1
Read Slack thread Collects the affected user, error details, and approximate time of the issue from the Slack thread where /debug was invoked
2
Investigate Simultaneously queries Sentry (error traces), GitHub (code/commits in the swarm repo), and Metabase (SQL on affected user data)
3
Classify issue Determines whether the issue is a user error or product bug, and whether the fix requires a script or a PR
4
Create Linear ticket If a PR is needed, creates a fully populated Linear ticket with root cause, steps to reproduce, and implementation notes
5
Draft email If the thread is in #bot-bugs-channel, saves a customer-facing email draft in Gmail for human review
6
Post report Posts a formatted debug report back into the original Slack thread

The complete skill-summaries.md file that Claude reads when this skill is invoked. Scroll to read the full instructions, or copy them to use in your own setup.

Skills & Commands Summary

---

#1. debug-slack · skill

Command: /debug

Summary: Full-stack debugging assistant that investigates bugs reported in Slack threads, classifies them, and posts a structured findings report — automatically.

How it works:

Reads the Slack thread where /debug was invoked to collect the affected user, error details, and approximate time of the issue
Simultaneously investigates using Sentry (error traces), GitHub (code/commits in the swarm repo), and Metabase (SQL queries on affected user data)
Classifies the issue as a user error or product bug, and determines whether the fix is a script or a PR
If a PR is needed, creates a fully populated Linear ticket with root cause, steps to reproduce, and implementation notes
If the thread is in #bot-bugs-channel, saves a customer-facing email draft in Gmail for human review before sending
Posts a formatted debug report back into the original Slack thread

Required connectors: Slack, Sentry, GitHub, Metabase, Linear, Gmail

Instructions file: .skills/skills/debug-slack/SKILL.md

---

#2. docx · skill

Command: *(auto-triggered)*

Summary: Creates, edits, and analyzes Microsoft Word (.docx) files with full support for formatting, tracked changes, tables, images, and comments.

How it works:

For new documents: generates .docx files using the docx JavaScript library, with explicit control over page size, margins, styles, lists, tables, images, hyperlinks, footnotes, headers/footers, and table of contents
For editing existing documents: unpacks the .docx ZIP archive to raw XML, applies changes directly (tracked changes, comments, insertions, deletions), then repacks and validates
Validates output after every creation or edit; auto-repairs common XML issues and flags schema violations
Converts .doc legacy files to .docx using LibreOffice before editing

Required connectors: None (runs locally)

Instructions file: .skills/skills/docx/SKILL.md

---

#3. pdf · skill

Command: *(auto-triggered)*

Summary: Handles all PDF operations — reading, creating, merging, splitting, rotating, watermarking, form-filling, encrypting, and OCR on scanned documents.

How it works:

Uses pdfplumber for text and table extraction with layout preservation
Uses pypdf for page operations: merging, splitting, rotating, watermarking, and password protection
Uses reportlab to create new PDFs from scratch (multi-page, styled documents with tables and images)
Uses pytesseract + pdf2image for OCR on scanned PDFs that contain no selectable text
Uses command-line tools (qpdf, pdftotext) for fast bulk operations
For form-filling, follows a separate detailed workflow documented in FORMS.md

Required connectors: None (runs locally)

Instructions file: .skills/skills/pdf/SKILL.md

---

#4. pptx · skill

Command: *(auto-triggered)*

Summary: Creates and edits PowerPoint presentations (.pptx), with design guidance for visually compelling slides and mandatory visual QA before delivery.

How it works:

Reading: extracts slide text via markitdown and generates thumbnail grids for visual overview
Creating from scratch: uses pptxgenjs (Node.js) with bold color palettes, varied layouts, and strong typographic contrast — never plain bullets on white
Editing templates: unpacks the .pptx ZIP, manipulates slide XML, then repacks
QA (required): converts slides to JPG images using LibreOffice + poppler, then uses a subagent with fresh eyes to inspect for overlaps, text overflow, low contrast, and leftover placeholder text
Iterates through a fix-and-verify loop until a full pass finds no new issues

Required connectors: None (runs locally)

Instructions file: .skills/skills/pptx/SKILL.md

---

#5. xlsx · skill

Command: *(auto-triggered)*

Summary: Creates, edits, and analyzes Excel spreadsheets (.xlsx) with formulas, formatting, financial color conventions, and automatic formula error checking.

How it works:

Uses pandas for data analysis, bulk operations, and simple exports
Uses openpyxl for formula-driven models, formatting, multi-sheet work, and financial conventions (blue for inputs, black for formulas, green for cross-sheet links)
Always writes Excel formulas (e.g. =SUM(B2:B9)) rather than hardcoding Python-calculated values, so spreadsheets stay dynamic
Runs scripts/recalc.py after every save to force LibreOffice to recalculate all formulas and scan for errors (#REF!, #DIV/0!, #VALUE!, #NAME?)
For financial models: enforces industry-standard color coding, parentheses for negatives, sourced comments on hardcoded values, and separate assumption cells

Required connectors: None (runs locally)

Instructions file: .skills/skills/xlsx/SKILL.md

---

#6. skill-creator · skill

Command: *(auto-triggered)*

Summary: Builds new skills from scratch and iteratively improves existing ones through a test-evaluate-refine loop with quantitative benchmarking and a browser-based review viewer.

How it works:

Conducts a discovery interview to understand the skill's purpose, trigger conditions, and expected outputs
Drafts a SKILL.md with name, description, and instructions — structured for progressive disclosure (metadata → body → bundled references)
Generates 2–3 realistic test prompts, runs them in parallel via subagents (with-skill vs. baseline), and launches an HTML eval viewer for human review
Grades outputs against quantitative assertions, aggregates benchmark stats (pass rate, tokens, duration), and surfaces patterns like non-discriminating assertions or flaky evals
Iterates on the skill based on feedback until the user is satisfied
Optionally runs a description optimization loop (run_loop.py) to maximize triggering accuracy on a held-out eval set
Packages the final skill as a .skill file for installation

Required connectors: None (runs locally via subagents)

Instructions file: .skills/skills/skill-creator/SKILL.md

---

#7. cowork-plugin-customizer · skill

Command: *(auto-triggered)*

Summary: Customizes an installed Cowork plugin for your organization's specific tools, team names, channels, and workflows — by searching Slack and docs to pre-fill answers automatically.

How it works:

Detects whether the plugin has ~~-prefixed generic placeholders (first-time setup) or is already configured (targeted edits)
Searches Slack, internal documents, and email to gather tool names, project identifiers, channel names, and team conventions — so it can fill in most answers without asking
Presents a todo list of customization items in plain language (never exposes placeholder syntax to the user)
Uses AskUserQuestion only when a specific value couldn't be found in internal sources
Connects relevant MCP integrations for any tools identified during setup
Packages the customized plugin as a .plugin file for re-installation

Required connectors: Depends on plugin being customized (Slack, Linear, etc.)

Instructions file: .local-plugins/.../cowork-plugin-customizer/SKILL.md

---

#8. create-cowork-plugin · skill

Command: *(auto-triggered)*

Summary: Guides you through designing and building a new Cowork plugin from scratch — covering discovery, component planning, implementation, and delivery as a ready-to-install .plugin file.

How it works:

Runs a five-phase guided workflow: Discovery → Component Planning → Design & Clarifying Questions → Implementation → Review & Package
Determines which component types are needed: skills (on-demand knowledge), commands (user-initiated actions), MCP servers (external integrations), agents (autonomous tasks), and hooks (event-driven automation)
Writes all plugin files following the standard directory layout (plugin.json, commands/, skills/, .mcp.json, README.md)
Uses ${CLAUDE_PLUGIN_ROOT} for all intra-plugin paths so the plugin is portable
Validates the plugin with claude plugin validate, then zips and delivers it as a .plugin file
Supports distributable "template" plugins with ~~category placeholders for organizations that want to share plugins externally

Required connectors: Depends on what the plugin integrates with

Instructions file: .local-plugins/.../create-cowork-plugin/SKILL.md

---

#9. schedule-task · skill

Command: *(auto-triggered)*

Summary: Saves a repeatable version of any task from the current session as a scheduled shortcut that can run on demand or automatically on a cron interval.

How it works:

Reviews the current session to distill the core task into a single, self-contained repeatable objective
Writes a second-person imperative prompt with all necessary context (file paths, URLs, tools, success criteria) so future runs work cold without session history
Assigns a kebab-case task name and determines whether the task is one-off or recurring
For recurring tasks, proposes an appropriate cron expression in the user's local timezone (not UTC)
Calls create_scheduled_task to register the shortcut for future use

Required connectors: Depends on what the task does

Instructions file: .claude/skills/schedule-task/SKILL.md

# Skills & Commands Summary

---

## #1. *debug-slack* · skill

**Command:** `/debug`

**Summary:** Full-stack debugging assistant that investigates bugs reported in Slack threads, classifies them, and posts a structured findings report — automatically.

**How it works:**
- Reads the Slack thread where `/debug` was invoked to collect the affected user, error details, and approximate time of the issue
- Simultaneously investigates using Sentry (error traces), GitHub (code/commits in the `swarm` repo), and Metabase (SQL queries on affected user data)
- Classifies the issue as a **user error** or **product bug**, and determines whether the fix is a script or a PR
- If a PR is needed, creates a fully populated Linear ticket with root cause, steps to reproduce, and implementation notes
- If the thread is in `#bot-bugs-channel`, saves a customer-facing email draft in Gmail for human review before sending
- Posts a formatted debug report back into the original Slack thread

**Required connectors:** Slack, Sentry, GitHub, Metabase, Linear, Gmail

**Instructions file:** `.skills/skills/debug-slack/SKILL.md`

---

## #2. *docx* · skill

**Command:** *(auto-triggered)*

**Summary:** Creates, edits, and analyzes Microsoft Word (.docx) files with full support for formatting, tracked changes, tables, images, and comments.

**How it works:**
- For **new documents**: generates `.docx` files using the `docx` JavaScript library, with explicit control over page size, margins, styles, lists, tables, images, hyperlinks, footnotes, headers/footers, and table of contents
- For **editing existing documents**: unpacks the `.docx` ZIP archive to raw XML, applies changes directly (tracked changes, comments, insertions, deletions), then repacks and validates
- Validates output after every creation or edit; auto-repairs common XML issues and flags schema violations
- Converts `.doc` legacy files to `.docx` using LibreOffice before editing

**Required connectors:** None (runs locally)

**Instructions file:** `.skills/skills/docx/SKILL.md`

---

## #3. *pdf* · skill

**Command:** *(auto-triggered)*

**Summary:** Handles all PDF operations — reading, creating, merging, splitting, rotating, watermarking, form-filling, encrypting, and OCR on scanned documents.

**How it works:**
- Uses **pdfplumber** for text and table extraction with layout preservation
- Uses **pypdf** for page operations: merging, splitting, rotating, watermarking, and password protection
- Uses **reportlab** to create new PDFs from scratch (multi-page, styled documents with tables and images)
- Uses **pytesseract + pdf2image** for OCR on scanned PDFs that contain no selectable text
- Uses command-line tools (`qpdf`, `pdftotext`) for fast bulk operations
- For form-filling, follows a separate detailed workflow documented in `FORMS.md`

**Required connectors:** None (runs locally)

**Instructions file:** `.skills/skills/pdf/SKILL.md`

---

## #4. *pptx* · skill

**Command:** *(auto-triggered)*

**Summary:** Creates and edits PowerPoint presentations (.pptx), with design guidance for visually compelling slides and mandatory visual QA before delivery.

**How it works:**
- **Reading**: extracts slide text via `markitdown` and generates thumbnail grids for visual overview
- **Creating from scratch**: uses `pptxgenjs` (Node.js) with bold color palettes, varied layouts, and strong typographic contrast — never plain bullets on white
- **Editing templates**: unpacks the `.pptx` ZIP, manipulates slide XML, then repacks
- **QA (required)**: converts slides to JPG images using LibreOffice + poppler, then uses a subagent with fresh eyes to inspect for overlaps, text overflow, low contrast, and leftover placeholder text
- Iterates through a fix-and-verify loop until a full pass finds no new issues

**Required connectors:** None (runs locally)

**Instructions file:** `.skills/skills/pptx/SKILL.md`

---

## #5. *xlsx* · skill

**Command:** *(auto-triggered)*

**Summary:** Creates, edits, and analyzes Excel spreadsheets (.xlsx) with formulas, formatting, financial color conventions, and automatic formula error checking.

**How it works:**
- Uses **pandas** for data analysis, bulk operations, and simple exports
- Uses **openpyxl** for formula-driven models, formatting, multi-sheet work, and financial conventions (blue for inputs, black for formulas, green for cross-sheet links)
- Always writes Excel formulas (e.g. `=SUM(B2:B9)`) rather than hardcoding Python-calculated values, so spreadsheets stay dynamic
- Runs `scripts/recalc.py` after every save to force LibreOffice to recalculate all formulas and scan for errors (`#REF!`, `#DIV/0!`, `#VALUE!`, `#NAME?`)
- For financial models: enforces industry-standard color coding, parentheses for negatives, sourced comments on hardcoded values, and separate assumption cells

**Required connectors:** None (runs locally)

**Instructions file:** `.skills/skills/xlsx/SKILL.md`

---

## #6. *skill-creator* · skill

**Command:** *(auto-triggered)*

**Summary:** Builds new skills from scratch and iteratively improves existing ones through a test-evaluate-refine loop with quantitative benchmarking and a browser-based review viewer.

**How it works:**
- Conducts a discovery interview to understand the skill's purpose, trigger conditions, and expected outputs
- Drafts a `SKILL.md` with name, description, and instructions — structured for progressive disclosure (metadata → body → bundled references)
- Generates 2–3 realistic test prompts, runs them in parallel via subagents (with-skill vs. baseline), and launches an HTML eval viewer for human review
- Grades outputs against quantitative assertions, aggregates benchmark stats (pass rate, tokens, duration), and surfaces patterns like non-discriminating assertions or flaky evals
- Iterates on the skill based on feedback until the user is satisfied
- Optionally runs a description optimization loop (`run_loop.py`) to maximize triggering accuracy on a held-out eval set
- Packages the final skill as a `.skill` file for installation

**Required connectors:** None (runs locally via subagents)

**Instructions file:** `.skills/skills/skill-creator/SKILL.md`

---

## #7. *cowork-plugin-customizer* · skill

**Command:** *(auto-triggered)*

**Summary:** Customizes an installed Cowork plugin for your organization's specific tools, team names, channels, and workflows — by searching Slack and docs to pre-fill answers automatically.

**How it works:**
- Detects whether the plugin has `~~`-prefixed generic placeholders (first-time setup) or is already configured (targeted edits)
- Searches Slack, internal documents, and email to gather tool names, project identifiers, channel names, and team conventions — so it can fill in most answers without asking
- Presents a todo list of customization items in plain language (never exposes placeholder syntax to the user)
- Uses `AskUserQuestion` only when a specific value couldn't be found in internal sources
- Connects relevant MCP integrations for any tools identified during setup
- Packages the customized plugin as a `.plugin` file for re-installation

**Required connectors:** Depends on plugin being customized (Slack, Linear, etc.)

**Instructions file:** `.local-plugins/.../cowork-plugin-customizer/SKILL.md`

---

## #8. *create-cowork-plugin* · skill

**Command:** *(auto-triggered)*

**Summary:** Guides you through designing and building a new Cowork plugin from scratch — covering discovery, component planning, implementation, and delivery as a ready-to-install `.plugin` file.

**How it works:**
- Runs a five-phase guided workflow: Discovery → Component Planning → Design & Clarifying Questions → Implementation → Review & Package
- Determines which component types are needed: skills (on-demand knowledge), commands (user-initiated actions), MCP servers (external integrations), agents (autonomous tasks), and hooks (event-driven automation)
- Writes all plugin files following the standard directory layout (`plugin.json`, `commands/`, `skills/`, `.mcp.json`, `README.md`)
- Uses `${CLAUDE_PLUGIN_ROOT}` for all intra-plugin paths so the plugin is portable
- Validates the plugin with `claude plugin validate`, then zips and delivers it as a `.plugin` file
- Supports distributable "template" plugins with `~~category` placeholders for organizations that want to share plugins externally

**Required connectors:** Depends on what the plugin integrates with

**Instructions file:** `.local-plugins/.../create-cowork-plugin/SKILL.md`

---

## #9. *schedule-task* · skill

**Command:** *(auto-triggered)*

**Summary:** Saves a repeatable version of any task from the current session as a scheduled shortcut that can run on demand or automatically on a cron interval.

**How it works:**
- Reviews the current session to distill the core task into a single, self-contained repeatable objective
- Writes a second-person imperative prompt with all necessary context (file paths, URLs, tools, success criteria) so future runs work cold without session history
- Assigns a kebab-case task name and determines whether the task is one-off or recurring
- For recurring tasks, proposes an appropriate cron expression in the **user's local timezone** (not UTC)
- Calls `create_scheduled_task` to register the shortcut for future use

**Required connectors:** Depends on what the task does

**Instructions file:** `.claude/skills/schedule-task/SKILL.md`

Release Notes