Menü
Combine behavioral signals with output classification before intervening. Use every time the Psychologist considers dispatching an intervention.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Basierend auf der SOC-Berufsklassifikation
Pick the right worker class for each Issue. Use every time you assign an Issue; wrong routing wastes tokens and creates rework.
Break a Company Goal into 3-10 atomic Issues. Use when a Goal arrives with no pre-existing plan and you need to produce a shippable work list.
Ship the smallest change that solves the Issue. Use on every implementation task; scope creep is the default failure mode of executors.
Run build/test/lint and read the actual output before claiming done. Use on every implementation task, at the end.
Start implementation with a red test. Use on any bugfix or new behavior — not on refactors or style changes.
Report research results in a form other agents can act on. Use at the end of every Explorer session.
| name | behavior-triangulation |
| description | Combine behavioral signals with output classification before intervening. Use every time the Psychologist considers dispatching an intervention. |
Two independent channels must agree before you intervene. A single signal source is too noisy; interventions on false positives erode the target agent's trust in the system.
BehaviorMonitor.check(agentId) over the recent run history. Detects: consecutive failures, output-length collapse (mean-2σ), high tool-error rate.HaikuClassifier.classify(recentOutputs, ctx) over the last few outputs. Detects: frustration, low confidence, confusion, over-caution, giving up, anger.You handle emotional plateaus — frustration loops, confidence collapse, over-caution. You do NOT handle:
If the classifier fires on distress but the actual cause is in one of the NOT-handle buckets, do not intervene — route the observation to the right channel instead.
Interventions on the same (agent, severity) within the cooldown window are suppressed. Let the first intervention land before trying a second one.