Trough of Disillusionment

Claude Code is a language model that predicts the most likely next words based on patterns. When it lacks specific information, it will generate plausible-sounding content rather than saying 'I don't know.' Understanding this tendency is the first step to preventing it from causing real damage in your business.

No. Fabrication is an inherent characteristic of how large language models work, not a software bug. Future models will improve, but the tendency will never fully disappear. That is why this programme teaches you to build systematic guardrails rather than hoping the problem goes away.

Watch for overly confident claims about specific numbers, dates, or facts that you did not provide. If Claude Code cites a statistic, URL, or data point that was not in your input files, treat it as suspect until verified. This module teaches you the specific patterns to watch for.

You can trust Claude Code for tasks where fabrication is obvious or harmless, such as drafting copy you will review, reformatting data you provided, or generating code you will test. The risk is highest when it presents facts, figures, or references. This module teaches you where to trust and where to verify.

An effective specification tells Claude Code exactly what to produce, what constraints to follow, what format to use, and what to avoid. Vague specifications produce vague results. The more precise your spec, the more predictable and reliable the output. We teach a structured template that covers all these elements.

Detailed enough that a competent stranger could follow them without asking questions. If your specification requires Claude Code to guess your intent, it will guess wrong at least some of the time. This module shows you how to find the right level of detail without over-engineering.

Yes. Well-written specifications become templates you adapt for new projects. The initial investment in writing a thorough spec pays off every time you reuse it. Most businesses build a library of specification templates within a few weeks of completing this module.

A specification defines what needs to be built and the rules it must follow. A skill defines how Claude Code should execute a repeatable task. Specifications are project-specific documents; skills are reusable tools. Often a specification references multiple skills during execution.

Separate the stable parts from the variable parts. Put fixed rules and formats in the specification and pass changing data as inputs. For example, a report specification defines the structure and brand rules permanently, while the actual data changes each month. This module covers parameterisation patterns in detail.

Without explicit brand rules in your CLAUDE.md, Claude Code defaults to generic styling: blue headers, standard fonts, and American spelling. It has no way to know your brand unless you tell it. One client received a 'professional blue' report when their brand explicitly bans blue. This module prevents that.

Extremely specific. Include exact hex colour codes, font names, heading hierarchy, table styling rules, and spelling conventions. Do not write 'use our brand orange' when you should write 'H1 colour: #d35a00.' Claude Code follows precise instructions well but interprets vague ones unpredictably.

This module helps you create them. Even if you have never documented your brand formally, you have preferences: colours on your website, fonts in your documents, the tone you use with clients. We walk you through extracting these into a format Claude Code can follow consistently.

Only if they are in your CLAUDE.md or a file that CLAUDE.md references. Claude Code reads these files at the start of every conversation. If your brand guidelines are in a separate brand-guidelines.md file referenced from CLAUDE.md, they will be loaded automatically every time.

Claude Code processes data through AI summarisation layers that can misread tables, swap columns, invert values, and present wrong numbers with complete confidence. We have seen it swap customer service hours with shop opening hours on a real client project. Every data point in client-facing work must be verified independently.

Never verify data using the same method that produced it. If Claude Code extracted a number from a webpage, do not ask Claude Code to check that same webpage again. Take a screenshot and read it with your own eyes, or cross-reference against the original source directly. This module gives you a complete verification checklist.

It takes far less time than fixing incorrect data after it reaches a client. A quick visual check of key figures takes two to three minutes. Rebuilding client trust after presenting wrong data takes months. The protocols in this module are designed to be fast and systematic, not burdensome.

For client-facing documents, yes. For internal working documents, verify the data points that drive decisions. The module teaches you a risk-based approach: high-stakes outputs get full verification, internal drafts get spot checks. But anything a client will see gets verified completely.

Flag it immediately with an '[UNVERIFIED]' tag, go back to the original source using a different method such as a screenshot or direct database query, and correct the record. Then update your specifications to prevent the same class of error recurring. This module includes a full incident response workflow.

It is a set of instructions you embed in your CLAUDE.md that explicitly tells Claude Code to never make up content, data, URLs, statistics, or quotes. Instead of fabricating, Claude Code should flag gaps, ask for sources, or mark content as unverified. It transforms the default behaviour from 'guess confidently' to 'admit uncertainty.'

It significantly reduces fabrication but does not eliminate it entirely. Claude Code follows CLAUDE.md instructions with high compliance, but no instruction is 100% effective against a language model's tendency to generate plausible content. The rule works best when combined with the verification protocols from Module 4.

This module provides the exact wording you need, tested across hundreds of real business tasks. You add a 'Critical Overrides' section to your CLAUDE.md with explicit rules about content extraction, source verification, and what to do when information is missing. We provide the template; you customise it for your business.

It makes Claude Code marginally more cautious, which is exactly what you want for business use. You may see more responses like 'I cannot verify this figure' instead of a confident wrong answer. That honesty is far more valuable than speed when your professional reputation is on the line.

Scepticism after poor outputs is healthy and expected. This module teaches you to acknowledge the failures openly, show what you have learned about preventing them, and demonstrate improved results on low-risk tasks first. Trying to dismiss valid concerns will make the resistance worse.

Not necessarily, and not all at once. Start with one or two early adopters who are genuinely interested, let them build skills and confidence, then expand. Forcing adoption on reluctant team members before the systems are robust enough creates exactly the failures that fuel scepticism.

Understand their concerns first. Often resistance comes from legitimate worries about job security, quality standards, or past bad experiences. Address those concerns directly. Some team members may be better suited to verification and quality assurance roles rather than direct AI interaction. Not everyone needs to prompt; everyone needs to understand.

Show them the fabrication examples from this phase. Decision-makers who understand the real risks of premature deployment become allies for doing it properly. Frame your careful approach as risk management, not resistance. This module includes a stakeholder communication template specifically for this conversation.

For most businesses, three to six weeks from the first significant AI failure to having robust enough systems to trust again. The businesses that get stuck are those that either give up entirely or refuse to acknowledge the problems. This module gives you a structured path through the doubt so you come out stronger.