Internal — engineering Live artifact — regenerates with the pipeline

Repo AI Health

How to read this. Every repo gets two scores. Agentic readiness (0–100) measures how well-equipped it is for agentic work — and it's content-weighted: a CLAUDE.md earns most of its points for carrying real build/test commands, an architecture map, conventions and "never do" warnings, not for merely existing. AI guidance required (0–100) measures how much guidance an agent needs to work well in that repo, based on the technology and the codebase involved — how big it is, how old, how actively it changes, and whether it's written in a legacy language. An old PHP monolith needs far more agent guidance than a fresh TypeScript microservice. AI health is the gap between the two: the repos to fix first are big, old, busy codebases with thin Agentic readiness.

The estate at a glance — AI guidance required vs Agentic readiness

Each dot is a repo. Bottom-right = critical gap (lots of guidance required, little in place — fix these first). Top-right = healthy where it matters. Dot size = file count. Hover for detail.

Estate-level actions

    What good looks like (in-house patterns)

    aiske/mc — CLAUDE.md + custom agents + commands + committed .mcp.json: the full pattern for a large legacy codebase.
    aiske/dp-svc-hwlake-bronze — CLAUDE.md + skills + agents: recurring data workflows captured as skills.
    aiske-sandbox/integration-platform — CLAUDE.md + skills from day one in a greenfield repo (not in the 6 above — kept as a reference pattern).
    Copying these beats inventing from scratch — the suggestions below reference them.

    Every active repo, ranked by priority

    Critical gaps first (most guidance required within the band), then needs-attention by gap size. Click a row for the full breakdown and suggestions.

    Honest caveats

    1. Content heuristics, not comprehension. The CLAUDE.md quality features (commands/architecture/conventions/warnings) are keyword heuristics — a well-written file could miss a keyword, and boilerplate could match one. Spot-check before acting on a single repo's score.
    2. "AI guidance required" is a model. Size + age + activity + language is a proxy for "how hard is this codebase for agents" — it can't see domain complexity or coupling.
    3. Scope = the 6 key repositories. mc, the app and the AI services are always included; the remaining slots go to the repos with the most active contributors. Archived repos and forks are excluded.
    4. This measures Agentic readiness, not usage. Pair with the adoption leaderboard (who uses the tooling) and the outcomes page (what ships).
    5. Repo activity is a lower bound. The Claude-assisted split counts only PRs whose commits keep the Co-Authored-By: Claude trailer. Squash settings, rebases and amended commits can drop it, so a repo's true AI-assisted share is at least what's shown — never less. Whole-PR lines are attributed to the AI-assisted class when any commit carries the trailer; PRs are bucketed by merge month over the last 6 months, and the most recent month is partial. The % of lines is volume-sensitive (one large PR can swing a month); the % of PRs is the steadier adoption signal.
    6. Scoring updated 7 Jul 2026. Three changes from a hands-on repo sweep: (a) bot-authored commits (GitOps image bumps, merge-queue, dependabot) are now discounted from the activity component — raw counts overstated some repos' human activity by up to 10×; (b) a .gitignore that blocks .claude/ config from being committed is now detected and penalised — it silently keeps agents/skills on one machine while the docs claim otherwise; (c) a CLAUDE.md that documents config not present in the repo is flagged. Grade movements on this date reflect the methodology change, not sudden repo regressions. Deep-dive findings shown on cards are point-in-time agentic reviews (dated) and refresh only when a sweep is re-run.