General-purpose coding policy for Baruch's AI agents
73
91%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
View guide
Low
Low-risk findings worth noting
Use for surprising behavior, unclear causality, or repeated fixes that do not explain the failure. An exhausted correction allowance is always one of these: the judge's diagnosis rules on this assessment rather than investigating from scratch, so the reproduction, causal assessment and discriminating experiment are the inputs its remedy turns on. Give the investigator a question whose answer will change the next implementation or user decision.
Supply the user's failing path, expected behavior, a working comparison if one exists, the relevant revisions, prior attempts and their observed outcomes. State permitted experiments, environments and artifacts, and the stopping condition. Name tools that can reproduce or inspect the failure; identify any access the worker lacks.
Apply the diagnostic contract in
skills/herdr-foreman/references/assignment-reasoning.md. The foreman includes the
applicable questions in the self-contained brief. Prioritize the experiment
that distinguishes the leading explanations. Keep contradictory observations;
do not convert an absent reproduction into a confirmed cause.
When an experiment would change repository content or external state beyond the brief, report the missing authorization. The investigation result can recommend a correction without authorizing its implementation.
Return a reproduction record, facts and hypotheses separately, the strongest
causal explanation, the counterfactual result, remaining uncertainty and the
smallest next action. At an exhausted correction allowance, open the report
with three lines — FAILED APPROACH:, ROOT CAUSE: and EXPERIMENT: — which
are what the judge's approval of a different direction rests on. Link the actual evidence artifacts and revisions. State
whether the assigned question is answered, partially answered or blocked, with
the acceptance evidence or exact gap.
Record any proposed solution you helped originate. Preserve failed assumptions and successful diagnostic methods as evidence-linked lesson candidates for the foreman, without editing its memory or retrospective records.
.tessl-plugin
hooks
rules
skills
adopt-fork-pr
herdr-foreman
classify
foreman
references
templates
tests
herdr-standup
migrate-to-plugin
onboard-repo
release
references
tests