Tiger Tracks · Eye of the Tiger · AI and Automation · October 2026

AI agents are moving from pilots into production. In McKinsey's 2026 survey, 40% of respondents at organizations with more than $1 billion in annual revenue report scaling AI agents, up from 27% a year earlier, while the share at smaller organizations held flat at 22% [1]. As more work runs through agents, the instructions behind them carry more weight.

AI agents rarely fail because the model is bad. They fail because the prompt is vague.

1. Vague Prompts Produce Vague Agents

A prompt is the instruction set your AI agent runs on. When that instruction is unclear, the agent fills in the gaps with assumptions, and those assumptions rarely match what you actually want. The result is output you have to rewrite, rework or discard entirely.

Most people treat prompting as typing a sentence and hoping for the best. Prompt engineering is a discipline. It involves understanding how language models interpret context, how to structure constraints and how to test variations systematically. Anthropic's own guidance frames the challenge well: it compares Claude to "a brilliant but new employee who lacks context on your norms and workflows," and notes that the more precisely you explain what you want, the better the result [2].

Claude Can Draft the Prompt for You

Most marketers do not have time to develop prompt engineering skills from scratch, and they do not need to. Claude reasons well about language, context and intent, so you can ask it to draft, critique and optimize a prompt on your behalf. Anthropic recommends a similar route for developers: its prompt engineering overview points anyone without a first draft to a metaprompt recipe in the Claude Cookbook that generates one, and it advises defining success criteria and ways to test against them before tuning a prompt at all [3].

2. Four Elements Define a High-Performance Prompt

Understanding these elements lets you direct Claude more precisely and audit the prompts it produces.

Role

Tell the AI who it is. "You are a direct-response copywriter with 10 years of experience writing Facebook ads for e-commerce brands" produces fundamentally different output than "write me an ad." Anthropic's documentation notes that setting a role in the system prompt focuses Claude's behavior and tone, and that even a single sentence makes a difference [2].

Context

Give the AI the situation it is operating in. Include the product, the audience, the platform, the goal and any constraints. The more specific the context, the less the model has to guess. Explaining why a constraint matters also helps, because Claude generalizes from the explanation [2].

Task

State the exact deliverable. Not "write something about our product" but "write three 125-character Facebook ad headlines targeting women aged 28 to 44 who have previously purchased skincare products."

Format

Specify the output structure. If you want a table, say so. If you want numbered options, say so. If you want the output in JSON for an API call, say so. Anthropic recommends including three to five relevant, diverse examples and wrapping distinct parts of a prompt in XML tags to reduce misinterpretation [2].

One caution applies to every element. Claude will generate a prompt that sounds professional even when your brief is incomplete, so always verify that the prompt it produces reflects your real constraints. A polished prompt built on a vague brief still produces vague output.

3. A Four-Step Loop Turns Claude Into Your Prompt Engineer

The workflow is straightforward, and you do not need to be a developer or a technical user to run it.

Step 1: Brief Claude on the Task

Start by describing what you are trying to accomplish in plain language. Tell Claude the tool you are using, the output you want and any constraints that matter. Do not worry about formatting this perfectly. Claude's job is to turn your rough brief into a structured prompt.

Example brief to Claude: "I need a prompt for my AI writing agent. I want it to write LinkedIn posts for Tiger Tracks. The posts should be direct and confident, not fluffy. Each post should have a hook, a 3 to 5 line body, and a question at the end. No hashtags in the body."

Step 2: Ask Claude to Generate Multiple Variants

Request at least three prompt variants. Different structural approaches produce meaningfully different outputs from AI agents, and having options lets you test rather than guess. Ask Claude to explain the reasoning behind each variant as well. Understanding why a prompt is structured a certain way builds your own prompt intuition over time and makes you a better briefer.

Step 3: Test Each Prompt Against Your AI Agent

Run each Claude-generated prompt through your actual AI agent and evaluate the outputs against quality criteria you set in advance. Anthropic treats clear success criteria and a way to test against them as prerequisites for prompt engineering [3]. Note which structural choices produced the best results.

Step 4: Return to Claude with the Results

Feed the outputs back to Claude and ask it to diagnose what worked and what did not. Claude can identify patterns in the failures and produce a refined prompt that addresses them. Anthropic describes the same pattern for its own models as self-correction: generate a draft, have Claude review it against criteria, then have Claude refine it based on the review [2]. This iteration loop is what separates professional prompt engineering from guessing.

4. A Prompt Library Compounds the Gains

One-off prompts are inefficient. The real leverage comes from building a reusable library of tested, optimized prompts for your most common tasks.

What to Include in Your Prompt Library

Identify the five to ten tasks you use AI agents for most frequently. For each one, work with Claude to develop a master prompt template with clearly marked variables. A template for writing ad copy might look like this:

"You are a direct-response copywriter specializing in [PLATFORM] ads for [INDUSTRY] brands. Write [NUMBER] headline variants for [PRODUCT NAME], a [PRICE POINT] [PRODUCT DESCRIPTION] targeting [AUDIENCE DESCRIPTION]. Each headline must be under [CHARACTER LIMIT] characters. Tone: [TONE]. Do not use [PROHIBITED PHRASES]. Return as a numbered list."

The bracketed variables are the only things that change between campaigns. The structural logic stays constant.

How Claude Maintains the Library

Ask Claude to act as the librarian. When you encounter a task that your existing templates do not cover well, brief Claude on the gap and have it draft a new template. When a template underperforms, bring the failure case to Claude and ask it to diagnose and update the template.

Prompt libraries also go stale, because models, platform requirements and audience language all change. Anthropic's current documentation, for example, says Claude 4.6 and later models use adaptive thinking, where Claude decides when and how much to think, and shows developers how to replace manual thinking budgets with an effort setting [2]. Schedule a quarterly review with Claude to audit your library and update templates whose output quality is slipping.

5. Agentic Workflows Need Engineered Prompts

Single-task prompts are the entry point. The larger payoff comes when you use Claude to engineer prompts for multi-step agentic workflows.

Reasoning Instructions

Claude can build prompts that instruct your AI agent to reason through a problem before producing an output, which helps with complex tasks like campaign strategy, audience segmentation or competitive analysis. Anthropic's guidance for its current models is that a general instruction such as "think thoroughly" often produces better reasoning than a hand-written step-by-step plan, and that asking the model to verify its answer against test criteria before it finishes catches errors reliably [2].

Role Stacking

Claude can engineer prompts that assign multiple roles to an AI agent within a single workflow. For example, a prompt can instruct the agent to first act as a market researcher, then as a copywriter, then as a brand compliance reviewer. Anthropic notes that its current models handle most multistep reasoning internally, but that explicit prompt chaining, splitting the task into sequential calls so you can log, evaluate or branch at any point, is still useful when you need to inspect each stage [2].

Constraint Layering

Claude is good at building prompts with layered constraints that prevent common failure modes. Instead of a single instruction like "write clearly," a Claude-engineered prompt might include: "Use sentences under 20 words. Avoid passive voice. Do not use the words 'leverage,' 'synergy,' or 'innovative.' Start with a verb." Adding the reason behind each rule makes it more robust, since Claude generalizes from the explanation [2].

Conclusion

The model behind your agent will keep changing; the quality of the brief is what you control. Teams that treat prompts as tested, versioned assets, drafted with Claude, measured against clear criteria and refined after every miss, get more from every AI tool they use. Claude can do much of the engineering, but people still define what good looks like, judge the output and own the result. That is the Human-Led, AI-Augmented approach.

References

1. McKinsey & Company, Tinkoff, D., Van der Veken, L., Chui, M., and Balakrishnan, T. (August 25, 2026). The state of AI in 2026: On the road to ROI. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
2. Anthropic. (Accessed October 2026). Prompting best practices (Claude Platform documentation). https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
3. Anthropic. (Accessed October 2026). Prompt engineering overview (Claude Platform documentation). https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview
4. Anthropic. (October 14, 2024). Improve your prompts in the developer console. https://claude.com/resources/articles/prompt-improver

Published by Tiger Tracks. Eye of the Tiger Intelligence Series.