Capability Levels
ails check reports a Level: L# <Name> line in the scorecard between Agent: and Scope:. The level is a read-out, not a gate — every rule fires when its match conditions apply regardless of which level your project is at. Use the level to self-locate; use the symptom table at the bottom to decide when to climb.
The Ladder
| Level | Name | What's added | Channel |
|---|---|---|---|
| L0 | System | System prompt only | attention |
| L1 | Primer | One instruction file (CLAUDE.md, AGENTS.md, .cursorrules) |
attention |
| L2 | Composite | Multiple files — user defaults, project overrides | attention |
| L3 | Scoped | Path-scoped rules (.claude/rules/*.md) |
attention |
| L4 | Delegated | Skills — procedures invoked on demand | attention |
| L5 | Abstracted | Sub-agents — child contexts called by the parent | attention (interface) |
| L6 | Governed | Hooks, MCP gates, deny-permissions | enforcement |
| L7 | Adaptive | Self-improving skills written by the agent | self-writing |
The ladder sorts by the channel each rung runs on: soft attention (L0–L5), hard enforcement (L6), self-writing memory (L7). Each rung adds a new diagnostic concern — scope leakage at L3, skill-instruction coherence at L4, governance-instruction alignment at L6, drift detection at L7.
Channels
- Attention — text the model reads and weights against everything else loaded. Fails probabilistically; competes for budget; decays with load. Fixes are content and ordering.
- Enforcement — hooks, MCP gates, deny-permissions. Acts outside the model's context. Fails deterministically when configured wrong, never silently. Fixes are scripts, schemas, permission rules.
- Self-writing — agent-authored instructions written between sessions. At read time these land in attention like anything else; at write time the user never saw the prompt that produced them. Fixes are review cadence and explicit auto-memory boundaries.
Detection
The displayed level is the highest architectural capability present, cumulative — every level below must also pass.
| Detected | Level |
|---|---|
| Auto-memory, learned rules | L7 (Adaptive) |
| Hooks, MCP servers, managed policies | L6 (Governed) |
| Sub-agent definitions | L5 (Abstracted) |
| Skill definitions | L4 (Delegated) |
Path-scoped rules (.claude/rules/ with paths:) |
L3 (Scoped) |
| Multiple main files, user defaults, overwrites | L2 (Composite) |
| Single main instruction file | L1 (Primer) |
| No instruction files | L0 (System) |
A project with skills but no path-scoped rules still displays L4 — the gate engine walks L1 → L2 → ... → L7 cumulatively. Hub-only repos that ship hooks but no instructions sit at L0 with an enforcement layer; the level reads the soft channel.
When to climb
Each rung exists because the rung below it fails in a specific way. The trigger is the failure, not a feature wishlist.
| From | To | Symptom that triggers the climb |
|---|---|---|
| L0 | L1 | Re-explaining the same project context every session |
| L1 | L2 | One file got long enough that important rules get ignored |
| L2 | L3 | Path-irrelevant rules pollute every task |
| L3 | L4 | The same procedure gets described inline across multiple rules |
| L4 | L5 | A procedure pollutes the parent's context with reasoning chains the parent doesn't need |
| L5 | L6 | A constraint must hold 100% of the time, not 95% |
| L6 | L7 | You keep correcting the same preference across sessions |
Climbing without a symptom adds structure the model has to navigate without solving a problem you had. Under-climbing is more common: "agent didn't run tests before pushing" reads like a prompt-engineering problem but is usually a missing L6 hook; "agent forgot we use Cloudflare Workers" reads like context drift but is usually a missing L7 memory entry.
← Score Guide · Capability Levels · FAQ →