Enterprise AI Portfolio Curation

See the Work. Standardize the Pattern.
Curate the Portfolio.

Skills Trace shows enterprise AI leaders the top recurring work patterns across the organization, finds where approved capabilities are missing or bypassed, and prioritizes what to standardize next — so executives can answer one question with confidence: how much of our AI work is covered by approved assets?

skillstrace.dev/dashboard/coverage

Top Work Patterns

weekly-update-draft
customer-response
ticket-triage
release-notes

Portfolio Coverage

67%

of top patterns covered (+8 pts QoQ)

Standardization Queue

weekly-update
Build
release-notes
Extend
ticket-triage
Merge

Bypass & Overlap

customer-response

Heavy bypass — approved agent skipped

release-notes / changelog

Overlap — two assets, same work

23 opportunities awaiting review

Your AI Portfolio Is a Black Box

You can see how much you spend on AI. You can see who is calling which model. But you cannot see the work itself — the recurring patterns hiding inside hundreds of thousands of prompts, or whether the capabilities and agents you have approved are actually being used.

What You Can See Today

  • Spend — Token costs by model and team
  • Latency — Provider response times, error rates
  • Adoption — Active users, request volume

What You Cannot

The Work Itself

What recurring jobs your teams ask AI to do, which of those have an approved capability or agent behind them, and which are still running ad hoc.

Questions your CTO and AI Program Manager cannot answer today

?What work are people actually doing with AI — across every team?
?Which recurring patterns deserve a standardized capability or agent?
?Where are approved assets being bypassed in favor of ad-hoc prompts?
?Which two approved capabilities quietly cover the same work?

Without portfolio visibility, every AI investment is a leap of faith.

A Curation Engine for Enterprise AI

Skills Trace turns raw gateway traffic into a curated portfolio of capabilities and agents. Built for the Digital CTO and AI Program Manager who own standardization, governance, and adoption.

Capture
Gateway-side telemetry from every prompt
Identity & Context
User → team → business unit attribution + approved registry
Intelligence
Cluster prompts into work patterns; match to approved assets
Governance
Detect bypass and overlap; record portfolio decisions
Experience
Coverage, queue, decisions, draft specs, end-user guidance

Five layers, one workflow — from raw traffic to portfolio decisions

Discovery Mode

For low-maturity orgs with zero approved assets

Find recurring work, generate a standardization backlog, draft specs that humans review before anything ships. Mode is inferred automatically — no flags, no settings table.

Portfolio Mode

For orgs with an existing capability and agent catalog

Match traffic against approved assets, surface bypass and overlap, route end users to the right capability, and rank what to standardize next.

Work Patterns

Recurring AI tasks inferred across teams — the demand side of your portfolio

Coverage

Percent of top work patterns served by an approved capability or agent — the headline executive metric

Standardization Queue

Prioritized opportunities with recommended action — build, extend, merge, retire — and confidence-bounded upside ranges

Bypass & Overlap

Approved assets being skipped for ad-hoc prompts, and approved assets quietly covering the same work

From Prompt Traffic to Portfolio Decisions

Skills Trace runs as a thin middleware inside your existing AI gateway. Prompts are never modified. Capture once at the gateway, curate everywhere.

Your AI Gateway

Capture Middleware

Skills Trace

Portfolio Surface

ClickHouse Telemetry Store

Multi-tenant, columnar, partitioned by tenant

<5ms

p99 capture latency

Zero

prompt content stored

Human

in every decision loop

The Curation PipelineFive steps, humans always in the loop

1.

Capture

<5ms p99

Gateway middleware ingests prompts and emits redacted, length-bounded telemetry

2.

Cluster

batched

Group similar prompts into Work Patterns with cluster signatures and representative examples

3.

Match

with confidence

Exact and near-neighbor matching against your approved capability and agent registry

4.

Score

snapshot-versioned

Rank standardization opportunities; classify bypass and overlap; estimate upside as ranges with confidence

5.

Decide & Draft

admin-gated

Humans record portfolio decisions and review draft capability or agent specs — never autonomous

Captured event

json
// Telemetry event written at the gateway
{
  "tenant_id": "acme",
  "user_id": "u_2yK...",
  "team": "support",
  "business_unit": "ops",
  "prompt_excerpt": "[redacted, length-bounded]",
  "model": "claude-sonnet-4-6",
  "ts": "2026-04-27T18:42:11Z"
}

Prompt content is redacted and length-bounded before storage. Only attribution metadata, model, timestamps, and a representative excerpt are kept — never raw user content.

One Surface for the Whole Portfolio

From the executive coverage chart down to a single work pattern’s org spread — every view is built for weekly triage and monthly portfolio review.

Portfolio Coverage

Executive

67%

of top work patterns covered

+8 pts vs last quarter

Headline executive metric: percent of top work patterns served by an approved capability or agent. Trended for the board.

Standardization Queue

Queue

23

opportunities ranked this week

Top item: weekly-update

Prioritized opportunities with recommended action — build, extend, merge, retire — and confidence-bounded upside.

Portfolio Decisions

Governance

47

decisions logged this quarter

12 awaiting follow-up

Recorded human decisions with action, alternatives, owner, and outcome — your audit trail and adoption story.

Bypass & Overlap

Risk

9

active findings

customer-response: heavy bypass

Approved assets being skipped for ad-hoc prompts, plus quiet duplication where two assets cover the same work.

Capture Where You Already Run

One Capture Layer. Every gateway. No agent changes.

LiteLLM

Python

Python callback that captures redacted telemetry from the LiteLLM proxy

Vercel AI SDK

TypeScript

LanguageModelMiddleware that captures gateway traffic from Vercel-powered AI apps

Express

Node.js

Drop-in capture middleware for Express-based AI proxy servers

Cloudflare Workers

Edge

Edge-native capture adapter for Cloudflare AI Gateway Workers

Deployment typically requires one configuration change. No agent instrumentation required.

Built for Enterprise AI Governance

Three pillars under the Digital CTO and AI Program Manager — standardize what repeats, govern the rest, and quantify the upside with ranges, not promises.

Standardize

Turn recurring work into curated capabilities and agents

  • Cluster prompts into Work Patterns
  • Exact + near-neighbor matching
  • Action tree: build, extend, merge, retire
  • Heuristic draft specs (LLM generators ready to plug in)

Govern

See what is bypassed, what overlaps, and who decided what

  • Bypass detection with severity classification
  • Overlap detection across approved assets
  • Portfolio Decisions with full audit trail
  • Admin-only governance writes (Clerk + RBAC)

Quantify

Confidence-bounded upside and an executive coverage metric

  • Coverage = % of top patterns with approved match
  • Upside ranges (never single-point ROI promises)
  • Org spread by user, team, business unit
  • Snapshot-versioned opportunity reads

Security & Privacy by Design

No prompt content stored — redacted, length-bounded excerpts only
Read-only inspection — never modifies, blocks, or alters prompts
Multi-tenant by design — every query scoped by tenant_id
Optimistic concurrency on every governance write

One Operating Loop for AI Curation

Weekly triage, monthly portfolio review. Skills Trace runs the loop from raw gateway traffic to executive coverage — and keeps a human in every decision.

01

Discover

Skills Trace

Cluster prompts into Work Patterns; attribute to org spread

02

Match

Skills Trace

Compare patterns against approved capabilities and agents (exact + near)

03

Decide

AI Program Manager

Humans rank opportunities and record portfolio decisions with audit trail

04

Draft

Skills Trace

Generate capability or agent specs (heuristic v1, LLM generators next)

05

Enable

Digital CTO

Coach end users toward the approved path; surface coverage to executives

One Capture Layer. Everything Else Follows.

Drop the middleware into your existing AI gateway. Within a week, you have a real picture of recurring work — and a backlog of standardization opportunities ready for review.

Lightweight gateway middleware

Adapters for Express, Vercel AI SDK, Cloudflare Workers, and LiteLLM. Telemetry lands in ClickHouse multi-tenant by design — every query scoped by tenant, every read snapshot-versioned.

  • No agent instrumentation required
  • No prompt content stored — redacted excerpts only
  • Multi-tenant from request one (tenant-scoped queries)
typescript
import express from "express";
import { createCaptureMiddleware } from "@skills-trace/express";

const app = express();

// One configuration. Capture redacted telemetry from every prompt.
app.use(
  "/v1/chat/completions",
  createCaptureMiddleware({
    tenantId: process.env.TENANT_ID,
    emitter: "https://telemetry.skillstrace.dev/api/telemetry",
    apiKey: process.env.SKILLS_TRACE_API_KEY,
  })
);

Stop Buying AI in the Dark. Curate It.

Every enterprise will run hundreds of capabilities and agents. The ones that win will be the ones whose CTO can answer, on demand, how much of the work is covered — and what is next on the queue.