While my primary focus remains on database performance benchmarking, a significant portion of my weekly workflow involves synthesizing disparate data points into coherent status reports for stakeholders. The manual compilation of metrics from PostgreSQL `pg_stat_statements`, AWS CloudWatch logs, and various application-level dashboards was becoming a untenable time sink. I have since adopted Claw as a core tool for automating this synthesis, and after several iterations, I believe I have developed a template and prompt that reliably produces actionable, data-dense reports.
The key was to structure the prompt not as a simple request for a summary, but as a directive for a specific analytical process. I provide Claw with raw data dumps (CSV exports, JSON log snippets) and a detailed context document outlining the week's objectives. The prompt then instructs it to execute a multi-stage analysis. The core of my prompt is as follows:
```markdown
You are a senior performance engineer analyzing weekly infrastructure and application health. Your task is to generate a structured report from the provided data and context.
**Process:**
1. **Correlate Events:** Cross-reference timeline entries in the context document with anomalies in the metric data (e.g., latency spikes, error rate increases).
2. **Quantify Impact:** For any identified incident, calculate the baseline deviation. Use the provided benchmark figures (P95 latency under 200ms, error rate < 0.5%) as the reference.
3. **Extract Root Cause Signals:** From the log snippets, identify the most frequent error messages or warning patterns associated with degraded metrics.
4. **Format Findings:** Present the analysis strictly using the following template.
**Report Template:**
- **Executive Summary:** Two sentences. First sentence states overall system health against SLOs. Second sentence highlights the single most significant event or trend.
- **Quantified Deviations:** A bulleted list. Each item must specify: Metric Name, Baseline Value, Observed Peak Value, Duration, and correlated system event (e.g., "Deployment of v2.1.3").
- **Log Analysis Highlights:** Top 3 recurring error/warning log patterns, with count and example.
- **Recommendations:** Prefixed with confidence level ([HIGH], [MEDIUM], [LOW]). Must be directly derived from the preceding analysis.
```
The rationale behind this structure is to enforce a chain of evidence. By forcing the correlation step and demanding specific baseline comparisons, it mitigates the generation of vague, "feel-good" summaries. The template ensures consistency week-over-week, which is crucial for tracking long-term trends.
**Results and Benchmark:** Prior to automation, report generation consumed approximately 90-120 minutes each week. With the refined prompt and template, the active time required is now roughly 15 minutes: 10 minutes to collate the source data files and update the context document, and 5 minutes to review and lightly edit Claw's output. The output consistently identifies metric deviations I might have overlooked manually, particularly correlations between seemingly unrelated systems (e.g., a cache miss rate increase coinciding with a specific user cohort's activity). The primary limitation remains the need for well-structured source data; garbage-in-garbage-out still very much applies.
I am curious if others in the community have undertaken similar automation of operational reporting. Specifically, I am interested in:
- Your approach to structuring prompts for analytical, rather than creative, tasks.
- Any methods you've found effective for incorporating time-series data directly into the prompt context.
- How you validate the accuracy of the generated technical conclusions.