메뉴
Read scenario scores and trajectories in scenario-dumps/, diagnose patterns across models, understand root causes, propose and apply changes to skill/prompt files.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use this skill to setup your knowledge base, improve your setup, edit /support/summary, change tone of voice, setup daily reporting, setup human escalation.
Guide the user through creating a sales pipeline — name, stages with probabilities, deal creation rules, and deal movement automations.
Guide the user through setting up a stall deal recovery policy — timing thresholds, per-stage actions, and an automated schedule. Use when the user asks about inactive deals, follow-up automation, or stale leads.
Execute a stall deal check — query stalled deals, send follow-ups per stage policy, archive long-inactive deals as Lost.
Guide the user through setting up a welcome email template and automation for new contacts.
How to log CRM activities (conversations, emails, calls) using erp_table_crud on the crm_activity table. Fetch when you need to record an interaction with a contact.
| name | improve-bot-by-looking-at-scenario-trajectories |
| description | Read scenario scores and trajectories in scenario-dumps/, diagnose patterns across models, understand root causes, propose and apply changes to skill/prompt files. |
| allowed-tools | Read, Glob, Grep, Edit, Write, Bash(ls *) |
You are diagnosing why a bot underperforms in scenario runs and fixing the root causes in skill or prompt files.
Pattern: scenario-dumps/BOT__SCENARIO-*-score.yaml
User will point you at score files.
Or just find the most recent runs by looking at creation timestamp of -score.yaml
Most likely you will see bot running different models, but some kind of A/B test is possible too, to study which code is better, in that case experiment name (goes right after bot version) is important.
Read all score pointed by human or belonging to the most recent experiment.
Trajectories -actual.yaml are bigger in size, in fact might be huge. Pick 1 or 2 to improve your understanding of the problems, think which will be most informative.
The -happy.yaml files are all equal within an experiment, so read one if you are curious, it's also mostly equal to scenario file in the bot, only formatting changes, and expansion of !PYTHON[] notation to include some json from source code so it does not have to be copy-pasted.
First, list every problem you can identify.
BTW not every trajectory is indicative of a problem, sometimes it's random noise.
Then be critical of your ability to fix things. What you think you can fix are only hypotheses at this point. Pick 1-3 points you are likely to successfully fix and go investigate the source code.
Look up bot code, skills, flexus-client-kit sources, other bots, confirm or disprove your potential solutions.
For each of 1-3 problems you have solutions at all, offer to the human alternatives how to fix it. Sometimes solutions are worse than the problems, right. Fix only what the human chooses.
