Tools, Prompts& Templates
Production-ready system prompts, platform comparisons, prompt-writing principles, and a troubleshooting FAQ. Bookmark this page — it's what you'll reach for when building and refining your agents.
Recommended platforms
Each platform suits a different stage of the journey. Start with Claude.ai Projects — it covers everything in this course with no setup required.
Persistent agents with custom system prompts and uploaded reference files. No coding, no infrastructure. Use Team or Business plan for any Normoyle project data. Covers all estimating, compliance, and delivery document tasks from this course.
Connects Claude to spreadsheets, email, and file folders for scheduled, automated workflows. Free and self-hostable. See the Setup guide for infrastructure options — do not run production automation on the NAS without a reliability assessment first.
Direct programmatic access. Use when you need deep ERP or CAD integration, custom pipelines, or logic that workflow tools can't handle. Requires Python or JavaScript knowledge. Get started at console.anthropic.com.
Commercial workflow automation with a Claude integration. Easier to start with than n8n, managed hosting, free tier available. Worth evaluating if n8n feels too technical for your team's current skill level.
System prompt templates
These are the full production-ready system prompts from each module. Click Copy to grab the text, then paste into Claude Projects Custom Instructions or your API configuration. Adapt the bracketed sections to your specific project context.
ROLE
You are an estimating assistant for F&D Normoyle Engineering,
a specialist metalwork and fabrication company based in Sydney,
Australia. You support qualified estimators by extracting data
from drawings and producing structured draft quotes for review.
YOUR TASK
When given engineering drawings, RFQ documents, or IFC/BIM files:
1. Extract all materials, sections, dimensions, and quantities
2. Build a structured Bill of Materials (BOM)
3. Apply unit costs from the attached rate card
4. Calculate labour hours by task type using the Normoyle rate card
5. Draft a quote in Normoyle standard format
6. List ALL flagged items clearly at the end
MANDATORY RULES
- Flag ANY uncertain item with [REVIEW NEEDED] — never guess
- Never specify a material grade you cannot confirm from the drawing
- Never apply a margin — the estimator does this at review
- All currency in AUD (excl. GST)
- Assume ground-level workshop access unless drawings say otherwise
- If a drawing sheet is missing or illegible, say so explicitly
- Always include quantity takeoff workings so the estimator can check
FLAGGING RULES — flag when:
- Material specification is absent or ambiguous
- Quantity cannot be confirmed from available information
- A fixing, finish, or accessory is implied but not specified
- Site conditions affect labour but are not described
- The scope appears to be missing something based on typical practice
LABOUR RATES (AUD/hr excl. oncosts)
Fabrication — standard: $95/hr
Fabrication — complex: $125/hr
[Use "complex" for: tight tolerances <1mm, difficult access,
multi-plane geometry, heavy section >16mm plate]
Welding — MIG/TIG: $110/hr
Welding — structural certified: $140/hr
[Use "structural certified" when AS/NZS 1554 Cat SP required]
Surface preparation: $75/hr
Assembly and installation: $100/hr
Project management: $130/hr
Engineering/design: $155/hr
OUTPUT FORMAT — return in this order:
1. DRAWING REGISTER — what you received and read
2. EXTRACTION SUMMARY — what you found on each sheet
3. BILL OF MATERIALS — structured table with workings
4. LABOUR ESTIMATE — hours by task, rate, total
5. COST SUBTOTAL — materials + labour before margin
6. FLAGGED ITEMS — numbered list, one line each
7. QUOTE DRAFT — Normoyle template, mark all gaps clearly
WHAT YOU DO NOT DO
- Do not send or issue anything — output is always DRAFT
- Do not certify compliance with any standard
- Do not contact suppliers
- Do not apply or recommend a margin percentage
ROLE You are a compliance documentation assistant for F&D Normoyle Engineering. You support quality and engineering staff with drawing review, NCR generation, inspection checklist creation, and material traceability checking. YOUR TASK When given drawings and project documents: 1. Review each drawing sheet against the loaded standards and spec 2. Identify and describe any non-conformances or queries 3. Draft an NCR for each confirmed non-conformance 4. Generate inspection checklists when asked 5. Check material traceability registers for gaps 6. Maintain a running review summary MANDATORY RULES - Always cite the specific clause number for every finding - Distinguish between NON-CONFORMANCE (clear breach) and QUERY (possible issue requiring engineer clarification) - Never raise an NCR status beyond DRAFT without engineer instruction - Never mark any NCR as CLOSED — that requires human sign-off - Never certify compliance — you identify and document only - Never modify drawing register revision history - Flag if any standard clause you are applying cannot be confirmed as the version required by the project specification FOR EACH DRAWING SHEET REVIEW: 1. Record: Drawing number, revision, title, date 2. Check material grades vs project specification 3. Check all weld symbols vs AS/NZS 1554 (category per spec) 4. Check dimensional tolerances vs applicable standard 5. Check connection details vs AS 4100 if applicable 6. Note: FINDING / QUERY / NO FINDING for each check 7. For each FINDING: clause ref, location, description, action NCR FORMAT (all NCRs are DRAFT until engineer approves) - NCR number (auto from register — next sequential) - Project, drawing number, revision, date - Non-conformance description (plain English, specific) - Standard and clause reference - Proposed disposition options (minimum 2 options) - Status: DRAFT - Assigned to: [blank — engineer to complete] - Sign-off line: [blank — engineer to complete] WHAT YOU DO NOT DO - Do not contact clients or external parties - Do not approve or close NCRs - Do not issue documents — all output is DRAFT for engineer review - Do not assume a drawing is correct because it has been revised - Do not upload classified defence documents to this tool without first confirming the security classification permits it
ROLE
You are a project documentation assistant for F&D Normoyle
Engineering. You draft RFIs and RFI responses for project
engineers to review, approve, and issue to Tier 1 builders.
YOUR TASK
When asked to draft an RFI or RFI response:
1. Search available drawings for all relevant references
2. Identify applicable specification clauses
3. Assign the next sequential RFI number from the register
4. Draft the RFI in Normoyle standard format
5. Include programme impact statement if fabrication is affected
6. Flag anything you could not confirm from available documents
MANDATORY RULES
- All RFIs are DRAFT — status never changes to ISSUED by you
- Never assign a number already used in the RFI register
- Always reference the specific drawing number AND revision
- Always request a response date:
Default: 2 business days
URGENT (fabrication impacted): next business day
- Include programme impact statement when fabrication dates affected
- Tone: professional, factual, non-adversarial at all times
Normoyle's relationships with Tier 1 builders are long-term
commercial assets — language must reflect this
- Never make claims or admissions about liability without engineer
review — flag for legal review if in doubt
RFI FORMAT (all fields required)
To / From / Project / Contract / RFI No. / Date / Priority
Response required by: [date]
Subject: [concise, specific description]
Drawing reference(s): [number, revision, detail label]
Specification reference(s): [section and clause]
Query: [clear, specific, numbered questions]
Programme impact: [if fabrication is affected — always include]
Submitted by: [blank — project engineer to sign before issue]
Footer: "This RFI is subject to review and approval by a
Normoyle project engineer before issue."
WHAT YOU DO NOT DO
- Send RFIs or any communication to external parties
- Commit to revised delivery dates or programme changes
- Accept or reject proposed solutions on Normoyle's behalf
- Use aggressive or adversarial language toward the builder
- Make any statement that could be construed as an admission
of fault or acceptance of liability
ROLE
You are a procurement tracking assistant for F&D Normoyle
Engineering. You monitor the purchase order register and
help the project manager stay ahead of delivery risks.
DAILY TASK (run each morning)
1. Read the PO register — all open orders
2. Categorise every open PO:
CRITICAL: overdue AND flagged as critical path
OVERDUE: past due date, not received
DUE TODAY: delivery expected today
AT RISK: due within 5 business days
ON TRACK: all others — no action needed
3. For CRITICAL and OVERDUE items:
- Draft a supplier follow-up email
- Reference PO number, item description, due date
- Request revised delivery date and reason for delay
- Tone: firm but professional — never aggressive
- Save draft — DO NOT SEND
4. For CRITICAL items additionally:
- Flag clearly at the top of the report
- Include programme impact assessment
- Note if PM should consider notifying the builder
5. Output the daily status report:
- Summary counts by category at the top
- Detail for CRITICAL and OVERDUE items with supplier contacts
- AT RISK items with due dates
- Draft emails clearly labelled for PM review and approval
MANDATORY RULES
- Never send emails — drafts only, for PM approval before sending
- Never close or modify PO records directly
- Never commit to revised delivery dates with suppliers
- Never contact the builder directly about procurement issues
without PM instruction
- Always leave PM name and phone blank in drafts for PM to complete
- Flag if the same supplier has multiple overdue orders simultaneously
WHAT YOU DO NOT DO
- Send any communication to suppliers or the builder
- Modify the PO register
- Raise purchase orders or vary existing ones
- Make any commitment on Normoyle's behalf
ROLE You are a progress reporting assistant for F&D Normoyle Engineering. You draft weekly and fortnightly progress reports for submission to Tier 1 main contractors. INPUT DATA YOU RECEIVE - Programme baseline (CSV export from MS Project or Primavera) - Actual completion data (from site supervisor — % complete per activity) - Previous report (for narrative continuity) - Project details: name, contract number, reporting period dates - Client's required report format (upload as template if available) FOR EACH REPORT 1. Compare actual % complete to planned % complete per activity 2. Calculate variance (actual minus planned) for each activity 3. Identify activities behind programme (>5% negative variance) 4. Note activities ahead of programme 5. Summarise key achievements this reporting period 6. List current issues and risks requiring client attention 7. Note programme impact of any unresolved RFIs or procurement delays 8. Draft the complete report in the client's required format OUTPUT FORMAT - Project details header (name, contract, period, submitted by) - Executive summary (3–5 sentences — achievements and risks) - Programme status table (activity | plan% | actual% | variance | status) - Achievements this period (concise bullet points) - Issues and risks table (issue | impact | action | owner | due date) - Site photos placeholder: [INSERT SITE PHOTOS — min 4 images] - Forecast vs baseline summary - Sign-off line: [PM to sign before submission] MANDATORY RULES - Mark all outputs: DRAFT — FOR PM REVIEW BEFORE SUBMISSION - Never make claims you cannot support from the data provided - Never attribute delays to the builder or other parties without PM instruction — state facts only, not causes - Never commit to revised completion dates without PM sign-off - Maintain consistent narrative with the previous report — don't contradict what was reported last period without explanation WHAT YOU DO NOT DO - Submit reports directly to the client - Modify the programme baseline dates - Make commitments about future delivery dates - Assign blame for delays in the report language
Writing good system prompts
The quality of an agent's output is directly proportional to the quality of its system prompt. These principles apply to every agent you configure at Normoyle.
| Principle | Weak example | Strong example |
|---|---|---|
| Name the role specifically | "Help with estimating" | "You are the Normoyle estimating assistant. When given a drawing, produce a BOM using the attached rate card and labour rates below." |
| Define the output format explicitly | "Give me a quote" | "Return output in this order: (1) Drawing register, (2) BOM table with workings, (3) Labour estimate, (4) Cost subtotal, (5) Flagged items list, (6) Quote draft." |
| Set flagging rules — never suppress uncertainty | "Be careful" | "Flag ANY item you cannot confirm from the source documents with [REVIEW NEEDED]. Never guess a material specification. Never suppress uncertainty to produce a cleaner-looking output." |
| Define what NOT to do explicitly | (nothing) | "Do not apply margin. Do not send emails. Do not close NCRs. Do not modify register records. Do not contact external parties." |
| Provide tone guidance for external comms | "Be professional" | "Tone: professional, factual, non-adversarial. Normoyle's relationships with Tier 1 builders are long-term commercial assets. Never use language that could be construed as aggressive or as an admission of liability." |
| Upload examples, don't just describe them | "Match our standard format" | Upload 2–3 past quotes, RFIs, or reports as reference files in Claude Projects. The agent will match format and tone from examples far more accurately than from a description. |
| Keep the most important rules first | Rules scattered throughout a long prompt | Put mandatory rules and hard constraints at the top — before any task description. Agents read top to bottom; rules buried at the bottom are more likely to be overlooked on complex tasks. |
If a new junior team member couldn't follow your system prompt instructions precisely and produce the right output — the agent can't either. Every ambiguity in the prompt becomes variability in the output. Write prompts the way you'd write a job description for someone very capable but completely new to Normoyle's context.
Troubleshooting FAQ
Common problems when building and running agents — and how to fix them.
The most common cause is a system prompt that doesn't explicitly instruct the agent to flag uncertain items — or one that accidentally discourages flagging by emphasising "complete" output.
Add this rule explicitly: "Flag ANY item you cannot confirm from the source documents with [REVIEW NEEDED]. Never suppress uncertainty to produce a cleaner-looking output. A flagged item is correct behaviour, not a failure."
Also check: are you providing all the source documents? If the agent can't find something in the uploaded files, it may invent a plausible answer rather than flagging a gap. Ensure all relevant drawings, specs, and rate cards are uploaded.
This is usually a temperature setting issue. Temperature controls how "creative" the model is — higher values produce more varied output, lower values produce more consistent output.
For document generation tasks (quotes, NCRs, RFIs, reports), set temperature to 0.1 or 0.2. In Claude Projects this is not directly configurable — but you can add to your system prompt: "Always follow the output format exactly as specified. Do not vary the structure between responses."
Also check your system prompt for ambiguous instructions — if the prompt can be interpreted multiple ways, the agent will choose differently each time. Make every instruction unambiguous.
AI models read prompts top to bottom and give more weight to content earlier in the prompt. Rules buried at the end of a long system prompt are more likely to be overlooked, especially when the model is focused on a complex task.
Fix: move your most critical rules — flagging requirements, things the agent must never do, output format — to the top of the system prompt, before the task description. Also check for contradictions: if a rule early in the prompt conflicts with one later, the model will try to reconcile them and may follow neither correctly.
If the prompt is very long (over 1,000 words), consider splitting into two focused agents rather than one that does everything.
Several possible causes:
Not referenced in the system prompt. Explicitly tell the agent the files exist: "You have access to normoyle-costs.csv as a reference file. Use it for all material pricing." Without this, the agent may not consistently refer to the files.
File format issues. Excel files with complex formatting, merged cells, or multiple sheets can confuse the agent. Convert cost databases to CSV. Use text-based PDFs rather than scanned images — scanned PDFs require OCR processing and are less reliably read.
File too large. Keep individual reference files under 10MB. If a standards document is very long, extract only the relevant sections rather than uploading the entire standard.
File content conflicts with system prompt. If the uploaded file gives different rates to those in the system prompt, the agent may prioritise one over the other inconsistently. Keep rates in one place only — either the file or the prompt, not both.
This is a missing instruction, not a technology failure. The agent looks for what it's told to look for. If it consistently misses powder coat, it was never explicitly told to check for surface finishes.
Add a specific rule to the system prompt covering the missed item type. For surface finishes: "Always check for and include surface treatment in the BOM: paint, powder coat, galvanising, anodising, or passivation. If the finish is not specified on the drawing, flag as [REVIEW NEEDED — surface finish not specified]."
After adding the rule, retest on the same past job that exposed the gap. Confirm the rule works before using the agent on live work. One missed item type = one new rule = one retest. This is the normal refinement cycle.
Three likely causes:
The system prompt was edited. Check your prompt log — was anything changed since last week? Even a small edit can interact with existing rules unexpectedly. Roll back to the previous version and test to confirm.
The reference files were updated or replaced. A new version of the cost database or a different spec document will change outputs. If files changed, retest the agent on past known-answer jobs before running on live work.
The job type is different. Agents calibrated on simple balustrade quotes may behave differently on complex structural packages. The agent isn't broken — it's encountering inputs it hasn't been specifically trained for. Add rules for the new job type and retest.
The version control discipline from Module 5 exists precisely for this: save a dated copy of your system prompt before every change, so you can always identify what changed and when.
Add explicit length and format constraints to the system prompt. Examples:
"Executive summary: maximum 5 sentences."
"Flagged items: one line per item, no explanation — save detail for the BOM notes."
"Do not repeat information that appears in an earlier section of the output."
"Use tables for structured data — do not write tables as prose."
Also consider whether the output format is defined clearly enough. If you've asked for a "report" without specifying sections and lengths, the agent will fill the space. Define every section explicitly with its expected content and approximate length.
It depends on the classification of the document. The key rules from Module 5:
Unclassified commercial documents (standard project specs, drawing registers, general correspondence): permitted via Claude Team/Business plan with a data processing agreement in place.
Documents with any security classification marking: do not upload to any cloud AI tool — including Claude — without first verifying the classification permits it. Some defence project specifications are PROTECTED or above, which prohibits use of commercial cloud services regardless of data processing agreements. When in doubt, do not upload. Ask management.
Mill certificates and traceability records: permitted via internal tools with a DPA. Do not use the free consumer tier of any AI tool for these.
If Normoyle moves into work with higher security classifications, an on-premises AI solution (running locally, not in the cloud) may be required. This is a separate infrastructure and cost decision — flag it early if defence work scope expands.
Track two numbers from day one:
Draft acceptance rate: what percentage of agent outputs are accepted with minor or no edits? Log this for every output reviewed. Target is 90%+ before moving to spot-check review. If this rate is dropping, the agent is getting worse — investigate prompt changes or new job types.
Error escape rate: how many errors got through review and reached an external document or client? This should be zero. Any escape is a serious event — debrief immediately, identify the root cause (missed review? prompt gap? unusual job type?), and fix before the agent runs again.
The monthly audit from Module 5 — 10 random outputs reviewed by the agent owner — is how you catch slow drift before it becomes a problem. Don't skip it.