sandbox to test UI labels and procedural steps against the live product.
- Provides validation reports and auto-updates drafts when gaps are found.
- Integrates directly with GUS, Salesforce’s JIRA equivalent.
- Creates work items for sprints, estimates story points, and suggests actionable ticket comments.
- Converts approved Slack text into production-ready Oxygen Author XML.
- Strictly validated against DITA and DTD constraints for CI/CD integration.
- Connects natural language generation with enterprise publishing requirements.
Beyond text generation, knowledge engineering
The architecture is defined by three capabilities that move well past simple LLM prompting.
Orchestrated multi-agent logic
Chaining specialised skills, Research, Governance, UI Testing, and Lifecycle Management, that function in tandem as an autonomous pipeline.
Structured XML generation
Connecting LLM generation with strict enterprise publishing requirements via DITA and DTD validation.
Integration maturity
MCP for live Confluence retrieval and Playwright for real-time UI validation: production-grade AI managing docs against live products.
System efficiency gains
The pipeline delivers a fundamental shift in technical writing velocity by automating the research and validation phases upstream of the draft. Writers move into content strategy and orchestration roles.
Automating multi-source research and synthesis eliminated the manual discovery phase entirely.
AI governance and UI validation pre-validate content before it reaches SME review.
Deployed across a writing team of eight, each skill covering a distinct DDLC phase.
Became a sought-after framework by Lead and Staff cloud writers across the org.
Beyond the Expo rating, this toolsuite generated organic demand, becoming a sought-after framework by Lead and Staff-level cloud writers across the organisation.
What this project validates
- 1The highest-value AI intervention in technical writing is not generation. It is automating the research and validation phases upstream of the draft.
- 2Slack as the deployment surface removes adoption barriers. Writers do not change their workflow; the AI meets them where they already work.
- 3Output constraints such as DTD validation are what separate a prototype from a production tool. LLM creativity must be bounded by publishing requirements.
- 4Human-in-the-Loop checkpoints at the right junctions, not everywhere, are the key to trust and adoption in enterprise environments.