sandeco-loop is an open skill released as a deliverable accompanying Stop Hand-Holding Your Coding Agent: Engineering the Loops that Replace Step-by-Step Prompting. It turns that paper’s principles and anti-patterns into a repeatable procedure for authoring a Loop Specification. The paper stresses that the skill does not run a loop; it writes the specification of one — a single document capturing the trigger, goal, check, stopping rule and memory, which a human can read, version and hand to an agent harness.
Capabilities
The skill works in stages. It first triages the task by asking whether the outcome of each turn changes the next action; if not, it says so and returns a simple scheduled prompt instead of a loop. For a task that survives triage, it conducts a short interview, one question at a time, covering the goal and whether success is verifiable, the concrete check, the trigger, the named stop states beyond success, the named skills and any nested sub-loops, where state lives between turns, and guardrails such as iteration and budget ceilings and points needing human approval.
A hardening pass then maps onto the paper’s anti-patterns: it insists on an external check rather than a self-score, keeps the maker distinct from the checker, requires terminal states in which an error or exhausted budget never counts as success, enforces one change per turn with the worst item first, caps nested sub-loops with a multiplicative ceiling and forbids a sub-loop from invoking its caller, puts state on disk, and records cost per accepted change as a health metric.
The output is a single <name>-loop.md document with a fixed skeleton — description and “use when”,
goal and verification, the steps of one turn, named stop states, guardrails, memory location,
sub-loops, a short “why it works” section, and an actuation clause. Actuation is either a /goal
command, when the work fits inside a single context window, or a fresh-context
Ralph Loop skeleton when it is long enough that a single context would degrade; the
skill can optionally also emit a /loop-<name> command. The paper’s worked example is a
test-coverage loop whose terminal states are success at full coverage, no-progress after two barren
rounds, and exhaustion after a turn ceiling.
Adoption & Ecosystem
The paper offers the skill as a bridge from principles to a writable artifact, not as an evaluated contribution, and reports no usage data for it; whether specifications written this way improve outcomes is left as an open empirical question. It is published in the author’s prompts repository on GitHub.