Surface Performance

Spreadsheets (CSV, XLSX, XLS)

If the user provides a spreadsheet file:

1. CSV -- Read the file directly. Parse headers and rows. If encoding issues arise, try latin-1 fallback.
2. XLSX/XLS -- Use a subagent to extract the data:
   • Read the file with the Read tool (it handles Excel files)
   • If the Read tool cannot parse it, ask the user to export as CSV

Inspect the first 5-10 rows to identify columns. Map detected columns to the fields needed for analysis (case type, open date, close date, disposition, revenue, attorney, referral source, etc.). Report which fields were found and which are missing.

PDFs

If the user provides a PDF with case data (e.g., a CRM report export, court docket summary):

1. Read the PDF using the Read tool (it handles PDF files)
2. Extract tabular data from the text -- look for consistent spacing, delimiters, or table structures
3. If the PDF is image-based or unreadable, delegate OCR to a subagent: launch an Agent (subagent_type: "general-purpose") with prompt: "Run /legal-toolkit:extract-text on {file_path} and write the extracted text to {parent_dir}/{filename}_ocr.txt." Use the OCR output text for analysis.

Text or Pasted Data

If the user pastes data directly or provides a text/markdown file, proceed to analysis immediately. Parse any tabular structure present (pipe-delimited tables, tab-separated values, etc.).

Data Validation

After parsing, report:

• Total records found
• Date range covered
• Fields detected vs. fields missing
• Any parsing issues (malformed rows, inconsistent date formats, etc.)

If critical fields are missing (no dates at all, no case identifiers), flag this and ask the user what data is available before proceeding.

What to Produce

1. Executive Summary

Three to five bullet points -- the most important things the firm owner needs to know from this data. Lead with what changed, what is working, and what needs attention. No jargon, no hedging.

2. Case Volume

Include:

• Cases opened
• Cases closed
• Net change (opened minus closed -- is the backlog growing or shrinking?)
• Cases currently open (total active inventory)

If data supports it, break down by month within the period. Flag any month with a volume drop or spike exceeding 25% from the average.

3. Time to Resolution

Average days from open to close, broken down by case type:

Median matters more than average here -- a single case that dragged on for 18 months will skew the average. Flag any case type where the average exceeds the median by more than 40% (indicates outlier cases pulling the number up).

4. Disposition Breakdown

How cases are resolving:

Map the firm's disposition labels to standard categories as closely as possible (Dismissed, Reduced charges, Settled, Judgment for client, Judgment against client, Mediated/arbitrated, Withdrawn, Other). Flag shifts -- if any category moved more than 5 percentage points, call it out.

5. Revenue Analysis

Also break down by attorney:

Do not rank attorneys by revenue -- present alphabetically. Revenue per case varies by case type, so an attorney handling complex litigation will naturally show higher per-case revenue than one handling routine matters. Note this context.

6. Intake and Conversion

If intake data is available:

Include: total leads/inquiries, consultations scheduled, consultations held (show-rate), cases signed, overall conversion rate (leads to signed cases). Flag if conversion rate dropped -- this is often the highest-leverage number in the firm.

7. Source / Channel ROI

If referral source data is available:

Rank by ROI if cost data is available, otherwise rank by conversion rate. Identify the top 3 sources and any sources that generate leads but never convert.

8. Anomaly Flags

Scan the data for anything unusual and list each finding:

• What: Describe the anomaly in plain language
• Data: The specific numbers that triggered the flag
• Possible explanations: Two or three hypotheses (do not pick one -- let the firm investigate)
• Recommended action: What to look into next

Examples of anomalies to watch for: sudden intake drop, attorney with disposition pattern significantly different from firm average, case type with resolution time trending upward, revenue per case declining for a stable case type, single referral source dominating intake, seasonal patterns, unusually high or low billable hours for a period.

9. Data Quality Notes

List any issues with the data that limit the analysis:

• Missing fields (e.g., no referral source on 30% of cases)
• Inconsistent labels (e.g., "Personal Injury" vs "PI" vs "Pers. Inj." used interchangeably)
• Date gaps or impossible dates
• Recommendations for what to start tracking or how to clean up data entry

Rules

• Use the periods the data supports. If you have 6 months of data, compare quarters. If you have 2 years, compare year-over-year. Do not force a comparison the data cannot support.
• Show the numbers. Every insight must reference specific data points. "Revenue is up" is useless -- "Revenue increased 12% ($47K to $53K) driven by a 3-case increase in personal injury closings" is actionable.
• Present, do not prescribe. Surface the data and flag what stands out. The firm owner decides what to do about it. Avoid language like "you should" or "you need to" -- use "this suggests" or "worth investigating."
• Do not fabricate benchmarks. If you do not have industry comparison data, do not invent it. Compare the firm against itself (prior periods). If the firm asks for benchmarks, note that firm-specific trends are more useful than generic industry averages.
• Format for a busy partner. Executive summary first, details below. Tables over paragraphs. Bold the numbers that changed significantly. The reader should get the headline in 60 seconds and dig into detail only where needed.
• Handle any practice area. This skill works for any type of law firm -- criminal defense, personal injury, family law, corporate, immigration, etc. Adapt disposition categories and terminology to match the practice area reflected in the data.
• Skip sections with no data. If the dataset lacks revenue fields, skip the Revenue Analysis section entirely rather than guessing. Note the missing data in the Data Quality section.

Connector Action: ~~chat

If a ~~chat connector (e.g. Slack) is available, offer to share the performance report:

"Want me to post the KPI summary to a Slack channel for firm leadership?" If yes, post the Executive Summary and top 3 anomaly flags as a formatted Slack message.

Accuracy and QA (Required)

Anti-hallucination rules (include in ALL subagent prompts):

• Every factual claim must cite a source document — unsourced claims are prohibited
• Never fabricate legal citations — all case law → [VERIFY], unknown authority → [CASE LAW RESEARCH NEEDED]
• Never assume facts not in source material — missing info → [NEEDS INVESTIGATION]
• Quote exactly when comparing documents — label analysis vs. facts distinctly

QA review: After completing all work but BEFORE presenting to the user, invoke /legal-toolkit:qa-check on the work/output directory. Do not skip this step.