working-with-llms
Mandatory workflow for creating LLM-facing content. Follow the 4-step process (objective → draft → verify → adjust) before writing any prompt, skill, tool description, or system instruction. Triggers on requests to create or revise skills, prompts, agent workflows, or any content that will be sent to an LLM repeatedly.
Best use case
working-with-llms is best used when you need a repeatable AI agent workflow instead of a one-off prompt.
Mandatory workflow for creating LLM-facing content. Follow the 4-step process (objective → draft → verify → adjust) before writing any prompt, skill, tool description, or system instruction. Triggers on requests to create or revise skills, prompts, agent workflows, or any content that will be sent to an LLM repeatedly.
Teams using working-with-llms should expect a more consistent output, faster repeated execution, less prompt rewriting.
When to use this skill
- You want a reusable workflow that can be run more than once with consistent structure.
When not to use this skill
- You only need a quick one-off answer and do not need a reusable workflow.
- You cannot install or maintain the underlying files, dependencies, or repository context.
Installation
Claude Code / Cursor / Codex
Manual Installation
- Download SKILL.md from GitHub
- Place it in
.claude/skills/working-with-llms/SKILL.mdinside your project - Restart your AI agent — it will auto-discover the skill
How working-with-llms Compares
| Feature / Agent | working-with-llms | Standard Approach |
|---|---|---|
| Platform Support | Not specified | Limited / Varies |
| Context Awareness | High | Baseline |
| Installation Complexity | Unknown | N/A |
Frequently Asked Questions
What does this skill do?
Mandatory workflow for creating LLM-facing content. Follow the 4-step process (objective → draft → verify → adjust) before writing any prompt, skill, tool description, or system instruction. Triggers on requests to create or revise skills, prompts, agent workflows, or any content that will be sent to an LLM repeatedly.
Where can I find the source code?
You can find the source code on GitHub using the link provided at the top of the page.
SKILL.md Source
# Working with LLMs ## Workflow Follow this sequence for all LLM-facing content. Do not skip steps. ### Step 1: State the Objective Before writing anything, state the desired outcome explicitly in your response: ``` **Objective:** [One sentence describing what the LLM should do when this content is applied] ``` This checkpoint is visible to the user. Every instruction that follows must directly serve this objective. ### Step 2: Draft Write instructions that serve the objective. Draft as you normally would, but do not present to the user yet - the draft must go through at least one iteration/refinement step before presenting. ### Step 3: Verify Before presenting to the user, ALWAYS launch a sub-agent to explicitly verify draft contents against these criteria: - Is this actionable? (Commands behavior, not describes principles) - Does the model need this? (Would it behave worse without it?) - Each instruction is imperative (do X) not descriptive (X is important) - No speculative "don't" instructions - only prohibitions earned by observed behavior - Context directly serves the objective, not "nice to know" - If guarding against a pattern, there's an explicit verification step, not just a prohibition ### Step 4: Make Adjustments Make any fixes identified in Step 3, then review again against the criteria. Only present to the user once the draft passes verification. ## Principles Reference Use these when evaluating instructions in Steps 2-3: **Token cost.** Context window is shared and expensive. Every token competes for attention. Bloated prompts dilute signal. **Actionability.** Instructions tell the model what to do. If it doesn't command action or inform a decision, delete it. **Positive focus.** Write what to do, not what to avoid. Vague prohibitions create uncertainty; clear directives give something to execute. **Earned negatives.** Add "don't" only after observing the unwanted behavior. Speculative guardrails waste tokens on problems that may never occur. **Verification over prevention.** To guard against patterns, add inspection steps to the workflow rather than hoping prohibitions work. --- **NOTE:** You tend to skip this workflow entirely, especially when creating skills alongside skill-creator. This is not background context to absorb - it is a procedure to execute. Output the `**Objective:**` checkpoint before drafting anything, and verify against the criteria outlined in Step 3 before presenting to the user.
Related Skills
generate-llms
Generate llms.txt and llms-full.txt files for AI agent consumption following the llmstxt.org standard. Use when updating site content that should be reflected in the llms files, or when building/deploying the site.
building-with-llms
Help users build effective AI applications. Use when someone is building with LLMs, writing prompts, designing AI features, implementing RAG, creating agents, running evals, or trying to improve AI output quality.
serving-llms-vllm
Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching. Use when deploying production LLM APIs, optimizing inference latency/throughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ/AWQ/FP8), and tensor parallelism.
bgo
Automates the complete Blender build-go workflow, from building and packaging your extension/add-on to removing old versions, installing, enabling, and launching Blender for quick testing and iteration.
humanizer-ko
Detects and corrects Korean AI writing patterns to transform text into natural human writing. Based on scientific linguistic research (KatFishNet paper with 94.88% AUC accuracy). Analyzes 19 patterns including comma overuse, spacing rigidity, POS diversity, AI vocabulary overuse, and structural monotony. Use when humanizing Korean text from ChatGPT/Claude/Gemini or removing AI traces from Korean LLM output.
huggingface-accelerate
Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
hr-pro
Professional, ethical HR partner for hiring, onboarding/offboarding, PTO and leave, performance, compliant policies, and employee relations. Ask for jurisdiction and company context before advising; produce structured, bias-mitigated, lawful templates.
hive-mind-advanced
Advanced Hive Mind collective intelligence system for queen-led multi-agent coordination with consensus mechanisms and persistent memory
hire
Interactive hiring wizard to set up a new AI team member. Guides the user through role design via conversation, generates agent identity files, and optionally sets up performance reviews. Use when the user wants to hire, add, or set up a new AI agent, team member, or assistant. Triggers on phrases like "hire", "add an agent", "I need help with X" (implying a new role), or "/hire".
hic-tad-calling
This skill should be used when users need to identify topologically associating domains (TADs) from Hi-C data in .mcools (or .cool) files or when users want to visualize the TAD in target genome loci. It provides workflows for TAD calling and visualization.
helix-memory
Long-term memory system for Claude Code using HelixDB graph-vector database. Store and retrieve facts, preferences, context, and relationships across sessions using semantic search, reasoning chains, and time-window filtering.
heath-ledger
AI bookkeeping agent for Mercury bank accounts. Pulls transactions, categorizes them (rule-based + AI), and generates Excel workbooks with P&L, Balance Sheet, Cash Flow, and transaction detail. Use when the user wants to do bookkeeping, generate financial statements, categorize bank transactions, connect Mercury, or produce monthly/quarterly/annual books. Triggers on: bookkeeping, P&L, profit and loss, balance sheet, cash flow, financial statements, Mercury bank, categorize transactions, generate books, monthly close.