Convert the request into a managed batch:

{{args}}

Create a short execution plan, group related work, identify dependencies, and decide what can run in parallel. Use subagents for independent investigation or verification. Keep each task small enough to review and test.

Output:

PLAN: ordered batch list. PARALLEL: items safe to run concurrently. GATES: validation required before proceeding. RISKS: the main ways the batch can go wrong.

Answer this Claude API request:

{{args}}

Because API details change, verify current facts with official documentation or the local codebase before giving precise model names, parameters, pricing, or SDK behavior. Provide concise implementation guidance and cite URLs when web tools influence the answer.

Coordinate browser-assisted work:

{{args}}

First check whether a browser/Chrome MCP server is configured. If available, use it to inspect, click, type, or capture page state. If unavailable, explain the missing connector and use web fetch/search only for public pages.

Debug this issue:

{{args}}

Work like an investigator:

1. Reconstruct the symptom and expected behavior.
2. Find the narrowest relevant code path.
3. Gather evidence from tests, logs, diffs, and nearby code.
4. State the most likely root cause and confidence.
5. Recommend the smallest fix and validation.

Do not make broad edits from inside this skill unless explicitly asked.

Delegate

Use this skill when the task is large enough to benefit from parallel or staged work.

Recommended pattern:

1. Open explore agents for independent evidence gathering.
2. Open plan when the implementation path is unclear.
3. Open implementer only after the desired change is specific.
4. Open review or verifier to check the result.
5. Merge child results by citing evidence, changes, risks, and blockers.

User task: {{args}}

Documents

Use this skill for reports, memos, forms, letters, and DOCX-like deliverables.

Work with a clear structure:

• Identify audience, purpose, required sections, and output format.
• Inspect any existing source files before editing.
• Prefer plain Markdown drafts when native document libraries are unavailable.
• Verify page headings, tables, references, and final file paths.

User task: {{args}}

Feishu / Lark

Use this skill for Feishu or Lark integration work.

Rules:

• Never hardcode app secrets, tenant tokens, or user tokens.
• Keep credentials in environment variables or local ignored config.
• Separate OpenAPI auth, resource IDs, and business logic.
• Add dry-run or read-only modes before write operations.
• Document required scopes and callback URLs.

User task: {{args}}

Design a recurring loop for:

{{args}}

Clarify the trigger, cadence, stop condition, output destination, and safety gate. Use automation tools when the loop should persist. If persistence is not available, create a tracked task plan instead.

Generate placeholder content for:

{{args}}

Match the requested domain, tone, and length. Avoid pretending the placeholder is final product copy. If no constraints are supplied, produce concise neutral English text suitable for UI mockups.

MCP Builder

Use this skill when creating or debugging MCP-style tool servers.

Checklist:

• Define tools, resources, and prompts separately.
• Keep schemas strict and small.
• Make startup errors visible.
• Support a simple local smoke test.
• Avoid leaking secrets through tool output.
• Prefer stdio or JSON-RPC-lines first, then add HTTP/SSE only when needed.

User task: {{args}}

PDF

Use this skill for PDF work.

Procedure:

• Inspect file existence, size, and page count when possible.
• Extract text before summarizing.
• For edits, keep the original file and write a new output path.
• Verify output page count and text extraction after conversion.
• If OCR or PDF libraries are unavailable, state the missing dependency and provide the exact next command the user can run.

User task: {{args}}

Presentations

Use this skill for decks, talks, and slide outlines.

Keep slides practical:

• Start with audience, goal, duration, and slide count.
• Use one message per slide.
• Prefer concise titles and speaker notes.
• Verify order, repeated headers, image paths, and export location.
• If native PPTX tooling is unavailable, produce a Markdown slide outline that can be converted later.

User task: {{args}}

Evaluate whether this should be saved as memory:

{{args}}

Save only durable, useful information:

• Stable user preferences.
• Project conventions.
• Reusable facts that will help future work.

Do not save secrets, transient tasks, private data that is not clearly useful, or facts that may change frequently. If saving, choose a short name, concise text, a kind such as preference, project, or fact, and the right scope.

Generate a complete skill package for:

{{args}}

Create a directory containing SKILL.md and any useful companion files. Keep the skill contract explicit: name, description, allowed tools, when_to_use, context, and concrete instructions. Validate that the new skill loads through skill_list or load_skill.

Schedule this future action:

{{args}}

Extract the exact time or recurrence, timezone if present, action prompt, and success condition. If timing is ambiguous, ask one concise clarification. Store the schedule through automation tools when possible.

Simplify the target while preserving behavior and intent:

{{args}}

Prefer removing accidental complexity, duplicated logic, unclear naming, and unnecessary branching. Keep changes small and validate if code changes are made.

Report what became simpler and what validation was performed.

Skill Creator

Use this skill when adding reusable agent behavior.

Skill design rules:

• A skill is guidance plus lightweight assets, not executable business logic.
• Use a tool when structured execution is required.
• Use MCP when an external service has multiple tools/resources.
• Keep frontmatter complete: name, description, and mode.
• Include trigger guidance, workflow, verification, and failure handling.

User task: {{args}}

Skill Installer

Use this skill when inspecting or installing skills.

Safety checklist:

• Inspect SKILL.md before trusting a skill.
• List companion files and any executable scripts.
• Prefer project-local installs under .lilbot/skills.
• Do not run installer scripts without explicit permission.
• Record source, version, and trust decision.

User task: {{args}}

Create or improve a LilBot skill from this workflow:

{{args}}

Target format:

---
name: short-kebab-name
description: One sentence describing the capability.
allowed-tools: read_file, grep_files
when_to_use: Use when ...
context: inline
---
Actionable instructions for the model.

Rules:

• Keep the skill specific enough to be useful.
• Include only tools that are actually needed.
• Use context: fork only for self-contained work that can run independently.
• Add companion files only when they reduce prompt size or improve accuracy.

Spreadsheets

Use this skill for XLSX, CSV, TSV, formulas, charts, and tabular reports.

Workflow:

• Inspect columns, row count, encodings, and missing values.
• Preserve source files; write cleaned outputs separately.
• Make transformations explicit and repeatable.
• Validate totals, formulas, date parsing, and header names.
• Prefer CSV for simple outputs when spreadsheet libraries are unavailable.

User task: {{args}}

Unstick this work:

{{args}}

Audit the current state:

1. What is actually known?
2. What assumption might be false?
3. What evidence is missing?
4. What is the smallest reversible next step?
5. Should a specialist subagent inspect a narrower question?

Return a short recovery plan with concrete commands/files to inspect next.

Inspect and update configuration for this request:

{{args}}

Find the relevant config file before editing. Preserve unrelated settings. Explain the exact setting changed and how to verify it. If the requested setting does not exist, propose the smallest compatible addition.

V4 Best Practices

Use this skill on multi-step work with DeepSeek V4 models.

Rules:

• Verify current facts and local file state before committing to a plan.
• Keep plans short, concrete, and status-tracked.
• Avoid stale references after file edits; re-read the area you changed.
• Prefer dedicated tools over shell when available.
• End with evidence, changed files, tests, risks, and blockers.

User task: {{args}}

Verify the work described by the user or by these arguments:

{{args}}

Produce a compact verdict:

STATUS: PASS, FAIL, or PARTIAL. EVIDENCE: commands run, files inspected, and exact outcomes. RISKS: unresolved edge cases or missing coverage. NEXT: smallest next action if not PASS.

Prefer concrete validation over reassurance. If tests cannot be run, explain why and substitute targeted code inspection with file references.

Help with keybindings:

{{args}}

Check existing configuration when available. Prefer familiar shortcuts, avoid conflicts, and note terminal limitations for modifier keys. Return a concise mapping table and any config changes needed.

Plugin Creator

Use this skill when the user wants a plugin or external tool wrapper.

Plugin contract:

• One narrow tool per script.
• JSON input on stdin and JSON result on stdout.
• Frontmatter or manifest includes name, description, schema, and approval.
• Fail closed on invalid input.
• Include a smoke test command and an example call.

User task: {{args}}