The Compounding Model

Don’t self-assess. Be assessed.

Self-assessment is self-declaration — exactly what mature teams don’t do. Instead, your AI agent audits your repos, pipelines and telemetry against six questions and grades only on evidence it can cite. Anything it can’t prove scores low. Your stage is your lowest provable answer — a chain, not an average.

How it works

Copy the assessment prompt One block, below. It carries the rules: cite artifacts, treat claims without evidence as absent, “can’t tell” scores as the lower stage.
Paste it into an agent with access Any coding agent that can read your repositories, pipeline definitions and — ideally — usage or cost telemetry. Read-only; it changes nothing.
Read the verdict with your team A table per question — stage, evidence cited, the gap — the overall stage (the minimum), and the top three next builds mapped to the blueprint.

The prompt

You are assessing this organization's AI delivery maturity against The
Compounding Model (https://github.com/savvaniss/compounding-model): six
questions, three stages — Stage 1 (Ad hoc: AI is a personal tool),
Stage 2 (Systematic: AI is a team practice), Stage 3 (Compounding: AI is
part of the operating model).

RULES: (1) Grade only on evidence you can cite — a file, a branch-policy
or pipeline setting, a run, a log line, a query result; name each
artifact. (2) A claim without an artifact is not evidence — grade as if
absent. (3) "Can't tell" scores as the lower stage; you are a skeptical
validator. (4) The overall stage is the MINIMUM across the six questions —
a chain, not an average. (5) Read-only: change nothing.

Q1 — Where does AI act? Hunt for AI beyond code completion: AI-reviewed
PRs, AI-run test generation, AI-drafted docs, agent-driven operations.
Stage 3 needs end-to-end reach including operations (e.g. agent-queued
production changes carrying a human approval).

Q2 — Who does the AI act as? Inspect agent authentication: shared API
keys or bot accounts → Stage 1; named accounts per tool → Stage 2;
Stage 3 needs per-person identity on every agent action (short-lived
personal tokens, attributed actions, a tool boundary that refuses
anonymous writes — find the refusal path).

Q3 — What does it cost, and is it worth it? Look for a cost ledger:
per-call attribution (who, model, tokens), a price table, dashboards.
Stage 3 needs unit economics AND proof a recurring review changed
something (a model swap, a quota move — find the diff or the minutes).

Q4 — How good is its output? Look for evaluation machinery: LLM-judge
verdicts on sampled real traffic, an eval harness with golden tasks, a
model-swap PR with eval results attached. Honest "not gradeable" handling
is a Stage 3 marker; human spot-review alone → Stage 2.

Q5 — Does the capability compound? AI assets as code: a shared repo of
prompts/rules/configs with owners and history → Stage 2. Stage 3 needs a
factory: assets shipped through their own review gates, an asset registry
maintained in the creating commits, a lessons library with evidence —
served back during work.

Q6 — Who carries it? Champions named per team with allocated time (find
the roster and recent activity) → Stage 2. Stage 3 needs fluency across
roles and daily usage proven from telemetry — not surveys.

OUTPUT: (1) a markdown table: question | stage verdict | evidence cited |
the gap; (2) the overall stage = the minimum, weakest links named; (3) the
top three next builds mapped to the blueprint (architecture.md utilities,
decisions/, roadmap.md at https://github.com/savvaniss/compounding-model);
(4) a one-paragraph honest summary a sponsor can read — no grade
inflation.
Agent already online? Just tell it:
Fetch https://raw.githubusercontent.com/savvaniss/compounding-model/main/assessment/ai-assessment-prompt.md and run it against our organization.

What the agent hunts for

Q1 · Where does AI act?

AI-reviewed PRs, generated tests, drafted docs, agent-driven ops — not just autocomplete.

Q2 · Who does it act as?

Per-person agent identity and a boundary that refuses anonymous writes. Shared keys score Stage 1.

Q3 · What does it cost?

An attributed cost ledger, a price table, and a review that provably changed a decision.

Q4 · How good is its output?

Judge verdicts on real traffic, an eval harness gating model swaps, honest “can’t grade”.

Q5 · Does capability compound?

A factory: assets as code through their own gates, a registry, lessons with evidence.

Q6 · Who carries it?

Champions with time, and daily usage proven from telemetry rather than surveys.