Decide · Enterprise Skill Library
The LLM reasons. Skills carry the SOP.
The LLM understands the goal. Skills encode how the work is done. The CWM predicts what will happen. Flash executes and learns.
A general model can reason about anything — which is exactly why an enterprise can't hand it the keys raw. A Skill is a reusable, verifiable, versioned unit of business work: the play, its steps, its caps and its stop conditions, written down once and governed like the asset it is. Skills are how Flash constrains general LLM reasoning into your standard operating procedure.
Rolling out with the governed agent stack.
The five-layer execution stack
LLM Reasoning
L1understands the goal, weighs the evidence
Skills · the SOP layer
L2encode how the work is done — steps, caps, gates
Consumer World Model
L3the member state & predictions skills condition on
Enterprise Tools
L4governed, audited actions — never raw APIs
Feedback & Learning
L5outcomes flow back — skills evolve on evidence
↻ L5 feeds L1–L3 — every outcome sharpens the next run
Why a skill layer
Between the goal and the tools sits the way you work.
An LLM alone improvises; an API alone is blind. The skill layer is what sits between them — so what runs is not “whatever the model came up with today,” but a named procedure your team can read, version and approve.
Layer 1
LLM Reasoning
Understands intent, reads evidence, makes the judgment calls a procedure leaves open — nothing more.
Layer 2
Skills
The SOP: ordered steps, tool allowlist, caps, gates and stop conditions — the part that must not be improvised.
Layer 3
Consumer World Model
The live member state — churn risk, purchase timing, predicted value — that skills declare they need before they run.
Layer 4
Enterprise Tools
Governed, audited platform actions. Agents reach them only through a skill — never as raw API calls.
Layer 5
Feedback & Learning
Outcomes are measured against each skill's declared metrics, and upgrades ride the version lineage — on evidence.
The library
31 expert skills, across the whole consumer lifecycle.
The library ships with the retention playbook already written: 31 built-in skills spanning six lifecycle stages — from the first welcome to the last-chance rescue, plus the cross-cutting disciplines every play runs through. Each is a full procedure, not a prompt: steps, state gates, tool allowlist, verifier rules.
Acquisition
5 skillsTurn a first touch into a member who sticks.
- Welcome & Onboarding Series — a paced first-week arc that lands the program's value without flooding the inbox
- First Purchase Conversion Nudge — moves a signed-up-but-never-bought member to a first order with one right-sized push
- Referral Growth Loop — invites proven happy members to refer — with the fraud guards on from the start
Also: Store Scan-to-Capture Optimizer · Ambassador Candidate Recruitment
Growth
5 skillsDeepen the habit while the relationship is young.
- Category Cross-Sell Expansion — extends a member into their predicted next category — only when the model actually predicts one
- Replenishment Timing Reminder — lands the refill nudge inside the member's own reorder window, not everyone's Tuesday
- Points Balance Activation — wakes up members sitting on a balance they've forgotten they have
Also: Tier Upgrade Momentum Push · eStamp Completion Accelerator
Maturity
5 skillsKeep your best members feeling like your best members.
- VIP High-CLV Appreciation — recognition for top predicted-value members — appreciation, not another discount
- Birthday & Milestone Celebration — the personal moments, executed on time with consent and frequency caps intact
- Points Liability Redemption Drive — converts dormant points liability into store visits instead of balance-sheet risk
Also: UGC Advocacy Cultivation · Store Visit Frequency Builder
At-risk
4 skillsAct while the member is still reachable.
- Churn Early Intervention — steps in when churn risk crosses the line — while a small gesture still works
- Cart Abandonment Recovery (3-Step) — a three-step ladder — remind, reason, then (only then) an incentive
- Engagement Fatigue Cooldown — the counter-intuitive one: detects over-messaging and deliberately backs off
Also: Tier Downgrade Save
Win-back
3 skillsOne fitting gesture — or a deliberate hold.
- Win-back Worklist Execution — works the high-value at-risk worklist one member at a time, inside the server-side guards
- Lapsed VIP Personal Rescue — a personal, high-touch save for the members worth the most — never a batch blast
- Dormant Segment Mass Reactivation — the honest budget-capped sweep for the long tail that doesn't merit 1:1 work
Cross-cutting
9 skillsThe disciplines every play above runs through.
- Coupon Budget-Safe Issuance — issues incentives against live per-store budgets with hard ceilings — never past the cap
- Compliance Outreach Pre-flight Gate — consent, quiet hours and frequency checked before any outreach skill sends
- Metric Anomaly Diagnosis — traces a moved metric to its cause through the governed semantic layer — no invented numbers
Also: Approval Worklist Adjudication · Fraud Worklist Triage · Hire-AI Brief Execution · Structured Audience Segment Builder · Synthetic Panel Pre-launch Rehearsal · Journey Drop-off Diagnosis
Built-ins are a starting point, not a ceiling — teams version, tune and extend the library through the same governed lifecycle.
Governed like an asset
A skill is a managed asset — with an audit trail to match.
The difference between a skill library and a folder of prompts is governance. Every skill in Flash lives under the same discipline as your money paths.
Skill lifecycle
Only published skills serve · editing means going back to draft · deprecated is terminal — a comeback is a new version, with lineage
A lifecycle, not a prompt file
Every skill moves through a state machine — draft → verified → published → evolving → deprecated. Only published skills serve. Deprecated is terminal: bringing a play back means a new version, with its history attached.
Versioned, with lineage
A skill upgrade is a new version row that points at its parent — carrying the evidence that justified the change. You can always answer "what exactly ran last quarter, and why did it change?"
Published by a second person
Submitting a skill for verification and approving it for publication are separate acts by different people — the platform rejects self-approval. No one quietly ships their own SOP.
The library can't be deleted out from under you
The 31 built-in skills are platform-owned: teams tune, extend and version them, but can't silently delete them. Your agents' operating procedures don't disappear because someone cleaned up.
A machine-checkable verifier per skill
Each skill ships with executable rules — preconditions, invariants, postconditions: budget caps, frequency caps, consent, idempotency. The machine checks them; they are never left to the model's goodwill.
LLM discipline, written into the skill
Every skill carries operating notes the model must follow: when a prediction is degraded, read the reason codes before acting; when state is unavailable, never guess — route to a human. Every number must cite a tool-returned fact.
State-conditioned execution
Agents never call raw APIs. They run skills — when the state supports it.
A Flash agent picks from a whitelist of published skills — and every skill declares, up front, the Consumer World Model state it needs to run responsibly: a churn risk above the line, a predicted next category that actually exists, a prediction fresh enough and confident enough to act on.
When the gate fails — the prediction is missing, stale, or below the skill's confidence floor — the skill doesn't run on a guess. It fails closed: the work routes to a human, with the reason attached. Degraded state isn't hidden from the model either — the skill's operating notes require it to read the reason codes and downgrade the play before it acts.
One invocation, gated
Agent intent
“save this at-risk member”
Skill: Churn Early Intervention
requires: churn risk ≥ threshold · fresh · confidence floor met
State holds
runs the steps — through allowlisted tools only, verifier rules checking caps and consent
Gate fails
no guessing — the case goes to a human worklist, reason attached
Related reading
Keep exploring the pages most related to this one.
Intelligent Decision
A next-best-action: who, which offer, which channel, when.
Read more CapabilityConsumer World Model
The core technology under every decision: a Behavior Model that learns how your members actually behave — trained per brand on their own event stream, honesty-gated against a transparent baseline — plus a Digital Twin that simulates what they'd do next, rehearsing campaigns (and the option of doing nothing) before a message goes out. Together, one complete, living model of your member base.
Read more CapabilityDashboard
Real-time scan-to-redemption data as KPI trends, funnel diagnostics, and a rule + LLM next-best-action panel — plus retention curves by signup cohort and a 14-day revenue outlook that's honest about being a projection, not a commitment.
Read more CapabilitySoTag for Slack
The Slack agent that hands governed, audited numbers to Claude in-thread: daily anomaly scan + human-approved coupon-send. Early access.
Read more CapabilitySoClaw — Hire your AI employees
Hire AI employees for your loyalty program — autonomous agents that each hold a real job: a Loyalty Manager that wins back at-risk customers, an Order Manager that clears the receipt and order review queue, a Risk Manager that protects points from fraud, and a Marketing Manager that turns a goal into a ready campaign. Each works one case at a time within limits you set and logs every decision; you manage them like a team with goals, budgets, tasks and feedback. Off by default, capped, instantly pausable. Early access.
Read more CapabilityAI Governance & Approvals
Every AI action runs on a per-action authorization gear — automatic, human review, or forbidden — with an approvals console, a complete decision log, one-click undo, and automatic demotion to human review when an action starts misbehaving.
Read more