AI workflow

Is your AI investment
compounding?

Or is it just running up the bill?

Every team is writing with Claude. Fewer have the telemetry to tell whether the spend flows into work that ships, or into context they re-explain every session. The signal is already there, spread across every tool the team ships through. The intelligence lives in the synthesis.

◆The gap

The invoice grows.
The output doesn't, unless you look.

Tokens spent (last 30 days)
142M+38%vs last month
PRs merged with AI assist
12±0vs last month
◆ The scissor · Tokens vs merged PRs · 90-day index
100AI token spend138 · +38%PRs merged100 · ±0T-5moT-4T-3T-2T-1Today

The invoice compounds. The output holds flat. I call this the scissor, and it shows up across cost and token economics, spec and context quality, and build vs buy.

Illustrative, drawn from shapes I keep running into. The specifics differ. The pattern rarely does.

◆What DAC reads

Two streams of telemetry
from the tools you already use.

Claude Code (CLI sessions, agent skills, MCP servers) and Claude Design (prompts, artifacts, review velocity) both emit telemetry. Normalized into one taxonomy, they answer whether the spend is compounding.

CLI telemetry
Claude CodeClaude Code9 signals
  • Session duration and active-time patterns
  • Token spend per session, per developer, per repo
  • Model mix (Opus, Sonnet, Haiku) across sessions
  • Tool-use pattern (Edit, Read, Bash, Grep)
  • Skill adoption and most-used agent skills
  • Plan-mode adoption before implementation
Design telemetry
Claude DesignClaude Design5 signals
  • Prompt-to-artifact ratio per shipped design
  • Iteration cadence across design files
  • Time from AI-generated design to human review
  • Token spend per designer, per artifact
  • Design-system adherence on AI output
See all data sources →
◆What it tells you

Five dimensions,
lit up by the same telemetry.

Claude signals map onto the same 88-unit scoring model as everything else, so they need no separate dashboard and no new vocabulary. Just the work the team is already doing, seen clearly.

  1. Spec & Context Quality

    Shows whether you re-explain context every session, or surface the pattern once.

  2. Cost & Token Economics

    Tracks spend to the person, repo, and task. Not just the monthly invoice.

  3. Build vs Buy

    Flags where AI compounds into shipped work, and where it just looks busy.

  4. Process Iteration

    Surfaces which /plan, /implement, and /review patterns actually ship faster.

  5. Knowledge Management

    Maps skill and MCP usage. Who reaches for what, how often, to what effect.

◆Diagnostic preview

Pattern:
context debt.

High token spend, low plan-mode adoption, and rising skill-adoption variance across the team. Named and scored, the pattern points at the few moves that actually close it.

◆ PatternUpdated hourly
Claude CodeClaude Code
7-day window12 contributors
Tracking

Token spend is up 38%. Plan-mode usage is down 6 points.

Context debtProcess driftSkill variance
DAC · pattern recognition

Diagnosis: Your team is reaching for more tokens per task without the spec and context quality to make them productive. The invoice scales. The merge rate does not. Three moves close it.

  1. Make /plan the default before /implement across the teamProcess iterationMeasured
  2. Standardize 3 agent skills for the top repeating workflowsKnowledge managementProjected
  3. Cap Opus usage on mechanical edits, route to HaikuCost and token economicsProjected

A real pattern. Naming it is the easy part. Catching it while there is still a quarter left to act on it is the part that needs instrumentation.