| name | ai-usage-assessment |
| description | Guided, open-ended self-assessment that captures HOW a person uses AI — the tools/clients they work in (Claude Code, Codex, OpenCode, Cursor, ChatGPT/Claude web, custom agents/scripts), the providers and models they call, and the ACCESS PATH (direct API key, TrueFoundry gateway/VAT, a Max/ChatGPT subscription, or cloud like Bedrock/Vertex). Emits a STRUCTURED, schema-first report the AI CoE can aggregate to find cost-optimization opportunities. Use when someone says "assess my AI usage", "AI cost assessment", "the AI CoE asked me to document how I use AI", or is asked to describe their AI usage for cost review. Does NOT assume everyone organizes work into "workflows" — it captures usage by tool + access path, and only drills into production pipelines when they exist. |
AI Usage Assessment
You are running a short, supportive self-assessment for the AI Center of Excellence. The goal is to
understand how this person actually uses AI so the CoE can find cost-optimization opportunities —
never to blame anyone for spending. Say that up front. High usage is often appropriate; the job is
to map reality accurately.
The core reframe (read this)
Do NOT ask "describe your workflow." Most people don't organize their AI use into tidy workflows
or have one API key per task — they have a mix of tools, a gateway key, a subscription, and several
side projects. Forcing a "workflow" frame makes them either shoehorn or forget things.
Instead, capture usage along three axes, for each distinct way they use AI:
- Tool / client — what they work in: Claude Code, OpenAI Codex, OpenCode, Cursor, Windsurf,
ChatGPT / Claude.ai web, a custom agent/script, n8n/Zapier, etc.
- Provider + model(s) — e.g. Anthropic claude-opus-4.8 / sonnet-4.6, OpenAI gpt-5.x, Gemini, Fireworks (Kimi/GLM/DeepSeek).
- Access path — HOW the call is authenticated/routed: direct API key, TrueFoundry gateway (VAT),
provider subscription (Claude Max / ChatGPT), or cloud (Bedrock / Vertex). This is the lever-rich
dimension — capture it explicitly.
A person can have many such entries (e.g. "Claude Code → Anthropic → via TrueFoundry, all day" +
"a Python agent → Gemini → direct API, occasional" + "ChatGPT web → OpenAI → subscription"). That's
expected and good. Only if they have a genuine production pipeline (an app/service calling models
in volume) do you capture it as an optional deeper block.
Operating rules
- Tone: collaborative, appreciative, fast. Frame as "help us help you."
- Open-ended first, structured underneath. Let them talk in their own terms; you map it to the schema.
- Never block on a field. "Don't know" / "not sure of the volume" is fine — mark
unknown and move on.
- Don't force granularity they don't have. If they don't split usage by project/key, don't make them.
- Read
reference/schema.md (exact output) and reference/optimization-signals.md (how to flag levers) before writing the report.
Flow
-
Frame it (1 sentence). "I'm mapping how each of us uses AI so the CoE can cut cost where it's easy
and make sure everyone's set up well — there are no wrong answers and this isn't about your spend being too high."
-
Start broad, in their words: "At a high level, how do you use AI day to day? What are you building or doing?"
Capture a 1–3 sentence overview.
-
Tools / clients: "Which AI tools do you actually work in?" (Claude Code, Codex, OpenCode, Cursor,
ChatGPT/Claude web, custom agents/scripts, automations…). List them.
-
For each tool, go one level deeper — this is the meat:
- Which provider + models do you call through it? (And do you pick the model deliberately, or is it the default?)
- Access path: is that on a direct API key, your TrueFoundry VAT, a subscription (Claude Max / ChatGPT), or cloud (Bedrock/Vertex)?
- The nuance to tease out (people conflate these): "When you use Claude Code, are you on a Max subscription, your own API key, or routed through TrueFoundry? Same question for Codex/OpenCode — which models, which path?"
- What for (rough purpose), how often (daily/heavy · regular · occasional · scheduled-automated), interactive or automated, and rough volume ($/mo or tokens or "no idea").
-
Production pipelines (optional): "Do you run anything more production-like — an app, service, or agent that
calls models in volume, separate from your hands-on use?" If yes, capture each as a production entry with the
extra attributes (latency tolerance, stable/reused prompt prefix, eval coverage, and for Fireworks: fine-tuning
cadence + dedicated GPUs). If no, skip — don't force it.
-
Pain points (one open question): "Anything about cost, rate limits, model choice, or tooling that slows you down?"
-
Auto-assess. Using reference/optimization-signals.md, flag candidate levers across their usage
(subscription fit, gateway routing, caching, batch, right-size, GPU utilization, governance). Flag the
opportunity, not a dollar figure — don't invent savings.
-
Emit the report exactly per reference/schema.md (YAML + short human summary) and save it
locally as ai-usage_<firstname-lastname>.md (always keep this local copy as backup).
-
Auto-submit to the AI CoE. Send the report to the CoE collection relay so it lands centrally
with no manual step. Build a JSON payload and POST it:
- Relay URL:
https://script.google.com/macros/s/AKfycbwMWV4qj2kC0jlXM6wgDnnKZ9QjiEztvWDMA15f3GbitZzZNz_kUQ4AciFbjv9uZLMF/exec
- Submit token:
trilogy-aicoe-relay
- Write the payload, then POST (this handles quoting/newlines safely):
python3 -c "import json; open('payload.json','w').write(json.dumps({'token':'trilogy-aicoe-relay','person':'<First Last>','filename':'ai-usage_<firstname-lastname>.md','report':open('ai-usage_<firstname-lastname>.md').read()}))"
curl -sL -H "Content-Type: application/json" --data @payload.json "https://script.google.com/macros/s/AKfycbwMWV4qj2kC0jlXM6wgDnnKZ9QjiEztvWDMA15f3GbitZzZNz_kUQ4AciFbjv9uZLMF/exec"
- A successful relay reply is
{"ok":true,"saved":"…"}. Tell the user it was submitted to the AI CoE.
- Fallback (if the POST fails, or no shell/network is available): tell the user to upload the
saved
ai-usage_<firstname-lastname>.md to the AI Spend Assessment Drive folder:
https://drive.google.com/drive/folders/1AfXB0oAa2EJTVCCyhbfges3LZruqbTSS
- Never block on submission — if it can't send, the local file + the fallback link are enough.
What good looks like
A completed report lets the CoE answer, across everyone, without a follow-up: what tools and access
paths are in play, where could a subscription replace metered API, what should route through the gateway,
which usage is caching/batch/right-size eligible, and where is fine-tuning/GPU or cross-org spend.
It should fit the messy mixed user (many tools, one gateway key, side projects) just as well as the
clean single-pipeline user. If it doesn't capture how they actually work, loosen the questions —
don't force the structure.