LLM Workshops: Practical Training on Claude, GPT, Gemini

In short

A hands-on workshop on the frontier large language models: Anthropic Claude, OpenAI GPT-4-class, Gemini. What they can and can’t do. How to prompt them effectively. How to deploy them in your business. Hosted by Gary Cheers, who works with these models every day.

What you get

• LLM capability overview (and the honest limits)
• Prompting craft, the patterns that actually work
• Comparison: Claude vs GPT vs Gemini for SME use cases
• Deployment patterns: when to use API vs ChatGPT vs Copilot
• Safety + governance for LLM use in regulated industries

Format

Half-day or full-day cohorts. POA. 8 to 12 attendees. Regional in-person or UK-wide remote.

Who this is for

The typical attendee is somewhere on a spectrum between “has used ChatGPT a handful of times and is unsure how to push further” and “uses it daily but suspects they are leaving half the value on the table”. The workshop is calibrated to that range. People who have never opened a chat interface are accommodated; people deeply technical get separately scheduled into a more advanced session.

The mix that lands best in a single cohort is a function or department team, not a randomly assorted cross-business group. Marketing teams, sales teams, support teams, operations teams. They share the same source material, the same workflows and the same constraints. The workshop produces a prompt library at the end that is specific to their work, which is much more valuable than a generic one.

What the day actually looks like

A full-day session breaks into four blocks, with a working lunch separating the morning from the afternoon.

1. Capability foundations. What modern LLMs actually do well. Reading and reasoning over long documents. Drafting and rewriting. Structuring messy information. Code generation. Image and document understanding through multimodal inputs. The honest limits: arithmetic on specific numbers, factual recall on obscure topics, anything where the answer needs a citation the model cannot retrieve.
2. Prompting craft. Six patterns that actually shift output quality: clear role specification, structured task description, few-shot examples drawn from the team’s own work, chain-of-thought when reasoning is required, structured-output mode for downstream use, and citation requirements where accuracy is critical. Attendees build prompts against their real materials, not toy exercises.
3. Tool selection. When ChatGPT, Claude, Gemini, Copilot or a custom build is the right answer. Each has a specific shape. The shape matters more than the brand. The cost shape matters too: an SME paying for both ChatGPT Team and Claude Pro across ten seats is well inside £400 per month, often a tenth of the productivity it enables.
4. Deployment and governance. When to keep work inside the consumer apps. When to move to API. When to add a retrieval layer. What to do about hallucinations in practice. The governance perimeter for regulated industries. The kind of work that should not go to a public model at all.

Attendees leave with three things: a written prompt library for their specific workflows, a tool-selection cheat-sheet for the most common task types, and an action list of three to five concrete experiments to run inside the following fortnight.

The 2026 LLM landscape, briefly

The frontier models in 2026 are clustered tightly enough that the brand choice matters less than the workflow fit. Anthropic Claude (the 4.x family) is the strongest reasoner over long context, particularly on documents and code. OpenAI GPT (the 5.x family) has the broadest plugin ecosystem and the strongest image and voice integration. Google Gemini has the best integration with Google Workspace and a long-context window that suits enterprise document work. Microsoft Copilot sits in the same family as GPT, with Office 365 binding it close to the work the team already does.

Most SMEs end up using two of the four routinely. The workshop covers the trade-offs honestly so the team chooses on capability rather than habit, brand loyalty or whichever vendor’s marketing got there first.

Hallucinations and the accuracy question

The single most common SME concern about LLMs is hallucination. The workshop treats it directly. The honest position in 2026 is that frontier models hallucinate substantially less than the 2024 generation, but they still produce confident wrong answers on specifics, particularly on figures, dates, citations and obscure facts.

The patterns that contain the risk are practical and learnable.

• Ground the model in retrieved documents. Give it the source material rather than asking it to recall.
• Constrain the output shape. Structured-output mode reduces freelancing.
• Require citations. Where the model cannot point at a source, the team treats the output as a draft, not a fact.
• Use the right tool for the task. Calculation goes to a calculator, not the model. Date arithmetic, the same.
• Human review in the loop. For anything that goes external or has consequences, a human still signs off.

Used inside these guardrails, the LLM is fit for production work. Used without them, it is a liability. The workshop spends a meaningful share of the day on this discipline because it is the single largest lever on whether the team comes back to the tool a month later.

Sector adaptations

The standard workshop sits at the general business level. Where the team is in a regulated industry, a sector overlay is added. Law firms get a section on SRA implications, conflict checking, and document review patterns. Accountants get coverage of ICAEW guidance, working papers and audit trail discipline. Healthcare-adjacent and financial services teams get equivalent treatment. The overlay is included at no extra cost when the cohort is from a single sector.

Pricing and packaging

The workshop sits inside the AI Training line on price-on-application terms, with cohort sizes between 8 and 12 attendees. Half-day or full-day shapes available. Remote across the UK or in-person across the North West and Midlands routinely; further afield by arrangement.

Two adjacent options where the workshop is one piece of a larger picture.

• Government-funded AI training. Eligible SMEs can offset 50 percent or more of the workshop cost through Help to Grow: Digital and equivalent regional schemes.
• Fractional CAIO from £3,500 per month. The workshop seeds the team capability; the Fractional CAIO sustains it with ongoing experimentation, governance and prompt-library maintenance.

What attendees leave with

Three concrete artefacts.

1. A prompt library specific to their work. Twenty to thirty prompts they built during the day, calibrated against their own materials, ready to use on Monday morning.
2. A tool selection cheat-sheet. Which model wins on which task, when to use the free tier, when the paid tier earns its keep. Bookmark-able and quick to reference.
3. An experiment list. Three to five concrete things to try in the fortnight after the session. Anchored to their actual workflows. Specific enough that the team can hold each other accountable.

The third artefact is the most important. Workshops that do not produce a written action list rarely change behaviour. Workshops that produce a specific list with named owners and dates do.

What the workshop deliberately avoids

The workshop avoids three things that tend to derail similar sessions.

• Vendor pitches in disguise. No sponsor relationships. No “this product is the answer” framing. The capability is discussed in terms of what it does, not whose logo is on the marketing.
• Coding deep-dives. The audience is business users, not engineers. Code shows up only where it illustrates a point a business user needs to understand. Engineers get a different, more technical session.
• Speculation about AGI or the future. Some sessions like this turn into philosophy. This one stays anchored to what is buildable today and how to get value from it inside the next quarter.

How the day compares to free online resources

A reasonable question. Plenty of free YouTube content covers LLMs in 2026. The honest difference is calibration and bespoke applicability. A free video is broadcast at everyone; the workshop is calibrated to your team, your sector and your specific workflows. The prompt library and experiment list at the end are specific to the work the team actually does, which a generic video cannot produce.

For teams whose budget genuinely does not stretch to the cohort cost, the government-funded AI training routes can offset a meaningful share through Help to Grow: Digital and equivalent regional schemes.

Common questions from cohort attendees

Three questions come up in nearly every cohort, worth flagging here in case they help calibrate expectations before booking.

The first is on data security. Where can the team safely put real client information into an LLM. The answer in 2026 is nuanced: paid tiers of Claude, ChatGPT and Gemini all offer zero-retention modes and enterprise-grade data handling, which makes most ordinary business data acceptable for use. Specifically sensitive categories (health data, legal privilege, financial regulatory data) still warrant additional care including possibly self-hosted models for the most sensitive cases. The session covers the decision tree explicitly.

The second is on cost runaway. Will API costs explode if the team uses LLMs heavily. Short answer: rarely at SME scale. The math is straightforward: a 10-person team using LLMs for 50 prompts per person per day on Claude’s standard tier comes in under £200 per month all in. Even heavy use cases (long-document analysis, bulk classification) tend to land under £1,000 per month unless the volume is genuinely industrial.

The third is on staying current. Models update every few months; how does the team keep up. The honest answer is that the foundations are stable: prompting craft, tool selection logic, hallucination management. The headline model version changes but the underlying discipline does not. The session focuses on the durable skills rather than the model-specific quirks that will be irrelevant inside six months.

The hour after the workshop

The strongest signal that a workshop will translate into adoption is what happens in the hour after it ends. Cohorts that immediately schedule a follow-up working session, block weekly time, and identify which workflows to extend usually see substantial behaviour change. Cohorts that walk out without a plan usually do not.

The Wingenious sessions build the planning into the day. The final 30 minutes are dedicated to writing down what each attendee will do in the next week, the next month and the next quarter. The plan goes home with them, copied to their manager where appropriate.

The 30-day follow-up support catches teething problems and reinforces the early wins. The shared Slack channel keeps the cohort connected to each other and to Wingenious. Office hours at week two and week four catch the questions that emerge once the work moves from theoretical to applied.

The role of Gary Cheers in the session

Sessions are hosted by Gary Cheers, founder of Wingenious. Thirty years in digital, three years deep in AI tooling, based in Wrexham, working with these models daily on live SME engagements. The point matters because the workshop content is calibrated against what is actually live in production for UK SMEs in 2026, not against vendor marketing materials or generic enterprise frameworks.

The session draws on patterns Wingenious sees across audit, feasibility and sprint engagements: which prompts actually work for SME operations teams, which tools genuinely deliver value at SME scale, which integrations recur, where the hallucination risk shows up in practice, and how teams actually adopt new workflows. The content stays grounded because the work generating it is current.

Where a session has technical or sector-specific content beyond the standard scope, additional Wingenious staff or invited specialists join. The session shape is consistent regardless; the depth on specific topics adjusts to the cohort.

What the prompt library at the end actually contains

A typical cohort produces 20 to 30 prompts across the day, covering the main task types the team encounters.

Drafting prompts: emails to specific audience types, internal updates, customer-facing communication, follow-up notes. Each one calibrated against the team’s existing voice and tone.

Summarising prompts: meeting notes, document digests, customer call summaries, weekly performance overviews. Each one structured to extract the specific information the team’s workflow needs.

Classifying prompts: support ticket triage, lead categorisation, sentiment analysis, document type identification. Each one tested against examples from the team’s real data.

Extracting prompts: contract clauses, invoice fields, application form data, structured information from messy inputs. Each one bracketed with the right confidence-checking discipline.

Reasoning prompts: structured analysis of options, decision frameworks applied to specific situations, scenario evaluation. Each one designed to surface the thinking rather than to produce a single answer.

The library is delivered as a Notion or Google Doc that the team can extend after the session. Most teams grow it from 20 prompts on day one to 60 or more within the first quarter.

Related capabilities

AI strategy workshops · AI tools and platforms training · AI leadership training

Related

Sectors where LLM workshops land best: law firms, accountants.