The Prompt: The Instruction That Shapes the Output

A prompt is the instruction supplied with every request, and in a running system it is production configuration rather than a message somebody typed. The instruction sets the role, names the material, states what to produce, states what may never be asserted, fixes the format, and says what to do when the material does not support an answer. The last of the six is the one most often missing, and the absence shows directly in the output.

Everything in that instruction is done again on every single request. Nothing carries over from one request to the next, so the instruction is supplied fresh each time alongside the material, and it is the whole of what the component has been told. Change it and the change lands on the very next request, on every channel at once, with no version of the component changing and often with nothing written down anywhere. Immediate effect plus no record is the combination that makes an instruction a control, and it is also the combination that makes almost nobody treat it as one.

What is a prompt, once something is actually running?

Consider the counter of a busy sweet shop in wedding season. Two kinds of instruction are in the room and they are easy to confuse. One is the note the owner hands the clerk about the customer standing in front of him: give this one the bulk rate, he has ordered before. The note about the customer is a message. A message applies to one person, at one moment, and then it is gone. The other is the laminated sheet taped to the wall above the till: who gets the bulk rate, in what order to check, what to write on the bill, and what to do when somebody asks for a thing the sheet does not cover. The laminated sheet is not a message to anybody. The sheet is the setting the counter runs at.

Now watch what happens when the owner climbs up on a Tuesday afternoon and edits line four of the sheet. Every bill from that moment carries the change. There is nobody to tell, so nobody is told. The clerk is the same clerk and the till is the same till. The bill records the price and not the sheet, so a month later, holding a bill from that Tuesday, nobody can say whether it came from the old line four or the new one. A promptThe instruction supplied with every request, telling the component what to do and what it may say. in a running system is the sheet and never the note, and every difficulty that follows comes from that one fact.

Take the case at hand. At Sumeru Bank Limited, invented, component 8 of the intake chain is a drafting assistant. The assistant produces the first draft of two things: the exception noteThe written record a desk officer makes on a file that stopped, saying what stopped it and what was decided. a desk officer writes on a stopped file, and the explanation paragraph inside a decline letter. A person reads, corrects and signs both. The assistant runs on 3,010 exception notes a month. The instruction supplied with each of those 3,010 requests is the same 340 words in six numbered parts, and the component itself is one version held still all month. A fixed instruction against a component held still is what makes the 340 words production configurationSomething that shapes behaviour on every run and can be changed without any new version of the component being built. rather than correspondence.

What actually goes into one, and in what order?

The instruction for component 8 is 340 words in six numbered parts. Three hundred and forty words is short. Read aloud the instruction takes under three minutes, and a reviewer can hold the whole of it in their head. Holding the whole instruction in one head is a property worth protecting rather than an accident. The six parts between them set the whole of the component's behaviour for that request. Nothing else is told to the component, and nothing else survives from the request before.

The order is not decoration. Each part is qualified by the one above it, so reading them out of sequence changes what they mean. Part 1 has to come first because every later part is an instruction to whatever part 1 said the component was, and part 6 has to come last because it is the exception to all five parts above it. Put the insufficiency case at the top and it reads as the main job; put the role at the bottom and the four parts above it have been addressed to nobody in particular.

What does part one do, and why does it come first?

Part 1 has two halves and they do different jobs. The first half names the role: a desk officer drafting an exception note on a stopped retail loan file. People write the role half easily. Acting as a desk officer is a costume rather than a constraint, so on its own that half is close to worthless.

The second half of part 1 is the load bearing half, and it fixes the standard of assertion: the draft may state only what the supplied material states. The standard of assertion is what every later part leans on. Part 4 can list what may never be stated only because part 1 has already said that everything stated must come from somewhere. Part 6 can tell the component to stop only because part 1 has established that stopping is preferable to asserting. Take out the second half and the other five parts turn into style preferences.

Think about the difference between a clerk told to act like a manager and a clerk told he may commit the shop only to what is written on the order form in front of him. The first is a manner. The second is an authority limit, and it is the only one of the two that anybody can check afterwards.

What do parts two and three do, and what is the window?

Part 2 names the material supplied with this request: this file and nothing else. Part 3 says what to produce from it: a first draft of the note for the officer who will sign it, covering what stopped the file, what the material shows, and the decision the officer is being asked to make. Parts 2 and 3 are the only two that point at something outside the sheet, and pointing outward is what gives them a size problem.

The material for an exception note at this bank averages about 4,800 words, and the longest five per cent run past 12,000. Everything supplied on a request has to fit inside the windowThe total amount of instruction and material a component can hold on one request, after which something has to go., and where the material exceeded it, the oldest part was dropped silently. Nine of the 200 notes in the first measurement, being 4.5 per cent, were drafted from a file that had been quietly shortened this way. Nothing in those nine drafts said so. A truncated file produces a confident note about the part that survived, and the silence about the truncation is the whole of the problem.

Two practical consequences follow. The first is that part 2 should name the material in a way a person can verify. A reviewer can then ask whether what part 2 promised is what actually arrived. The second is that the instruction and the material compete for the same room, and a longer instruction is not free. How the material is chosen and found is covered separately, and what matters here is only that some material is supplied and that its size is finite.

What does part four do that part one has not already done?

Part 4 lists what may never be stated: no reason the file does not carry, no rate or period or timescale, and nothing about what happens next where no written procedure says. Part 4 looks redundant beside part 1, and it is not, for the same reason a house rule list is not redundant beside a general instruction to behave. Part 1 sets a standard. Part 4 names the specific things that this bank has been bitten by, in the words its reviewers use.

Part 4 is also the part with a matching check, and the check is what makes part 4 operational rather than decorative. Verification of a draft at this desk is five numbered checks, and check 3 is the one that asks whether anything has been asserted that the material does not contain. Check 3 asks a reader to notice something that is not there rather than to compare two things that are, so it is the slowest of the five. A part 4 written in the reviewer's own vocabulary is what makes that check finishable in the time the desk actually has.

Why spend words of a 340 word instruction on the format?

Format looks like the least important part on the sheet. The required statements live inside the format, so it is not the least important at all. The note has to name which of the six exception causes stopped the file, in the fixed vocabulary the desk counts by, and it has to state what the officer is being asked to decide. A note missing a required statement is defective however well supported every sentence in it is, and a month of such notes quietly breaks the desk's own counting.

A delivery challan has a fixed set of boxes for the same reason. The driver may write a beautiful description of the goods; if the box naming the receiving branch is empty, the paperwork does not do its job, and nobody discovers that until the month end reconciliation. Part 5 turns the note from a written paragraph into a record with fields, and records with fields are the only things a desk can count.

What does part six say, and why was it missing until month nine?

Part 6 says what to do when the supplied material does not support an answer: say so, name what is missing, and produce no draft. Part 6 is four lines. The four lines were added in month 9, five months after the chain went live, and until then the sheet ended at part 5.

Ask why nobody noticed. The answer is uncomfortable. A sheet with five parts does not look incomplete. Each of the five is present, each does its job, and a reviewer reading it will find nothing wrong with any line. The missing part is invisible in the instruction and visible only in the output. An omission is not a gap in the wording, it is a behaviour in the world. Why a component produces something on every request rather than nothing is set out under how a component produces text from an instruction; given that, an instruction with no part 6 has left the insufficient case unaddressed, and the component fills it the only way it can.

The everyday version is a helpline script with no line for the call it cannot handle. Every other branch is written out and the operator sounds competent throughout. Ending the call with an admission was never one of the options printed in front of the operator, so the one caller in twenty with a problem outside the script still receives a confident answer.

What did adding part six actually buy, once somebody measured it?

The drafting assistant was measured twice at this bank, on 200 notes each time, and the two measurements sit either side of the change. In the first, in month 6, 23 of the 200 drafts contained a statement that was not in the file, being 11.5 per cent. Twenty one of those were caught in verification and 2 reached a customer, and 21 plus 2 is 23. There was no part 6, and declining was not among the things the component had been told it could do, so it declined to answer 0 times out of 200.

In the second measurement, in month 10, the unsupported count stood at 7 of 200, being 3.5 per cent, and the component declined 12 times, being 6.0 per cent. The unsupported rate improved by 8.0 percentage points. Three separate changes were made between the two measurements and the bank never separated them, so the 8.0 points cannot be handed to part 6. A retrieval arrangement was added in month 8. In month 9, part 6 was added and, in the same week, a connection was built letting the component read the file's structured fields instead of restating figures out of prose.

One thread can be pulled clean. Of the 23 first-measurement errors, 6 were restated figures and 17 were other statements, and 31 of the 200 notes restated a figure at all, so 6 of 31 were wrong, being 19.4 per cent. In the second measurement 34 notes restated a figure and none was wrong. No other change touched figures, so the connection to the structured fields is the cause of that particular improvement. The rest is shared between the retrieval and the instruction, and this bank never split it. A deployment looks like this whenever three good ideas ship in the same quarter.

And one movement belongs to part 6 alone, by construction. Before part 6 no instruction permitted declining at all, so nothing else could have moved the declining rate from 0 to 12 of 200. The measured effect of part 6 is the appearance of a behaviour that did not previously exist, not a fall in errors that three changes share.

Is declining to answer a failure, or the arrangement working?

Declining is the arrangement working, and it is not free, and a report that gives only the first half of that sentence has misled its reader. Both halves are simple enough to hold together. Each of the 12 declined notes is a note where the material genuinely did not contain what was needed, and the alternative to declining was not a correct draft but an unsupported one. Before part 6, those cases did not disappear. The insufficient cases became drafts that asserted something the file did not say, and the drafts went into the pile verification had to catch.

Now the cost. A declined note arrives at the desk with nothing on it. The officer writes it from nothing. Writing a note from nothing took 18 minutes at this bank before the component existed at all, so the 12 declined notes in a round of 200 cost about 216 minutes. Scale that to the month and it is about 181 notes and roughly 3,250 minutes, being about 54 hours. Set against a fully loaded cost of Rs 9,00,000/- a year a post and a working month of 8,400 minutes, that is about 0.39 of one post, or roughly Rs 3,48,000/- a year of desk time that exists because part 6 exists. Both of those are arithmetic on this bank's own assumed figures rather than measurements.

None of that is an argument against part 6. The cost is an argument against a report with one line in it. A report that leads with the unsupported rate falling from 11.5 per cent to 3.5 and then stops leaves a reader free to conclude that the change cost nothing. A decline to answerProducing no draft at all, because the supplied material does not support one. is a transfer of work from the checking step to the writing step. A report that shows only the checking step shows only the half that improved.

Who can change a prompt, and what should that need?

At most places the honest answer to the first half is: whoever can edit the file. The instruction usually sits in a configuration file beside the other settings, and the people with access to that file are the people who run the service. Nothing about the arrangement announces that this particular file is different from the one holding timeouts and retry counts.

One question settles whether it is different, and it is not a question about storage. Ask whether changing the instruction changes what reaches a customer: if it does, it is a controlSomething a firm relies on to keep an outcome correct, and therefore governs, records and reviews., whatever file it happens to live in. Where it sits, what format it is in, whether it is code or text, and who is accountable for the repository are all beside the point. The instruction for component 8 shapes the explanation paragraph inside a decline letter that a real applicant reads. A decline letter is as close to a customer as anything in the chain gets.

The process that follows from a yes is not heavy. A named person accountable for the wording, a change recordWho altered the instruction, when, what it said before, and who approved the change. holding who changed it, when, what it said before and who approved it, a version identifier travelling with each request so a finished note can be tied back to the words that produced it, and a review when it changes materially. Four lines of process around 340 words is the whole of it.

What happens when nobody treats the instruction as a control?

Nothing dramatic happens, and the absence of drama is the difficulty. At Sumeru Bank Limited the instruction was changed 11 times between month 4 and month 12. Seven of those changes carry a record. Four do not: there is no note of who made them, and no copy of what the wording said before. The instruction sat in a configuration file that nobody had classified as anything in particular, and everyone with access to that file could edit it, exactly as they could edit any other setting beside it.

Who actually reads the instruction, and what do they do with it?

Four people at this bank have a use for those 340 words, and none of them is the person who wrote them.

Ismail Sheikh, who heads the exception desk, reads part 6 as a rostering line. Six per cent declined is about 181 notes a month arriving with nothing on them, and he needs those 54 hours in the roster whether or not anybody else calls them a cost. For an operations head the instruction is not text, it is a forecast of how much writing his desk will do next month.

Neelima Rao, in the risk function, does the independent validation. Her first request is not a sample of outputs, it is the instruction and its change record. A sample of outputs from a stretch where the wording moved four times without a record is a sample of nothing in particular. She reads 340 words in a few minutes, against the eleven working days the review of the scoring model took, and the reading tells her what the outputs are even supposed to be.

Revathi Balan, head of retail credit, is accountable for what goes into a decline letter that an applicant reads. She did not write part 4 and probably could not name it. If a paragraph in one of those letters states a reason the file does not carry, the question she will be asked is not which model produced it, and the answer she needs is a wording and a version.

Ashok Pillai, in technology risk, ran the register sweep. His problem is the governing one: whether the instruction is an item in its own right or a property of component 8. Registered as a property, the instruction inherits component 8's classification, and that classification was the one that scoped it out. Registered as an item, somebody has to own up to the wording. The instruction is not a document any of these four would have asked for a year earlier, and it is the first document all four of them now want.