General-purpose coding policy for Baruch's AI agents
73
92%
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
The lead applies this reference at intake, when composing a bug-fix brief, and when assessing findings before a correction. The worker receives the relevant questions in its self-contained brief. Reasoning stays with the lead and workers; the dispatch utility does not infer intent or causality from prose.
Record the operator's actual request and later decisions separately from the lead's proposed implementation. Name the observable behavior or deliverable the task must provide, its acceptance criteria, and any explicit limits. Use the existing task identity and authorization record throughout corrections.
Resolve ordinary implementation choices within that accepted behavior. A hard implementation, an additional necessary file, or a required regression test does not by itself create a new permission requirement. A diagnostic recommendation alone does not authorize implementation when the request was investigation only. State investigation-only intent explicitly in the role's task text and select the developer template's investigation branch. Its deliverable is a report; implementation, push, PR and release instructions do not apply to that branch.
Gate investigation reports against the requested knowledge and evidence, including required independent checks. The implementation phases in Review Before PR apply only to authorized code changes. Preserve blocking-finding, correction-allowance, and judge rules for the investigation. A judge ruling returns the investigation to its knowledge-deliverable gate; it never supplies code-release authority.
Read the full finding and identify the behavior that implementing it would commit the project to provide. Record the applicable acceptance criterion and classify the proposed change as one of:
A reviewer's label or confidence does not amend the accepted contract. Scope
classification also does not dismiss a blocking finding: a contested verdict,
lead override, or bot disagreement follows the existing judge path in
rules/agent-team-operation.md. The lead cannot waive a finding by calling it
an expansion. An agreed correction still obeys the fix allowance and release
gates; this reference adds no attempts or substitute judge trigger.
When an operator decision is needed, save an attention item before presenting it. State the original requirement, proposed expansion, smallest compliant option, consequences of each choice, and a recommendation. Reuse a prior decision when its scope still covers the action. Continue independent authorized work while the affected task waits.
When successive findings concern the same underlying design, compare their causes and the actual progress before proposing another fix. Record whether the changes close distinct defects or keep compensating for an unresolved assumption. Put that evidence in the next brief, existing judge dispute, or operator proposal as applicable. Preserve the cumulative attempt count.
For a bug task, ask the developer and tester for a reproduction matching the actual user path. Record expected and observed behavior, setup, inputs, and repeatability. If a faithful reproduction is unavailable, state the limitation and what the substitute does and does not establish.
Separate the initiating trigger, any condition that hides or exposes the fault, and the symptom the user sees. Compare the failing path with a demonstrated working path and locate the earliest relevant difference. Inspect history when it can explain that difference; proximity to a recent commit is not causal proof.
Name the leading explanation and an observation that would disprove it. Run the smallest feasible counterfactual, changing one relevant condition at a time, and retain contradictory results. Explain how the proposed cause accounts for both paths. Label facts, hypotheses, and unresolved uncertainties separately.
Before accepting a diagnosis, the lead checks that its cause explains the reproduction and comparison evidence. A missing experiment is an explicit gap, not an invented pass. Obtain a focused investigation when the gap could change what should be fixed. Keep bounded research inside an already authorized bug fix when it does not require a separate deliverable or decision.
When implementation is authorized, turn the reproduction into an appropriate regression test. The tester verifies the failing behavior before the fix and the expected behavior after it where feasible, and records any limitation. Retain the evidence in the normal report so the retrospective can examine failed assumptions and successful diagnostic methods.
.tessl-plugin
hooks
rules
skills
adopt-fork-pr
herdr-standup
herdr-teamlead
references
teamlead
templates
tests
migrate-to-plugin
onboard-repo
release
tests