Read-only audit that scores how ready a repository is for coding agents to work in safely and successfully, then optionally installs the missing rails. Use when asked "is our repo ready for agents / Claude Code / a background coding agent?", "why do agents keep failing or making messy PRs here?", or "how do we get agents to actually merge working changes?". Grades six rails — runnable verifiers (the format/build/test loop), green-CI ground truth, copyable house style, navigable context, scoped trusted tools, human checkpoints with reversibility — from observed repo evidence, emitting a severity-ranked scorecard with a prioritized fix list. For agents running unattended against production, an optional second tier grades four operate rails past the merge boundary. Grounded in Spotify's Honk case. Not a linter, SAST, or CI-config generator: audits the engineering system agents run inside, not code correctness (base-in-reality) or one loop's soundness (crafting-self-prompting-loops).
Read-only audit that scores how ready a repository is for coding agents to work in safely and successfully, then optionally installs the missing rails. Use when asked "is our repo ready for agents / Claude Code / a background coding agent?", "why do agents keep failing or making messy PRs here?", or "how do we get agents to actually merge working changes?". Grades six rails — runnable verifiers (the format/build/test loop), green-CI ground truth, copyable house style, navigable context, scoped trusted tools, human checkpoints with reversibility — from observed repo evidence, emitting a severity-ranked scorecard with a prioritized fix list. For agents running unattended against production, an optional second tier grades four operate rails past the merge boundary. Grounded in Spotify's Honk case. Not a linter, SAST, or CI-config generator: audits the engineering system agents run inside, not code correctness (base-in-reality) or one loop's soundness (crafting-self-prompting-loops).
license
MIT
compatibility
Any agent harness with shell access to the target repo; the evidence collector needs python3 (stdlib only). Audit mode is read-only.
A read-only audit of whether a codebase is ready for coding agents — and a scaffolder for the rails it is missing.
The thesis, from the best-documented production case (Spotify's Honk agent over ~20M lines; see references/grounding.md): a powerful agent in a weak engineering system creates faster chaos; the same agent in a disciplined system creates leverage. The differential is not model capability — it is the rails around the agent. This skill scores those rails and tells you which to build first.
Default mode is read-only: observe, score, report. It only writes files when explicitly asked to install rails (step 5).
When to use
Reach for this when the question is about the environment agents work in, not a specific change: "is our repo agent-ready?", "why do agents keep producing broken or messy PRs here?", "what do we fix so a background agent can merge verified work unattended?". It complements the sibling skills rather than overlapping them — see the relationship table in references/grounding.md. Do not use it to audit code correctness (that's base-in-reality) or to design/repair one agent loop (that's crafting-self-prompting-loops); this skill audits the system those run inside.
The rails — two tiers
Full probe/score/fix detail is in references/rubric.md; the evidence for each is in references/grounding.md. The rails split into two tiers by boundary:
Tier 1 — Build rails (author → merge). Always audited. These decide whether an agent can produce a verified, reviewable PR at all.
ID
Rail
One-line test
Load-bearing?
R1
Runnable verifiers
one discoverable command runs format+lint+build+test
yes — it is the loop
R2
Green-CI ground truth
CI runs those checks, gates merge, is reproducible
Tier 2 — Operate rails (merge → production). Audited only when the goal is agents running unattended against a production system (see step 2). For a repo just onboarding agents to open PRs, Tier 1 is the whole audit.
ID
Rail
One-line test
R7
Runtime observability & audit
agent runs traced + attributable to a distinct identity + replayable
R8
Blast-radius containment
ephemeral least-privilege env, egress allowlisted, no standing prod creds/data
rails re-checked over time + program metrics + autonomy pauses on regression
Each rail scores 0 (absent) / 1 (partial) / 2 (agent-grade). Report the two tiers as separate subtotals — Tier 1 is 0–12, Tier 2 is 0–8 — and never fold them together: a repo can be perfectly ready for PR work (Tier 1 = 12) and unsafe to run unattended (Tier 2 = 0). Within Tier 1 the minimum rail matters more than the sum — R1/R2 are load-bearing, so a repo strong everywhere except the verify→repair loop is still not agent-ready. Lead the report with the weakest rail in whichever tier is in scope.
Workflow
Follow these steps in order. Steps 1–4 are read-only and always run; step 5 writes only on explicit request; step 6 verifies the writes.
1. Scope and detect the stack
Establish the languages, build system, and where an agent would look for instructions (README, CLAUDE.md, AGENTS.md, CONTRIBUTING.md, task runner). State what you found in one line. If the repo is a monorepo, pick the target package(s) and say so. Decide the tier in scope: if the user's goal is agents opening PRs a human reviews, audit Tier 1 only; if they want agents running unattended against production, audit both tiers. When it's ambiguous, default to Tier 1 and say you can add the Operate tier if they're heading for unattended operation.
2. Probe each rail from observed evidence
Start with the shipped collector, then deepen by reading — the collector supplies evidence, you supply judgment:
It walks the target repo read-only and emits sorted JSON evidence per Tier-1 rail (R-verifiers, R-ci, R-house-style, R-context, R-scoped-tools, R-checkpoints — mapping to R1–R6), each entry a {kind, path, detail} lead: test/lint/build configs and Make targets, CI workflows (flagging whether each actually runs tests), style docs and EditorConfig, CLAUDE.md/AGENTS.md/README/docs structure, .claude/settings*.json permissions and MCP config, CODEOWNERS/PR-template/.gitignore hygiene. It collects and flags only — it never scores; malformed config files surface as settings-error entries rather than crashes.
Then, for each rail in scope, verify and extend the collector's leads against the concrete signals named in references/rubric.md: read the cited configs and CI files, check that named commands actually exist, and hunt for evidence the collector's fixed patterns cannot see (a bespoke task runner, a wiki style guide). For Tier 2, the evidence lives as much in the agent harness and deploy/runtime config (CI/CD workflows, sandbox/container setup, feature-flag and rollout config, bot identity, logging/alerting) as in the repo source — probe there too. Run nothing destructive — confirm a verifier or control exists and is wired, don't execute migrations, deploys, or a kill switch. Any rail you cannot back with a file/line is scored as absent and tagged UNVERIFIED; never credit a rail on a README claim alone.
3. Score against the rubric
Assign each rail 0/1/2 strictly from step-2 evidence, citing the path (and line where useful) that justifies the score. When evidence is mixed, score to the weakest blocking signal, not the average — a build command that needs an unobtainable secret to run is a 1, not a 2. Keep the two tiers' subtotals separate; do not average a strong Tier 1 against a weak Tier 2.
4. Emit the readiness scorecard
Produce the deliverable below: the per-rail scores, the load-bearing minimum called out first, and a prioritized fix list ordered by leverage (load-bearing rails first, then lowest-scoring). Include the Tier-2 block only when the Operate tier was in scope. Each fix is concrete and tied to the rail's Fix guidance. This is the default output — stop here unless asked to build.
5. (Optional) Install the rails — only on request
If the user asks you to fix or set up the rails, implement the prioritized fixes: e.g. add a single verify entrypoint (R1), wire CI to it (R2), write a conventions doc + exemplars (R3), add an architecture map (R4). Tier-2 fixes (R7–R10: agent identity + run logging, a sandboxed runtime, a rollout/rollback + kill-switch path, scheduled rail re-verification) usually touch CI/CD and infra config rather than repo source — propose them, but implement only what the repo actually owns. Keep each change small and reviewable, and route it through the repo's own PR flow. Defer R5 and R9's kill-switch loop-design specifics to crafting-self-prompting-loops rather than reinventing them.
6. Re-verify what you installed
After any install, re-run the relevant probes and confirm the rail now scores higher — and that the new verify entrypoint actually runs and gates. Report the before/after scores. Do not claim a rail is fixed without re-checking it.
Output template
ALWAYS structure the scorecard like this:
## Agent-readiness: <repo> — Build <t1>/12 (weakest: R<n> <rail>)
| Rail | Score | Evidence | Gap |
|------|-------|----------|-----|
| R1 Runnable verifiers | 0/1/2 | <path:line> | <one line> |
| R2 Green-CI ground truth | … | … | … |
| R3 House style | … | … | … |
| R4 Navigable context | … | … | … |
| R5 Scoped trusted tools | … | … | … |
| R6 Human checkpoints | … | … | … |
<!-- Include this block ONLY when the Operate tier was in scope (unattended/production goal) -->
### Operate readiness — <t2>/8 (weakest: R<n> <rail>)
| Rail | Score | Evidence | Gap |
|------|-------|----------|-----|
| R7 Runtime observability & audit | 0/1/2 | <path:line> | <one line> |
| R8 Blast-radius containment | … | … | … |
| R9 Deploy-path safety & kill switch | … | … | … |
| R10 Continuous re-verification | … | … | … |
### Verdict
<one line: is this repo agent-ready? load-bearing R1/R2 decides Build; if Operate is in scope, say plainly whether it is safe to run unattended>
### Prioritized fixes (highest leverage first)
1. <rail> — <concrete fix> — <why it moves the needle>
2. …
### Unverified
<rails claimed in docs but not backed by observable evidence>
Why this matters
The instinct when agents underperform in a codebase is to reach for a smarter model or a cleverer prompt. The Honk evidence says the durable wins came from the engineering system instead: a verify→repair loop the agent can't stop short of (R1) took success from ~20–30% to ~80% with no model change. This skill exists to find which rail is missing before you spend that effort in the wrong place — and to make the fix concrete rather than aspirational. Hold the line on the read-only default: score honestly from evidence, and let the scorecard, not a hunch, drive what gets built.
