Ops

Ops

Delivery lifecycle tracking, few-shot pattern capture, and business intelligence for EaaS operations.

/ops few-shot   — Guided few-shot capture (requirement → correction → lesson)
/ops close      — Log a completed delivery to SQLite
/ops insights   — Query patterns from past deliveries
/ops status     — Quick stats (revenue, hours, margins)

---

Data Paths

| Asset | Location | |-------|----------| | Deliveries DB | /Users/lucas/Sites/vlt-data/db/deliveries.db | | DB Schema | /Users/lucas/Sites/vlt-data/db/schema.sql | | JSON Export | /Users/lucas/Sites/vlt-data/exports/deliveries.json | | Few-shots | /Users/lucas/Sites/vlt-data/few-shots/{category}/{slug}.md |

---

/ops few-shot

Step-by-step guided capture using AskUserQuestion. One question at a time, never batch.

Flow

Step 1: "What was the task?" — free text (requirement)

Step 2: "What did Claude generate?" — free text, or offer "auto-detect from last git diff":

git diff HEAD~1 --stat
git diff HEAD~1

If user picks auto-detect, summarize the diff as the generated output.

Step 3: "What did you correct?" — free text (the fixes applied)

Step 4: "What's the lesson?" — free text (reusable insight)

Step 5: "Category?" — present options:

• layout
• component-usage
• theming
• data-display
• forms
• navigation
• other

Step 6: "Save to memory MCP too?" — yes / no

Write Few-Shot File

Generate slug from task description (lowercase, hyphens, max 50 chars). Ensure category directory exists.

mkdir -p /Users/lucas/Sites/vlt-data/few-shots/{category}

Write to /Users/lucas/Sites/vlt-data/few-shots/{category}/{slug}.md:

## Requirement
{answer 1}

## Generated
{answer 2}

## Corrected
{answer 3}

## Lesson
{answer 4}

Memory Store (if yes)

Call mcp__memory__memory_store with:

• content: Condensed version: "Few-shot [{category}]: {lesson}. Task: {requirement}. Fix: {correction summary}."
• tags: few-shot, {category}, relevant keywords from the task

---

/ops close

Log a completed delivery. Step-by-step using AskUserQuestion. One question at a time.

Flow

Step 1: "Client/project ID?" — free text (becomes id field)

Step 2: "Industry?" — options: saas, consumer, fintech, healthcare, ecommerce, other

Step 3: "Price?" — numeric (USD)

Step 4: "Hours spent?" — numeric

Step 5: "Theme config" — first attempt auto-detection:

grep -r "createTheme" --include="*.ts" --include="*.tsx" -l .

If found, read the file and extract the theme object. Present it for confirmation. If not found, ask manually. Full schema:

Colors (4): primary, secondary, neutral, background
  Each: { hue, sat, isNeutral, isTinted, highContrast }
Size: sm | md | lg
Spacing: compact | regular | relaxed
Typography: { baseFontSize, bodyFont, headingFont, subheadingFont, codeFont }
Radius: { base }
Border: { base }
Shadow: { intensity }
Preset: string

Step 6: "Blocks reused?" — comma-separated list or "none"

Step 7: "New blocks worth extracting?" — comma-separated list or "none"

Step 8: "Bugs found? Specs updated?" — free text or "none"

Step 9: "Capture few-shots?" — yes chains to /ops few-shot, no continues

Step 10: "Notes?" — free text or "none"

Write to SQLite

Generate a UUID for the id if the user gave a project name (use it as prefix):

ID="{project-name}-$(date +%Y%m%d)"

Insert into deliveries DB:

sqlite3 /Users/lucas/Sites/vlt-data/db/deliveries.db "
INSERT INTO deliveries (id, date, industry, price, delivery_hours, modules_selected, modules_rejected, theme_config, blocks_reused, blocks_extracted, specs_updated, bugs_found, few_shots_captured, notes)
VALUES (
  '{id}',
  date('now'),
  '{industry}',
  {price},
  {hours},
  '{modules_selected_json}',
  '[]',
  '{theme_config_json}',
  '{blocks_reused_json}',
  '{blocks_extracted_json}',
  '{specs_updated_json}',
  {bugs_count},
  {few_shots_count},
  '{notes}'
);
"

Update Module Affinity

For each pair of modules in modules_selected, increment co_occurrence_count in module_affinity:

sqlite3 /Users/lucas/Sites/vlt-data/db/deliveries.db "
INSERT INTO module_affinity (module_a, module_b, co_occurrence_count)
VALUES ('{a}', '{b}', 1)
ON CONFLICT(module_a, module_b)
DO UPDATE SET co_occurrence_count = co_occurrence_count + 1,
              updated_at = datetime('now');
"

Then recalculate co_occurrence_pct for all rows:

sqlite3 /Users/lucas/Sites/vlt-data/db/deliveries.db "
UPDATE module_affinity
SET co_occurrence_pct = ROUND(
  co_occurrence_count * 100.0 / (SELECT COUNT(*) FROM deliveries),
  1
);
"

Export JSON

After insert, regenerate the export:

sqlite3 -json /Users/lucas/Sites/vlt-data/db/deliveries.db "SELECT * FROM deliveries ORDER BY date DESC;" > /Users/lucas/Sites/vlt-data/exports/deliveries.json

Confirm

Print summary:

Delivery logged: {id}
Industry: {industry} | Price: ${price} | Hours: {hours}h
Blocks reused: {count} | Extracted: {count} | Bugs: {count}
Export updated: deliveries.json

---

/ops insights

Query the SQLite DB and present formatted results. No user interaction needed — just run and display.

Queries

Module Affinity (top 10 co-occurrences):

SELECT module_a, module_b, co_occurrence_count, co_occurrence_pct
FROM module_affinity
ORDER BY co_occurrence_count DESC
LIMIT 10;

Theme Trends by Industry:

SELECT industry,
       COUNT(*) as deliveries,
       json_extract(theme_config, '$.size') as common_size,
       json_extract(theme_config, '$.spacing') as common_spacing
FROM deliveries
GROUP BY industry
ORDER BY deliveries DESC;

Hours & Pricing:

SELECT
  COUNT(*) as total,
  ROUND(AVG(price), 2) as avg_price,
  ROUND(AVG(delivery_hours), 1) as avg_hours,
  ROUND(AVG(price / delivery_hours), 2) as avg_hourly_rate,
  ROUND(MIN(price), 2) as min_price,
  ROUND(MAX(price), 2) as max_price
FROM deliveries;

Most Reused Blocks (flatten JSON arrays, count occurrences):

SELECT value as block, COUNT(*) as times_used
FROM deliveries, json_each(deliveries.blocks_reused)
GROUP BY value
ORDER BY times_used DESC
LIMIT 15;

Output Format

Present each section as a markdown table with a header. Keep it scannable.

---

/ops status

Quick stats. No interaction — query and display.

SELECT
  COUNT(*) as total_deliveries,
  ROUND(SUM(price), 2) as total_revenue,
  ROUND(AVG(delivery_hours), 1) as avg_hours,
  ROUND(AVG(price / delivery_hours), 2) as avg_hourly_rate,
  ROUND(SUM(price) / SUM(delivery_hours), 2) as blended_rate
FROM deliveries;

Display as:

Deliveries: {total} | Revenue: ${total_revenue}
Avg hours: {avg_hours}h | Avg rate: ${avg_hourly_rate}/h | Blended: ${blended_rate}/h

---

Rules

1. One question at a time. Always AskUserQuestion sequentially, never batch.
2. Auto-detect first. Theme config, git diffs — try to pull from context before asking.
3. Escape SQL values. Single quotes in user input must be escaped before SQLite insertion.
4. Create directories on demand. Few-shot category dirs, exports dir — mkdir -p before writing.
5. JSON arrays for list fields. modules_selected, blocks_reused, etc. — always valid JSON.
6. Export after every write. Regenerate deliveries.json after any DB mutation.
7. Memory is optional. Only store to MCP when user explicitly opts in during few-shot capture.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vlting
• Source: vlting/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.