Give the AI only the manual's inspection, lubrication and maintenance chapters, and make it cite a manual page for every line it writes. A competent person then checks each line against its page, trials the sheet on one machine and approves it before anyone relies on it. The AI drafts and organises; it never approves.
Manuals are written for engineers and run to hundreds of pages. The model's real value is cutting that down to the dozen pages an operator needs and putting them in the order they walk round the machine. Its characteristic failures are small and dangerous: an interval attached to the wrong service, a unit quietly converted, a footnote about severe conditions dropped. The method here is built around catching exactly those.
Pick the pages before you upload anything
Start with the right manual, not the nearest one. It must match the machine's model and serial range, because makers can revise intervals and checks between ranges. The operator's manual usually holds the daily checks and the basic maintenance chart; the service manual goes deeper and is mostly for the fitter. Attachments such as a breaker, auger or quick hitch have manuals of their own.
Then cut the document down to the chapters that matter: the pre-operation or walk-round inspection, the maintenance schedule or interval chart, the lubrication and fluid specifications, and the safety decals and warnings. Cutting is not only for focus. It is also a limit issue. Anthropic's help article (updated July 2026) says Claude analyses text and visuals in PDFs of 100 pages or fewer, reads text only from 101 to 1,000 pages, and rejects anything larger; a chat can take up to 20 files of 500 MB each, but files kept in a Project are capped at 30 MB each. A grease-point diagram on page 140 of a 300-page manual is a visual, so a chapter extract of under 100 pages is the safer input. Other assistants have their own limits, so check the one you use.
Keep a small source register alongside, one row per machine, so that every checklist can be traced:
| ID | Machine | Manual and edition | Serial range | Pages used | Attachments |
|---|---|---|---|---|---|
| EX-01, EX-02 | 8 t tracked excavator | Operator's manual, ed. [date] | [range] | 12; 96 to 148 | Quick hitch, bucket set |
| MX-01 | 1.8 t mini excavator | Operator's manual, ed. [date] | [range] | 70 to 112 | Breaker |
| DM-01 | 3 t site dumper | Operator's manual, ed. [date] | [range] | 54 to 90 | None |
| DM-02 | 9 t site dumper | Operator's manual, ed. [date] | [range] | 62 to 104 | None |
| RL-01 | Smooth-drum roller | Operator's manual, ed. [date] | [range] | 48 to 82 | None |
| SS-01 | Skid-steer loader | Operator's manual, ed. [date] | [range] | 88 to 130 | Bucket, forks |
| TH-01 | Telehandler | Operator's manual, ed. [date] | [range] | 102 to 160 | Forks, bucket |
This example fleet of eight machines has seven manuals and therefore seven checklists. That number matters later, when we get to apps. Before relying on a scanned manual, run a quick test: ask the model to quote one interval word for word and compare it with the printed page. If it cannot, the extract is not readable and every line built on it is suspect.
A prompt that forces a page reference on every line
The prompt does three jobs: it restricts the model to your pages, it demands the manual's own words next to every line, and it keeps the model's general knowledge out of a safety document. The "verbatim" column is what lets a competent person check twenty lines in twenty minutes.
You are converting an equipment manual into a pre-start checklist for site operators.
SOURCE: the attached extract of the operator's manual for [make, model], serial range
[range], edition [date]. Use ONLY the attached pages.
TASK: a table with columns: area | check | what a fail looks like | who (operator or fitter) |
manual page | verbatim source text.
RULES
- Every row needs a manual page and the exact source wording. If you cannot find a page,
leave the row out.
- Copy every number (pressures, torques, intervals, levels, capacities) exactly as printed,
in the manual's units. Do not convert.
- Order the rows as an operator walks round the machine, front left first, cab last.
- List the manual's warnings and "do not operate" statements separately, verbatim, with pages.
- List anything that depends on conditions (dust, mud, breaker work, cold) separately,
verbatim, with pages.
- Do not add checks from general knowledge. If you think one is missing, put it at the end
under SUGGESTED, labelled "not from the manual", with no numbers.A first-draft excavator checklist, with invented pages
Here is part of a first draft for a tracked excavator, shown with the verbatim column left out and the stop column you add before review. Every page number is invented and the manual's own figures are left out on purpose; a real sheet carries the real values, copied exactly, in its verbatim column.
| Area | Check | A fail looks like | Who | Page | Stop if it fails? |
|---|---|---|---|---|---|
| Undercarriage | Tracks, rollers, sprockets, track tension | Cuts, missing bolts, tension outside the manual's range | Operator | 112 | Competent person decides |
| Engine bay | Engine oil level | Dipstick reading outside the marks | Operator | 118 | Competent person decides |
| Engine bay | Coolant level | Below the mark on the reservoir | Operator | 119 | Competent person decides |
| Hydraulics | Hydraulic oil level, in the position the manual states | Outside the marks | Operator | 121 | Competent person decides |
| Hydraulics | Hoses and fittings at boom and stick | Wet patches, chafing, weeping joints | Operator | 124 | Competent person decides |
| Attachment | Quick hitch locked; pins and retaining devices in place | Any pin missing or hitch not fully engaged | Operator | 131 | Yes, set by the competent person |
| Cab | Seat belt, mirrors, lights, horn, reversing alarm | Any not working or damaged | Operator | 98 to 101 | Competent person decides |
| Decals | Safety decals present and legible | Missing, painted over, unreadable | Operator | 12 | Competent person decides |
| Lubrication | Daily grease points listed in the manual | Dry joints, old grease, no grease at a listed point | Operator or fitter | 142 | Competent person decides |
The last column is deliberate. Whether a failed check stops the machine is a safety decision, not a drafting one. The AI leaves it blank for the competent person to fill from the manual's warnings and your risk assessment; the quick hitch row shows how a filled one looks. By "competent person" I mean someone with the training and experience to judge the machine and its checks, as your own safety rules define it: in a small firm that may be the lead fitter, an outside plant inspector or your safety adviser.
Reading the draft against the page: what the competent person ticks
For each row, open the cited page, compare the verbatim text with the manual and tick the row with initials and a date. Rows without a tick do not reach the operators' sheet. A machine's checklist should take about twenty minutes this way (time your first one), which is the point of the verbatim column. Five faults are worth looking for deliberately (all examples invented):
| Fault | How it shows | Guard |
|---|---|---|
| Interval on the wrong line | The manual's table puts the hydraulic filter at 500 hours; the draft says 250, from the next column over | Read the verbatim cell against the printed table row, not from memory |
| Unit swapped or converted | Printed "250 N·m (184 lb·ft)"; the draft shows "250 lb·ft" | The prompt forbids conversion; the checker confirms every number has the manual's unit |
| Dropped condition | An interval that halves when the machine runs a breaker appears only as the standard interval | The separate conditions list must be non-empty for any machine with attachments |
| Merged checks | "Check fluid levels" replaces three rows with different procedures | One check per row, each with its own page |
| Wrong serial range | The manual covers two ranges and the draft mixes them | Put the serial range in the prompt and the source register |
The third fault is the quietest. Interval notes tied to working conditions often sit in footnotes or boxed text, and a summary tends to keep the headline and lose the condition. That is why the prompt asks for conditions as a separate verbatim list, and why a breaker on the mini excavator should send you looking for one.
The same manual-to-checklist method, applied to a different kind of equipment, appears in building a gym equipment maintenance checklist from manuals and building service schedules from manuals for farm equipment. The pattern is the same; the stakes on a civil site are higher, which is why the page reference is mandatory here.
From the maintenance chart to a service interval sheet
The manual's maintenance chart lists services by hour interval: a short daily or 10-hour item, then longer intervals, each with its tasks. The model's job is to lay that chart out as a sheet. The spreadsheet's job is to do the sums. Each service row carries the interval and the calendar limit copied from the manual, with the page, and the sheet works out what is next due. Check whether your manual pairs hours with a calendar limit ("whichever comes first"). Where it does, the sheet needs both.
An example with invented intervals and readings:
| Machine | Service interval | Hour meter now | Last done at | Next due at | Hours left | Average h/week | Weeks left |
|---|---|---|---|---|---|---|---|
| EX-01 | 250 h | 4,312 | 4,150 | 4,400 | 88 | 28 | 3.1 |
| DM-01 | 500 h | 2,873 | 2,500 | 3,000 | 127 | 22 | 5.8 |
| RL-01 | 250 h | 1,096 | 1,000 | 1,250 | 154 | 12 | 12.8 |
The columns after "last done at" are formulas, and so is the due date, which needs today's date and so does not appear in the table. In the real sheet, let C be the interval, D the hour meter now, E the hours at the last service, H the average hours a week, I the date of the last service and J the calendar limit in months. Weeks left is hours left divided by average hours a week. Then:
F2 (next due at, hours): =E2 + C2
G2 (hours left): =F2 - D2
K2 (due by, date): =MIN(TODAY() + G2 / H2 * 7, EDATE(I2, J2))The MIN picks whichever comes first, the usage-based date or the calendar limit, so a lightly used roller still gets its service on time. Copy the intervals and calendar limits from the manual's table with their pages. Do not let the model estimate the usage rates: those come from your hour-meter readings, written down weekly.
Hour-meter typos are an easy data fault, and worth guarding against deliberately. An operator types 4,132 for EX-01 when the meter reads 4,312. The sheet now shows 268 hours to go instead of 88, and the machine runs 180 hours past its service before anyone notices. Two rules catch it: highlight any reading lower than last week's, and any weekly jump bigger than the machine could plausibly have run. For a machine that averages 28 hours a week, a jump over 90 deserves a look. With last week's reading in D2 and this week's in D3, the check is a single cell:
=IF(OR(D3 < D2, D3 - D2 > 90), "check reading", "")Ask for a photo of the meter with each weekly entry, so a flagged reading can be settled by looking rather than by argument.
One more sorting job suits the model. A maintenance chart mixes tasks for different people, and the AI can sort every row into a tier for you to confirm:
| Tier | Who | When | Where it comes from | Where it lives |
|---|---|---|---|---|
| Pre-start | Operator | Every shift, before work | The walk-round chapter | Cab sheet or app |
| Operator-level periodic tasks | Operator, sometimes with a fitter | At the short intervals the manual gives, such as greasing | Maintenance chart, short intervals | Weekly sheet, signed off |
| Interval services | Fitter | At each hour interval | Maintenance chart, longer intervals | Service interval sheet |
| Calendar tasks | Fitter or inspector | On the dates the manual or your safety rules give | Manual and safety rules | Same sheet, calendar column |
Moving a task from the fitter to the operator is a competence decision, not a drafting one, so the competent person confirms every row's tier. Pre-start checks also do not replace the periodic inspections your own safety rules require, and the calendar tier is where those belong.
The AI helps in two places. It builds the sheet's structure from the chart, and every Monday it can turn the rows that are due within, say, 25 hours or 14 days into a short reminder email to the fitter. The dates and the due flags come from the sheet. If you ask the model to "work out what is due" from a pasted list, it will answer fluently, and a wrong date reads as plausibly as a right one.
A defect log that forces a decision
A defect log that records faults without a decision attached becomes a diary. Give each entry these fields: ID, date and time, machine, hour-meter reading, who reported it, the defect in the reporter's words, a photo reference, severity, action, whether the machine was tagged out, parts ordered, date fixed, who checked the repair, and the date it returned to service.
The severity scale is a competent person's to set, and the examples below are headings for them to replace. For instance: S1 stop (do not operate), S2 restrict (operate only within stated limits) and S3 monitor. Anything the manual labels "do not operate" is S1 by definition. One rule goes into the prompt as well: the AI may suggest a higher severity than the reporter's wording implies, but it may never lower one.
Here is the rule at work. The operator writes: "Wet patch on the left stick cylinder hose again, wiped it yesterday." Asked to categorise and suggest a severity, the AI returns "Hydraulics, hose leak, S3 monitor (minor weeping)". The competent person sets S1 until the hose is replaced, because the leak came back within a day, and records the reason. The AI did not have that rule, so the instruction to the model says that a repeat of the same fault on the same machine within a week is raised one level automatically and flagged to a person.
The log also pays off in aggregate. Once a month, paste the entries (without names) and ask for repeats by machine and component, which parts keep failing and what is still open. A realistic summary reads:
Open S1: none. Open S2: DM-02 rear light cluster, reported Tuesday, part ordered. Repeats: EX-01 left stick cylinder hose, three entries in six weeks, replaced once. Pattern: four of this month's nine defects were hydraulic hoses on the two excavators. Suggested for the competent person: review hose routing and inspection frequency on EX-01 and EX-02.
Two guards keep that summary honest. It must list every open S1 and S2 item verbatim, never "minor issues". And its suggestions go to a person as questions, not as instructions. The wider technique of mining logs for repeat failures is covered in AI in plant engineering: breakdown logs and maintenance plans.
Paper, a form or an inspection app
Where the check lives decides whether it gets done. Three options, with the facts as last read:
| Option | Cost | Good for | Watch for |
|---|---|---|---|
| Laminated sheet and a daily paper log in the cab | Printing | Sites without signal; operators who dislike phones | Lost pages, illegible writing; photograph or scan the log weekly |
| Google Form feeding a sheet | Free with a Google account | Quick start; the owner can set allowed file types, number and size for photos | Google's help says responders must sign in to a Google Account to upload a file |
| Inspection app such as Mitti (by SafetyCulture) | Free for up to 10 seats with 5 active inspection templates; Premium $24 a seat a month billed annually, $29 monthly, with unlimited templates | Photos, issues and reports in one place; its AI Assistant is described as creating checklists, training and information in seconds | The pricing page does not mention offline use, and the home page mentions it only for documents; ask whether inspections can be completed offline before buying for sites with no signal |
Do a quick sum before choosing an app. The invented fleet needs seven templates, and the free plan allows five active ones, so two machine types would stay on paper or you would pay for Premium. If all of six operators and two supervisors need Premium seats (check how your account would count them), that is eight seats at $24: $192 a month on annual billing, or $2,304 a year. Whether that beats paper for a firm of this size is the owner's call. Another option is to keep the free plan for the five busiest machines and run the other two on paper, then review after a quarter.
The app's AI helps with drafting too. Mitti's page also describes an "AI Issue Capture" feature that turns photos or voice notes into issues. Treat both as drafting aids on the same terms as any other AI output: nothing goes live without the competent person's tick. Seat limits and plan details change, so check the current page.
Approval, version control and when a sheet must be reviewed again
Each approved checklist carries an approval block, printed on the sheet and stored in the file: machine and template ID, manual edition and serial range, prepared by (including that it was AI-drafted), verified by (initials, date), approved by (the competent person's name, date), trial machine and date, revision number and next review date. A sheet without a signed block is a draft, whatever it looks like.
Re-review it whenever any of these happen:
- The maker issues a new manual edition for that machine.
- A machine of the same type arrives with a different serial range.
- A new attachment is fitted, such as a breaker or an auger.
- An incident, near miss or repeated defect points at a missing check.
- Operators report that an item is unclear, impossible or ignored.
- The scheduled review date passes.
Retire superseded sheets by removing them from the cabs the same day and archiving the old version with its approval block. For hired-in plant, build a sheet from the hire company's manual and run the first check on delivery, not after the first shift. The wider question of how a construction company writes these rules into a policy is in an AI policy for a construction company, and the fleet register itself is covered in an AI asset register: spreadsheet or software.
A trial week on one machine
Do not roll a new checklist out across the fleet. Pick the machine used most, and have its regular operator run the sheet for a week while the competent person watches at least one check from start to finish. Look for four things: items operators skip, items that cannot be done safely or are hard to reach, items with no description of what a fail looks like, and wording that sounds like the manual and not like the machine. Fix the sheet, bump the revision number and then copy the method to the next machine.
After a month, measure two numbers. The first is completion: days a check was done divided by days the machine worked. The second is where defects are found: before use, at the pre-start check, or after something has already gone wrong. An illustrative month for one excavator: 22 working days, 21 checks completed (95%), 6 defects found at pre-start and 1 found in use. The one found in use goes straight back to the competent person as a question: which check should have caught it, and does the sheet have it?
Questions before you feed a manual to an AI tool
Can the AI read a scanned manual?
Possibly, but nothing guarantees it. Anthropic's upload article does not mention scanned PDFs, and a scanned page is only a picture, so on a long file read as text only there may be nothing to read. Test it: ask the model to quote one interval word for word and compare it with the printed page. If it cannot, get a text-based PDF from the maker or the dealer.
What if an older machine has no manual?
Ask the manufacturer or dealer for a copy, quoting the model and serial number from the plate. Do not let the AI write checks from general knowledge and present them as the maker's. If none exists, your competent person writes the checks from the risk assessment, and the AI can help format the sheet, not source it.
Do pre-start checks replace periodic inspections?
No. A pre-start check is the operator's look before work. Periodic inspections, and any examinations your safety rules require for lifting or pressure equipment, are separate duties. Ask your competent person or safety adviser which apply to each machine, and keep them on the same service sheet so none is forgotten.
Should the AI diagnose faults from defect photos?
It can describe what a photo shows and sort it into a category. It should not diagnose, and Anthropic's documentation says its image outputs can be wrong on low-quality, rotated or very small images and should be verified. A fitter diagnoses; the AI helps with the log and the weekly summary.
Further reads
- Plant Hire Software: What Small UK Yards Should Compare Before Buying — What to compare before buying plant hire software
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Anthropic help article on uploading documents (updated 23 July 2026) and vision documentation; Mitti (by SafetyCulture) pricing and home pages, October 2026; Google Forms help page on question types. Every machine, interval, page number and fleet figure is invented for illustration; real values come from your manual.