building-kb-pipelines

Designs and operates FlashFusion's Zero-Admin KB Generation pipeline — automated ticket-to-article for MSP/ITSM environments. Use when scoping a KB automation engagement, designing the ticket-to-artic

SKILL.md

GitHub

Building KB Pipelines

FlashFusion's Zero-Admin KB Generation service: ITSM tickets in, draft KB articles out, human reviews before publish. Reference deployment — Loxihart MSP, ~90-second generation, ~$1–5/month on AWS Lambda + Bedrock. Architecture is stack-agnostic. This SKILL.md drives the session; deep content lives in the reference files it routes to.

Evals

Should trigger:

  1. "A client wants their resolved tickets to auto-generate KB articles. Help me scope it."
  2. "Design the ticket-to-article pipeline — what are the stages and where do the quality gates go?"
  3. "Build the KB automation phase plan with the pilot gate and kill criteria."

Should NOT trigger:

  1. "Write a KB article on resetting a VPN token." → single article on demand, use generating-kb-articles.
  2. "Score and rank the client's AI opportunities." → that's use-case-discovery, not pipeline design.

Principles (non-negotiable)

  1. Draft mode is non-negotiable. All articles publish to Draft. A human reviews before it goes live.
  2. Router quality gate first. Not every ticket becomes an article — the router decides.
  3. Benchmark Zero required. Measure current KB state, volume, and cost before building.
  4. Five sections required. Incomplete articles route to the KB GAP queue, never publish.
  5. PII redaction is Phase 0. Strip PII before generation, always.

Decision Tree — route to the right reference file

What's the task?
  ├─ "How does the pipeline work / what are the stages / router criteria?"
  │     → references/pipeline-architecture.md
  ├─ "What goes in an article / the five sections / validation rules?"
  │     → templates/article-schema.md
  ├─ "What's the rollout plan / phase gates / kill criteria?"
  │     → workflows/phase-plan.md
  ├─ "What stack / infra / cost / which LLM?"
  │     → references/stack-options.md
  └─ "Scope a new KB engagement end-to-end"
        → run the Golden Path below (touches all four)

Load only the reference file the current task needs. Do not dump all four.

Intake — collect before scoping

Ask for these. Do not design on blanks; insert [FILL: …] and flag it.

Field Why it matters
ITSM platform Webhook + API shape (e.g., Freshservice, ServiceNow, Jira SM). [FILL: client ITSM platform]
Ticket volume (resolved/month) Sizes cost, router throughput, and pilot duration
Existing KB state Benchmark Zero — article count, freshness, gaps
Article-worthy ticket types Seeds the router's include criteria
PII / data sensitivity Sets Phase 0 redaction scope and any compliance constraints
Cloud / stack preference AWS, Azure, or other; drives references/stack-options.md
Reviewer + publish owner Who clears the Draft queue, and how often
Definition of a "good" article Calibrates the validator and pass-rate gate

Golden Path — scope a KB engagement

  1. Intake — collect the table above. Capture the ITSM platform and ticket volume first; everything sizes off them.
  2. Benchmark Zero — document current KB count, freshness, top recurring ticket themes, and the cost of the status quo (analyst time writing articles manually). No baseline, no build.
  3. Define the router gate — from references/pipeline-architecture.md, set include/exclude criteria for which resolved tickets become article candidates.
  4. Lock the article schema — confirm the five required sections and validation rules from templates/article-schema.md. Confirm Draft-only publishing.
  5. Set Phase 0 PII redaction — define what gets stripped before any generation. [FILL: client-specific PII fields / compliance regime].
  6. Choose the stack — pick deployment from references/stack-options.md (reference: AWS Lambda + Bedrock). Estimate monthly cost from ticket volume.
  7. Build the phase plan — instantiate workflows/phase-plan.md with this client's gate thresholds and kill criteria.
  8. Produce the scoping deliverable — assemble the output template below and hand back.

Scoping Output Template

Deliver one document with these named sections:

  1. Engagement Header — client, ITSM platform, ticket volume, reviewer/owner, date.
  2. Benchmark Zero — current KB state, monthly article-worthy volume, status-quo cost. Each figure [SOURCED] or [GENERATED].
  3. Pipeline Design — the stage diagram and the router include/exclude criteria (from pipeline-architecture).
  4. Article Schema — the five sections + validation/quality gate; confirmation of Draft-only.
  5. PII & Compliance — Phase 0 redaction scope; any regulatory constraint. [FILL: …] where unknown.
  6. Stack & Cost — chosen stack, estimated monthly cost, scaling note.
  7. Phase Plan — table of phases with duration, gate threshold, and kill/extend criteria.
  8. Open Items — every [FILL: …] in one list.

Self-Verification Checklist

Before returning the scope:

  1. Intake table complete or every gap carries an explicit [FILL: …].
  2. Benchmark Zero present — no proposal on an unknown baseline.
  3. Router gate has explicit include AND exclude criteria.
  4. Article schema confirms five sections and Draft-only publishing.
  5. Phase 0 PII redaction defined (or flagged [FILL: …]).
  6. Phase plan includes a written kill criterion per gated phase.
  7. Cost figures labeled [SOURCED] or [GENERATED]; no invented pricing.
  8. SKILL.md routed to reference files rather than restating their full content.

CLAIMS

  • Reference deployment (Loxihart MSP, ~90s generation, ~$1–5/mo on AWS Lambda + Bedrock) is carried forward from the original skill. [SOURCED — prior skill version]
  • Draft-mode-only publishing with mandatory human review is a FlashFusion service principle. [GENERATED]
  • Exact router scoring thresholds, the specific Bedrock model, and PII redaction implementation are not specified here and must be supplied per client. [GENERATED — see [FILL] markers]

COUNTEREXAMPLE

  • Do NOT use for writing a single KB article (→ generating-kb-articles) or for opportunity ranking (→ use-case-discovery).
  • Do NOT auto-publish: a pipeline that skips the Draft queue violates Principle 1 regardless of pass rate.

CONTRADICTIONS

  • ! "Stack-agnostic architecture" vs. a concrete AWS Lambda + Bedrock reference — resolve by treating AWS as the reference implementation in references/stack-options.md, not a requirement.