Why most prompts fail — and the Lean parallel
In Lean manufacturing, waste is defined as anything that doesn't add value. Motion without output. Steps that don't move the product forward. Rework because the specification wasn't clear at the start.
Poor prompting is the same problem. Vague inputs create rework. You get something back from Claude, it's not quite right, you describe what you wanted, Claude tries again, still off — four exchanges later you have something you could have gotten in one if you'd briefed correctly at the start.
The discipline of good prompting is the same discipline as good operations: define the spec before you start the work. That's all this module is. And like most operational disciplines, it feels like extra effort the first few times and then becomes reflexive.
Claude produces work that is exactly as scoped as your prompt. A vague prompt isn't Claude's failure — it's an underspecified order. You wouldn't hand a contractor a two-word work order and expect a finished job. Don't do it with Claude either.
The three failure modes
Most bad outputs trace back to one of three causes:
- Missing role. Claude doesn't know what lens to apply. "Write a summary" versus "Write a summary for a plant manager who needs to present to the board next week" are completely different tasks.
- Missing context. Claude doesn't know your operation, your client, your standards, or your constraints. It fills the gaps with generic assumptions — which is why you get generic output.
- Missing standard. Claude doesn't know what "done" looks like. Without a defined output standard, it'll produce something reasonable by its own measure — which may not match yours.
Operators know this pattern: when you skip the specification step, you pay for it in rework. The same is true here. Spending 3 minutes writing a tight prompt saves 20 minutes of back-and-forth corrections. The math always works out.
The RCTS framework — your prompt structure
Every prompt you write from here forward should have four elements. Not all four need to be long — some tasks are simple enough that a sentence handles one of them. But all four should be present in your thinking before you hit send.
RCTS is a mental checklist, not a rigid template. A tight, well-written paragraph can cover all four elements. What matters is that before you send any prompt, you can answer: Does Claude know who it is? Does it know the situation? Does it know the deliverable? Does it know the standard? If all four are yes, send it.
Building your system prompt
Your system prompt is the standing briefing you give Claude in every conversation within a Project. It handles the parts of RCTS that don't change — your role, your operation type, your standards — so you don't have to repeat them every time.
Think of it as the first page of a briefing binder you'd hand any new contractor: here's who we are, here's how we work, here's what we expect. Once it's in the system prompt, Claude reads it before every conversation. You stop repeating yourself.
What belongs in a system prompt
- Who you are and what kind of work you do
- Your operation type, industry, and scale
- The audience for your deliverables (clients, leadership, field teams)
- Your quality standard — what "professional" means in your context
- Any standing constraints: length, format, tone, language
- What Claude should always do (e.g., "flag any assumptions you're making")
- What Claude should never do (e.g., "don't use corporate jargon or filler phrases")
A sample system prompt — adapt this to your practice
WHO I AM: I'm an operations consultant working with manufacturing, distribution, and service businesses. I build deliverables for clients — Excel trackers, SOPs, leadership briefs, forecasts, and work order reports. My clients are plant managers, GMs, and business owners. YOUR ROLE: You are my working partner on client deliverables. You are a skilled operations practitioner — not a generalist AI assistant. Think and write like an experienced operations professional, not like a tech writer or a chatbot. MY QUALITY STANDARD: Every output must be production-ready or close to it. That means: - Clear, direct language. No filler phrases like "certainly," "great question," or "as an AI." - Correct professional formatting for the document type. - Specific and operational — not vague or theoretical. - Appropriate for a client to see without me rewriting it. STANDING INSTRUCTIONS: - If you make assumptions to fill in missing context, flag them explicitly so I can correct them. - If a task is ambiguous, ask one clarifying question before proceeding — don't guess. - When producing documents, use the format appropriate to the deliverable type. SOPs get numbered steps. Briefs get headers. Trackers get structured tables. - For bilingual requests, produce English first, then Spanish immediately after. - Never pad output. If the task is a 200-word brief, write 200 words of substance — not 300 words with filler. WHAT I'M BUILDING: [Update this section each time you start a new project or client engagement — describe the specific operation, client context, and what you're producing for them.]
The [WHAT I'M BUILDING] section at the bottom is the one you update project to project or even session to session. Everything above it stays constant. This is the power of the system prompt: your standing standards are set once, and you only have to specify what changes.
How to install your system prompt
- Go to claude.ai → Projects and open your Project
- Click Edit Project or the settings gear
- Paste your system prompt into the Project Instructions field
- Save — it now applies to every conversation in that Project
Context — what to include and how to structure it
Context is the most underused element in most prompts, and the most important. Claude is only as useful as the information it has to work with. Giving it the right context is the difference between a useful first draft and a generic one you spend an hour fixing.
Four types of context that change everything
| Type | What it is | Example |
|---|---|---|
| Operational context | The specifics of your operation — size, type, processes, vocabulary | "3-shift plastics plant, 120 employees, 4 production lines, running SAP for maintenance" |
| Audience context | Who will read or use the output — their role, knowledge level, priorities | "This brief is for the COO — she wants numbers, risks, and recommended actions. No background she already knows." |
| Document context | Existing materials Claude can reference or build on — SOPs, data, prior reports | Upload the existing SOP and say "revise this to reflect our new changeover sequence" |
| Constraint context | Limits and requirements — length, format, non-negotiables | "One page max. No jargon — this is for shop-floor employees. Must include a Spanish version." |
Uploading documents as context
One of the highest-leverage moves in Claude is uploading an existing document and asking Claude to work from it. Instead of describing your operation from scratch, hand Claude your current SOP, your tracker template, your client brief — and let it build on what already exists.
- Upload files using the paperclip icon in any conversation
- Tell Claude explicitly what to do with the document: "Use this as the baseline. Revise Section 3 to reflect the new procedure."
- You can upload multiple files — data files, templates, reference documents — in a single session
- In Projects, files uploaded to the Project are available in every conversation automatically
More context is almost always better than less. Don't worry about "overwhelming" Claude with background — the models handle long context well. The risk is almost always too little context, not too much. When in doubt, include it.
Examples as guardrails — showing Claude what good looks like
The single most powerful prompting technique most people never use: give Claude an example of the output you want.
Words like "professional," "concise," and "clear" are interpreted differently by everyone. An example is unambiguous. If you have a past SOP that's exactly the format you want, upload it. If you have a brief that hit the right tone, paste a paragraph in. Examples function as guardrails — they constrain Claude's interpretation and pull the output toward your actual standard.
Write a leadership brief on last month's production performance. Keep it professional and concise.
Write a leadership brief on last month's production performance. Match this structure and tone exactly: "PRODUCTION SUMMARY — MARCH 2026 Overall output: 94% of target. Assembly A over by 4%, Packaging line under by 6%. Key risk: Packaging line downtime trending up — 3rd consecutive month. Root cause not yet isolated. Action required: Maintenance RCA due April 18." One page max. Numbers first, then risk, then action. No narrative padding.
You're not giving Claude a finished template to copy — you're giving it a calibration signal. Even a rough example that captures the right length, tone, and structure will dramatically improve the output. Tell Claude "match this format" and show it something close to what you want.
Negative examples work too
You can also tell Claude what you don't want — especially useful when you need to avoid a specific output style:
Write the SOP for our receiving inspection process. Do NOT produce this kind of output: "This Standard Operating Procedure is designed to ensure that all receiving inspection activities are conducted in a consistent and systematic manner, leveraging best practices from industry standards to optimize quality outcomes..." DO produce this kind of output: "1. Verify shipment count against PO. Log any discrepancy in RecLog before moving product. 2. Inspect outer packaging for damage. Photograph and quarantine any damaged units. 3. Pull 5% sample for dimensional check. Record results on Form RX-7." Direct. Numbered. Actionable. Written for someone who already knows the job.
Iterating without restarting
When Claude's first output isn't quite right, most people do one of two things: accept it anyway, or start over with a new prompt. Both are wrong. The right move is to refine in the same conversation — and to be specific about what needs to change.
Claude holds the full context of the conversation. It remembers what it produced and can adjust precisely. Think of it like directing a revision: not "do this again but better" — "change this specific thing, keep everything else."
High-leverage refinement instructions
-
1
Scope corrections
"Cut Section 2 entirely — the client already knows that background. Move straight from the performance summary to the risk items."
-
2
Tone corrections
"The tone is too formal for this audience. Rewrite it the way I'd talk to a plant manager I've worked with for three years — direct, no ceremony."
-
3
Format corrections
"Restructure this as a numbered checklist, not paragraphs. Each item should be one sentence — action + responsible party."
-
4
Fact injections
"The actual downtime figure was 14.2%, not 12%. Update that number and revise the risk assessment to reflect it."
-
5
Expansion or compression
"The risk section is too thin. Expand it to cover the top three risks with a severity rating and owner for each. Everything else stays as-is."
If Claude has fundamentally misunderstood the task — wrong document type, wrong audience, wrong framing entirely — it's faster to start over with a corrected prompt than to iterate toward the right answer from a wrong baseline. The test: are you adjusting details, or correcting direction? Adjusting: iterate. Correcting direction: restart with a better prompt.
Your prompt library — the compounding asset
The prompts you write for your most common deliverables are reusable. Once you've written a tight prompt for an SOP, a production report, or a leadership brief — and it produced good output — save it. That's your template for the next time. A prompt library compounds: every good prompt you save means the next version of that deliverable takes less time.
Three starter prompts to build from
ROLE: Operations consultant preparing a status update for a manufacturing client's leadership team. CONTEXT: [Paste current week's production data or describe the key numbers here] Client: [Name]. Operation: [Type]. Reporting period: [Week of X]. Audience: Plant GM and ownership group. They want facts, variances, risks, and actions — not narrative. TASK: Write a one-page production status brief covering: 1. Output vs. target (by line if applicable) 2. Key variances and probable causes 3. Top 1–3 risks requiring attention 4. Recommended actions with owners and due dates STANDARD: One page. Numbers-first format. Bullet points under each heading. No filler language. Professional enough to send directly to the client.
ROLE: Operations documentation specialist writing for a [manufacturing / service / field] operation. CONTEXT: [Describe the process: what it is, who performs it, how often, what equipment or materials are involved, any safety requirements, common errors or failure modes] TASK: Write a Standard Operating Procedure for [process name]. Include: - Purpose and scope (2–3 sentences) - Required materials, tools, or PPE - Step-by-step procedure with responsible party for each step - Quality checkpoints or inspection criteria - Common errors and how to prevent them - Sign-off section STANDARD: Produce the English version first, then the Spanish version immediately after. Use numbered steps. Language must be plain — written for someone who knows the job but is reading this for the first time. No corporate filler.
ROLE: Senior operations advisor writing a memo to a client's executive leadership. CONTEXT: [Describe the situation: what happened, what decision needs to be made, what the options are, any relevant data or background] Audience: [Title(s) of recipients]. Their priorities: [what they care about]. TASK: Write an executive memo covering: - Situation summary (3–4 sentences) - Key data or findings - Recommended course of action with rationale - Risks of the recommendation - Requested decision or next step STANDARD: One page. Direct. Written for a decision-maker who has 5 minutes, not 30. Lead with the recommendation, not the background. Professional letterhead-quality language.
These are starting points. After you use each one, refine it based on what worked and what didn't. Save yours somewhere you'll actually use it — a Project document, a Notes file, a shared drive folder.
Your System Prompt + Starter Prompt Library
This module has two deliverables. Both are things you'll use immediately and keep updating throughout the class.
Deliverable 1 — Your system prompt
Using the sample from Section 3 as a base, write a system prompt that reflects your actual practice. Fill in: WHO I AM: Your real role, operation type, and what kind of deliverables you produce. YOUR ROLE: How you want Claude to show up — the expertise and voice it should bring. MY QUALITY STANDARD: What "production-ready" means for your work. Who sees it. What level of polish. STANDING INSTRUCTIONS: What Claude should always do. What it should never do. WHAT I'M BUILDING: Your current engagement or project context. Install this in your Claude Project today. You'll refine it as you go — start with something real, not something perfect.
Deliverable 2 — Your first prompt template
What do you produce most often for clients or your team? An SOP? A weekly report? A brief? A tracker? Pick one.
Role, Context, Task, Standard — written for that specific deliverable type, with your real operation as the context. Use the library examples as a starting point.
Use the prompt on a real task. Iterate once based on what you'd change. Then save the refined prompt — that's the beginning of your library.
One sentence: what would you add to this prompt to make the next version even tighter? Write it in the prompt file. That's how the library compounds.