Talking to LLMs

Talking to LLMs

This skill helps you write effective prompts, commands, and agent instructions.

Core Principle

LLMs are intelligent agents, not script executors. Talk to them like senior engineers.

Anti-Patterns

Over-Prescriptive Instructions

Bad:

Step 1: Run `sentry-cli issues list --status unresolved`
Step 2: Parse the JSON output
Step 3: For each issue, calculate priority score using formula...
Step 4: Select highest priority issue
Step 5: Run `git log --since="24 hours ago"`
...700 more lines

This treats the LLM like a bash script executor. It's brittle, verbose, and removes the LLM's ability to adapt.

Excessive Hand-Holding

Bad:

If the user says X, do Y.
If the user says Z, do W.
Handle edge case A by doing B.
Handle edge case C by doing D.

You can't enumerate every case. Trust the LLM to generalize.

Defensive Over-Specification

Bad:

IMPORTANT: Do NOT do X.
WARNING: Never do Y.
CRITICAL: Always remember to Z.

If you need 10 warnings, your instruction is probably wrong.

Good Patterns

State the Goal, Not the Steps

Good:

Investigate production errors. Check all available observability (Sentry, Vercel, logs).
Correlate with git history. Find root cause. Propose fix.

Let the LLM figure out how.

Provide Context, Not Constraints

Good:

You're a senior SRE investigating an incident.
The user indicated something broke around 14:57.

Frame the situation, don't micromanage the response.

Trust Recovery

Good:

Trust your judgment. If something doesn't work, try another approach.

LLMs can recover from errors. Let them.

Role + Objective + Latitude

The best prompts follow this pattern:

1. Role: Who is the LLM in this context?
2. Objective: What's the end goal?
3. Latitude: How much freedom do they have?

Example:

You're a senior engineer reviewing this PR.           # Role
Find bugs, security issues, and code smells.          # Objective
Be direct. If it's fine, say so briefly.              # Latitude

When Writing Claude Code Commands

Commands are prompts. The same rules apply:

Bad command (700 lines):

• Exhaustive decision trees
• Exact CLI commands to copy
• Every edge case enumerated
• No room for judgment

Good command (20 lines):

• Clear objective
• Context about what tools exist
• Permission to figure it out
• Trust in agent judgment

When Building Agentic Systems

Same principles scale up:

Bad agent design:

• Rigid state machines
• Exhaustive action lists
• No error recovery
• Brittle integrations

Good agent design:

• Goal-oriented
• Self-correcting
• Minimal constraints
• Natural language interfaces

The Test

Before finalizing any LLM instruction, ask:

"Would I give these instructions to a senior engineer?"

If you'd be embarrassed to hand a colleague a 700-line runbook for a simple task, don't give it to the LLM either.

Remember

The L in LLM stands for Language. Use it.