Self-Assessment

Three rubrics, one for each scope. Each places the subject on a Spine level (L1–L5) and scores each Axis (0–4). The Spine level is capped at L3 if Verification is at the bottom — the core rule: no real agentic maturity without verification.

Note

You can run all three by hand today with the tables below. In an AIOS workspace these are designed to be driven by a guided self-assessment and an automated codebase scan; this page is the rubric they share.

---

1. Individual (workstation)

Spine placement — pick the highest statement that is reliably true under pressure (not your best day):

• L1 — I type requests into a chat and use what comes back; I rarely read diffs.
• L2 — I keep reusable prompts (roles, examples, format constraints) and review diffs before accepting.
• L3 — I maintain a personal/project CLAUDE.md, curate my tools/MCPs, and manage context deliberately.
• L4 — I run agents in a loop against a written spec + a check they can run, and I review every diff. A meaningful share of my work is delegated-and-verified.
• L5 — I orchestrate multiple/parallel agents, run my own evals, and feed learnings back across sessions.

Axis scores (0 = absent, 4 = strong):

Axis	0	2	4
Verification	Eyeballs output	Runs tests/build manually before merge	Agent runs its own check; personal eval harness
Context hygiene	One long session for everything	Clears between tasks	Curated CLAUDE.md + subagents isolate exploration + compaction
Autonomy / leash	Approves everything, or blindly auto-accepts	Auto-accepts low-risk behind a check	Dials leash per risk; earns it via proven checks; sandboxes the rest
Learning	Repeats the same corrections	Adds fixes to CLAUDE.md sometimes	Corrections + memory feed back systematically; builds skills
Cost & governance	No cost sense	Aware of token cost	Token-efficient tooling; respects tier/permission discipline

A useful north star: the delegation ratio — the share of your work that is delegated-and-verified to agents. The “and-verified” is what separates L4 from fast L1.

---

2. Codebase (repo) — the automatable one

This scope is mostly binary, file-existence checks. Score = the % of checks passed, mapped to a level. A repo must pass ~80% of a level’s checks to claim it.

Pillar	Concrete signal (pass/fail)
Agent instructions	A CLAUDE.md / AGENTS.md exists at root, is non-trivial, references example files
Testing	A test suite exists; CI runs it on PRs (the agent’s feedback loop)
Build / validation	Reproducible build; linter + formatter + type-checker configured (guardrails)
Docs	README + setup/contributing docs present
Dev environment	One-command setup (a Makefile, devcontainer, or script)
Code quality	Pre-commit / pre-push hooks enforce style
Observability	Logging/monitoring so agent actions are inspectable
Security / governance	Secret scanning; ignore-hygiene; access conventions
Evals (unlocks L4+)	An eval suite or LLM-as-judge harness lives in the repo

Level mapping:

Level	Meaning
L1 Functional	A human can build and run it
L2 Structured	+ CLAUDE.md, linter/formatter, basic tests
L3 Agent-ready (target)	+ CI-on-PR, type-checker, one-command setup, secret scanning, docs
L4 Measured	+ eval harness, observability, pre-commit gates
L5 Self-improving	+ a compounding CLAUDE.md, shared skills/commands, agent-readable docs

The one metric to subordinate

”% of code AI-authored” is an activity metric, not a quality one. Track it if you like, but never let it outrank eval coverage and agent-readiness. A repo that is 90% AI-authored with no tests is L1, not L5.

---

3. Team / org

A maturity radar, not a single number. Dimensions borrowed from DORA’s 2025 AI Capabilities Model:

Dimension	Signal
AI stance	A documented, communicated policy on permitted tools/usage
Data ecosystem	Internal data/docs are high-quality, unified, and AI-accessible
Version control & batch size	Strong VCS discipline; small, frequent changes
Internal platform	Shared agent harnesses/skills; a quality internal platform
Throughput & stability	Delivery metrics tracked — watch change-fail rate as throughput rises
Trust	The share of developers who trust AI-generated code (closing the gap is maturity)
Adoption depth	% of repos at L3+; % of workflows with eval coverage
Org learning loop	Do evals/learnings feed back into shared prompts, skills, and conventions?

---

From assessment to practice

1. Place the subject (Spine level + Axis scores), applying the Verification cap.
2. Prescribe — take the weakest axis and pick the matching practices from the Pattern Library.
3. Re-assess over time — quarterly for people, per-PR for repos, per-quarter for teams — and track the progression.