$ cat ~/tips/

Tips

Prompts, workflows, and strategies I actually use. No fluff — just things that work.

5 tips

PromptPromptingSaves tokens

Zero-Filler Mode

When → Starting a session where you want concise, tool-first responses with no narration.

From now on: no filler words. Short 3–6 word sentences. Run tools first, show result, stop. No narration. Instead of 'The solution is to use async' → 'Use async.'

Why → LLMs default to verbose, explanatory output. This prompt overrides that and can cut response length 40–60%, saving tokens and reducing noise.

PromptWorkflowSaves tokens

Context Snapshot (COMPACT)

When → Hitting context limits or transferring work to a new chat session.

Summarize this conversation: 5–7 bullet points. Include all decisions, file paths, code snippets, and current state. Format for paste into a new session.

Why → LLMs lose context at window limits. A structured snapshot preserves decisions and state across sessions without re-explaining everything from scratch.

StrategyModelsSaves tokens

Model Routing

When → Every task. Default to the lightest model that handles the job.

  1. 1.Opus → architecture, judgment calls, long-range planning, complex debugging
  2. 2.Sonnet → implementation, analysis, code generation, summaries
  3. 3.Haiku → quick lookups, simple transforms, lightweight automation

Why → Haiku isn't worse—it's faster and lighter. Most coding tasks don't need Opus-level reasoning. Routing by complexity saves tokens and speeds up iteration without quality loss.

WorkflowToken SavingSaves tokens

PDF Pre-Processing

When → Sending document content to Claude—PDFs, long specs, research papers.

  1. 1.Upload the PDF to a free-tier LLM (Gemini, ChatGPT free, etc.)
  2. 2.Prompt: 'Extract core information. Remove all filler, formatting, and boilerplate. Return condensed plain text with key points only.'
  3. 3.Paste the condensed output into Claude

Why → Raw PDFs contain formatting tokens, headers, footers, and whitespace that inflate context. Pre-condensing with a free model strips noise and can reduce token usage 50–70%.

StrategyWorkflow

Usage Window Reset

When → Planning a heavy Claude Code session—large refactors, multi-file changes.

  1. 1.Send a lightweight message 2–3 hours before your real work session
  2. 2.This starts the 5-hour usage window early
  3. 3.When heavy usage begins later, the window resets sooner
  4. 4.Example: 6am → send 'hi' | 8am → start work | ~10am → potential reset

Why → Claude's rate limits operate on a rolling window from first message. Front-loading the timer means your reset aligns with when you actually need it.