Loop check readiness
A coding agent only succeeds when it can get a fast, deterministic, agent-runnable pass/fail
signal for the changes it makes. Whether it can comes down to two things meeting: what the
repo affords, and what this agent can actually do with the tools and skills it has. This skill
audits both halves and reports where they fail to meet. It judges the repo's standing capacity,
not one specific bug — and it proposes, it never scaffolds.
The rubric
A signal is good when it is:
- Present — a real command the repo ships, not one you invented.
- Fast — tight enough to run every iteration.
- Deterministic — same input, same verdict; matches the CI gate; no flake.
- Agent-runnable — you, headless, with the capabilities you have, can run it end to end and
read the result. A step that needs a human — interactive auth, a click, the user's
credentials, a GUI that won't launch headlessly — breaks the loop even when it's set up for a
person. And you have to discover and choose the path from a cold start: a loop that only works
if you already know the trick — which surface ships, which port to attach, the bring-up steps —
is not agent-runnable until that knowledge is written where an agent reads it (
AGENTS.md /
CLAUDE.md, a self-describing run script), not left to tribal knowledge.
That last pillar is where repo and agent meet, and it's the one people forget. Count the
capabilities you actually have — including user-scope skills you carry between repos, like
visual-loop or browser-driver tooling. A repo is a good sign when it lets you exercise them: a
documented run command, a headless path, auth you can seed. It's a bad sign when its only route
to a signal is one you can't take.
What to check
- Dev loop — tests, build, type-check, lint. Is each a real, discoverable command? Where
cheap and safe, run it once to confirm it's real and time it; declared is not working. In a
monorepo the real harness is per-package, not at the root.
- Visual loop (only if the repo renders UI) — can you actually get pixels, headless, with no
human in the auth path? If the UI renders but nothing you have can capture it, that's a gap —
"the code looks right" is not evidence. And can a cold agent tell which surface ships — when
signals are ambiguous (an Electron main process and a web dev server), the shipping surface
must be written down, not left to a guess that lands on the wrong one.
- Parallel-agent safety (when work fans out — parallel critics, a committee, acceptance) — can
several agents drive the app at once without colliding? Look for isolation the app affords:
per-run ports/sessions, isolated user-data/runtime roots, no single hard-coded debug port or
shared fixture state. A loop that only holds for one agent at a time caps fan-out — flag it.
Discover, don't invent: grade whatever the repo ships (build files, CI config,
CLAUDE.md/AGENTS.md, any harness-quirks memory), in any stack. Never run something with side
effects or that won't exit on its own.
Report
For each loop, a one-line verdict — healthy / weak / missing — then:
- Gaps — no usable signal, or the only path needs a human the agent can't be.
- Opportunities — a signal exists but is slow, flaky, ungated, or needs a capability you have
that the repo doesn't make reachable.
Pair every finding with a concrete remediation, grounded in a real file:line or a named
missing file. A missing or unreachable signal is a finding, never a silent "looks fine."
1---2name: loop-check-readiness3description: Read-only audit of whether a coding agent can actually get a fast, deterministic, agent-runnable feedback loop in a repo — judging both what the repo affords (tests, build, type-check, lint, CI, and a reachable, drivable UI for visual checks) and whether the agent can close that loop unattended with the tools and skills it has (no interactive auth, human clicks, or non-headless steps). Reports gaps and opportunities with a remediation for each; it inspects and reports only, changing no files and building no loop. Use when starting work in an unfamiliar repo, when the user asks whether a project is set up for success or has a good feedback loop, or to audit or grade its feedback signals before relying on an agent loop.4allowed-tools: Read, Grep, Glob, Bash5added_in: 0.28.06metadata:7 source: https://skillmd.com/skills/superliaye/loop-check-readiness8---910# Loop check readiness1112A coding agent only succeeds when it can get a **fast, deterministic, agent-runnable pass/fail13signal** for the changes it makes. Whether it can comes down to two things meeting: what the14repo affords, and what *this* agent can actually do with the tools and skills it has. This skill15audits both halves and reports where they fail to meet. It judges the repo's standing capacity,16not one specific bug — and it proposes, it never scaffolds.1718## The rubric1920A signal is good when it is:2122- **Present** — a real command the repo ships, not one you invented.23- **Fast** — tight enough to run every iteration.24- **Deterministic** — same input, same verdict; matches the CI gate; no flake.25- **Agent-runnable** — *you*, headless, with the capabilities you have, can run it end to end and26 read the result. A step that needs a human — interactive auth, a click, the user's27 credentials, a GUI that won't launch headlessly — breaks the loop even when it's set up for a28 person. And you have to *discover and choose* the path from a cold start: a loop that only works29 if you already know the trick — which surface ships, which port to attach, the bring-up steps —30 is not agent-runnable until that knowledge is written where an agent reads it (`AGENTS.md` /31 `CLAUDE.md`, a self-describing run script), not left to tribal knowledge.3233That last pillar is where repo and agent meet, and it's the one people forget. Count the34capabilities you actually have — including user-scope skills you carry between repos, like35visual-loop or browser-driver tooling. A repo is a good sign when it lets you exercise them: a36documented run command, a headless path, auth you can seed. It's a bad sign when its only route37to a signal is one you can't take.3839## What to check4041- **Dev loop** — tests, build, type-check, lint. Is each a real, discoverable command? Where42 cheap and safe, run it once to confirm it's real and time it; declared is not working. In a43 monorepo the real harness is per-package, not at the root.44- **Visual loop** (only if the repo renders UI) — can you actually get pixels, headless, with no45 human in the auth path? If the UI renders but nothing you have can capture it, that's a gap —46 "the code looks right" is not evidence. And can a cold agent tell *which surface ships* — when47 signals are ambiguous (an Electron main process *and* a web dev server), the shipping surface48 must be written down, not left to a guess that lands on the wrong one.49- **Parallel-agent safety** (when work fans out — parallel critics, a committee, acceptance) — can50 several agents drive the app at once without colliding? Look for isolation the app *affords*:51 per-run ports/sessions, isolated user-data/runtime roots, no single hard-coded debug port or52 shared fixture state. A loop that only holds for one agent at a time caps fan-out — flag it.5354Discover, don't invent: grade whatever the repo ships (build files, CI config,55`CLAUDE.md`/`AGENTS.md`, any harness-quirks memory), in any stack. Never run something with side56effects or that won't exit on its own.5758## Report5960For each loop, a one-line verdict — healthy / weak / missing — then:6162- **Gaps** — no usable signal, or the only path needs a human the agent can't be.63- **Opportunities** — a signal exists but is slow, flaky, ungated, or needs a capability you have64 that the repo doesn't make reachable.6566Pair every finding with a concrete remediation, grounded in a real `file:line` or a named67missing file. A missing or unreachable signal is a finding, never a silent "looks fine."