Prompt Engineering for Writers, Part 1: Weak vs Professional Prompts, COSTAR, a Five-Email Worked Example, the Meta-Prompt, and Three Prompting Modes

Context: only relevant background ('B2B SaaS for marketing managers at mid-market firms; LinkedIn in Q4; 40% churn trial-to-paid'). Objective: the success condition (raise trial-to-paid; position the CEO; capture emails). Steps: numbered reasoning - analyse pain points, identify benefits, create emotional hooks, structure the argument, write CTAs - self-reasoning improves output. Tools: frameworks, data, formats (McKinsey framework, HTML template, SWOT). Actions: the output spec ('5 email templates with subject lines, preview text and send timing'). Reflection: 'before finalising, check brand consistency, clear value proposition and a compelling CTA, and suggest one improvement per email'. Worked example in Claude: project-management SaaS losing 65% of trials; five emails (welcome, three-click setup, team momentum, pro tip, first win) with self-scores; 'consider the brand voice to be that of Asana'. Note: this differs from the published CO-STAR (Style, Tone, Audience, Response).

Prompt: 'I want to build a prompt using the COSTAR framework used above. I want to write a LinkedIn post about how AI can improve content creation. Ask me clarifying questions if needed.' He answers: education sector teaching professionals GenAI; thought leadership on ideation; professional but conversational, data-driven, storytelling, educational; audience 5+ years' experience; CTA comment and share; personal experience of weekly posting. Copy the generated COSTAR prompt into a new chat; the post ('The blank page used to paralyze me...') needs few edits. Define COSTAR in the same chat or spell out its parts - every LLM has its own version; re-paste if it drifts. Screenshot the example to reuse.

Rabbit hole (learning): start broad, pick a thread, go deeper; a five-level pyramid two levels broader and two narrower than the target. Brainstorming: not 'help me brainstorm product ideas' but 'based on my previous products, target market and these business goals, brainstorm ideas that lead to this outcome' - COSTAR with two or three paragraphs of context. Automation: run the 2-3-stage, 3-4-prompt workflow manually five to ten times, refine, reverse-engineer a repeatable prompt, then move it into Claude Projects, a custom GPT or a Gemini Gem with its context (brand guidelines, product specs, audience research; for fiction, world and characters).