Finding Out Why a Prompt Stopped Behaving

The situation

A support tool summarises every closed conversation into a JSON object: a one-line outcome, the products discussed, a sentiment label, and a boolean for whether a refund was promised. A Lambda parses that object and writes a row into a reporting table. It ran for four months and nobody thought about it once.

Two weeks ago the reporting table started missing rows. The parser now throws on roughly one call in six, and the responses that fail are not malformed JSON. They are prose. “The customer contacted us about a delayed delivery and was offered a replacement, which they accepted.” Perfectly good English, entirely unparseable, and no pattern anyone can see in which conversations produce it.

Nobody edited the template. Amazon Bedrock Prompt Management shows the same published version the application has referenced since March, and the deployment history for the service is empty for the period. Pasting the system promptSystem promptThe instruction block that frames the model’s behaviour for a session, separate from the user’s messages. into the console and running it against a sample transcript produces clean JSON ten times out of ten. The behaviour that fails in production does not reproduce anywhere a developer can watch it.

What actually matters

The template is not the prompt. What reaches the model is the template, plus the values filled into its input variables, plus whatever the retrieval step returned this time, plus as much conversation history as the assembler fitted into the window, plus any tool definitions. Exactly one of those is under version control, and it is the one everybody looks at first because it is the only one with a URL. Running the template in a console reproduces a fraction of the real input, so of course it behaves; the console is testing a different prompt from the one that failed. Before any diagnosis is possible, the rendered prompt has to stop being a transient string inside a request handler and start being an artefact somebody can read after the fact.

Once retrieved passages and user text sit in the same channel as your instructions, nothing in the input marks one as instruction and the other as data. Take a support transcript that quotes a customer email saying “please reply in plain English, no technical jargon”. That is instruction-shaped text, arriving after your formatting rule, from a source the template never anticipated. Given two directions that conflict, the output follows the later and more specific one, which here is the customer’s. That is prompt confusion, and the template cannot explain it because the template is not wrong. It also explains a rate rather than a switch. A cause that fires only when a conversation happens to contain instruction-shaped text produces one failure in six, not a clean break at a deploy boundary. That is why the timeline of code changes has nothing in it. The related deliberate case, where the instruction-shaped text was planted, is the subject of indirect prompt injection in a retrieval system; the accidental case has the same mechanism and none of the malice.

Format failures need a machine detector. A person reading a sample of outputs is a poor instrument for format inconsistencies, because prose reads well and the eye skims past a missing brace far more readily than it skims past a wrong number. The downstream parser is a detector of sorts, but it fires late, in another service, with no memory of what was sent. Schema validation applied to the response as it arrives turns the shape of the output into a per-response verdict and a metric. “One in six” becomes a number on a dashboard rather than a guess assembled from support tickets. That validator also has to run in production and not only in tests, because the inputs that break the format are production inputs.

Two properties decide which instruments are worth having. The first is attribution. Whatever you turn on has to carry a correlation identifier from the inbound request through assembly, retrieval and the model call. That is what turns “this reporting row is missing” into “here are the exact bytes that produced it”, rather than a search through a log group by timestamp. The second is whether it works backwards. Instrumentation added today explains failures that happen after today, which is tolerable at one in six and useless at one in five hundred. Against both sits the cost of recording. Recording the rendered prompt means recording the customer’s words verbatim, so the log group holding it inherits every obligation the source conversation carried. Keeping personal data out of prompts and logs is decided at the same moment as the decision to log at all, not tidied up afterwards.

What we’ll filter on

1. Rendered or template only: does the instrument show the exact bytes sent to the model, or just the wording you authored?
2. Attribution: can one bad answer be tied to one run, with the variables, retrieved passages and history that produced it?
3. Retroactive: does it explain a failure that has already happened, or only the next one?
4. Change detection: does it show what differs between a configuration that behaved and one that does not?
5. Cost and exposure: what does it store, for how long, and how much personal data does it copy in the process?

The landscape

Structured logging of the rendered prompt to Amazon CloudWatch Logs. The assembler emits one JSON log line per invocation. It carries the correlation identifier, the prompt identifier and version, the variable values, the identifiers and text of the retrieved passages, the number of history turns included, the token count, and the final rendered string. This is the only instrument that shows what the assembler actually produced, because it sits inside the assembler. It costs ingestion and storage on every request, it copies the conversation verbatim into CloudWatch Logs, and it needs a retention policy and a redaction step decided up front. Sampling cuts the cost, and cuts the chance that the sampled run is the one that failed. The usual shape is to log every request during an investigation and drop back to failures only afterwards.

Bedrock model invocation logging. Enabled once per account in each Region, this writes the request and response payloads for every Amazon Bedrock invocation to S3 or to CloudWatch Logs without touching application code. It captures what Bedrock received, which is the rendered prompt as sent, and what came back, and where the destination is CloudWatch Logs, the generative-AI observability view reads those records, so the input and output of a given request identifier are readable in the console without writing a query. Three limits shape how you use it. It records nothing about how the prompt was assembled, so a retrieved passage in the payload is text with no provenance. Bodies over 100 KB are not inline; they land as separate objects under the data prefix of an S3 bucket, and a CloudWatch destination has one only if you configured an S3 location for large data delivery. Attribution is opt-in. Each record carries Bedrock’s own requestId rather than yours, so joining it to your correlation identifier means matching on time and content unless you send request metadata: up to sixteen key-value tags on a Converse or InvokeModel call, written into the log entry under requestMetadata and filterable in Logs Insights. It is the quickest route to the rendered prompt when the assembler was never instrumented.

AWS X-Ray spans around assembly, retrieval and the model call. A trace per request, with a span for each stage, turns a bad run into one object. The retrieval span carries the query and the document identifiers returned. The assembly span carries the token counts and which sources were included. The model span carries latency, stop reason and validation result. That is the prompt observability pipeline, and the spans land in AWS X-Ray with CloudWatch Transaction Search as the search surface over them. Wiring is the same as for any other generative-AI workload on this stack: Transaction Search enabled once for the account, the AWS Distro for OpenTelemetry in the application, and identifiers propagated inward, all covered in tracing an agent’s decisions in production. Transaction Search ingests every span rather than a sample and indexes 1 per cent of them as trace summaries by default, adjustable upward, so “show me the runs where retrieval returned document 4417” is a filter rather than a grep. Two numbers bound what spans will do here. A segment document is capped at 64 KB, so the rendered prompt does not fit and should not go there. Trace and service map data is retained for thirty days, which covers this investigation and not a quarterly one.

Version comparison in Amazon Bedrock Prompt Management. A version is a snapshot of the working draft taken at a moment and numbered from one upward; iteration continues on the draft, not on a published version. The snapshot covers the whole variant, model identifier and inference configuration included, not only the wording. The console compares two selected versions by showing their JSON side by side, highlighting the fields present in one and missing from the other, and it will run both against the same test variables. That makes version comparison the fastest way to eliminate or confirm an authored change, and to be sure the running application is invoking the version you think it is, since the version is a suffix on the prompt ARN it calls. The workflow that keeps this useful is described in managing prompts as a first-class resource. Pointing production at a version rather than a draft is also what makes rolling a wording change back a repoint rather than a redeploy.

Schema validation in the response path. A JSON Schema for the expected object, applied to every response before it leaves the service, with the outcome emitted as a metric dimensioned by prompt version. This is the detector that turns format inconsistencies from anecdote into rate, and it is the gate a retry or a fallback hangs off. It says nothing about cause. It tells you which runs to go and read, which is a different job from telling you why they failed. Where the output has to be structured rather than merely checked, constraining the model with tool use does more than validating after the fact, as in getting structured JSON out with tool use.

Bisecting against a saved test set. A fixed set of inputs with known-acceptable outputs, run against one configuration at a time, changing exactly one thing between runs. That is template testing, and it is why prompt testing frameworks exist. A harness renders the prompt from a named template and a named set of inputs, calls the model, validates the output, and reports a pass rate you can compare across runs. Building the set is the slow part, and building a golden dataset is the prerequisite for every other use of it too, including the regression gate in catching a regression after the deploy. It is the only instrument on this list that answers “did the fix work” rather than “what happened”.

Evaluation

Side by side

No row is a diagnosis. The first two produce the evidence and the third organises it. The fourth eliminates the authored configuration in about a minute, the fifth tells you which runs to fetch, and the sixth is how you establish cause and then prove the fix. The interesting column is the retroactive one, because it sets your first move. If invocation logging is already enabled in the Region, the failed calls from the last two weeks are already on disk and the investigation starts with reading them. If it is not, the first change is turning capture on and waiting, and at one call in six that wait is minutes rather than weeks.

Where the failure actually lives

Once the rendered prompt is in hand, the diagnosis is a sequence of eliminations against a saved input, each one changing a single component of the assembled prompt.

The ordering matters more than the individual tests. Eliminating the authored template first takes a minute and removes the component everybody suspects; removing retrieved passages next is the change most likely to move a format failure, because that is the component carrying text nobody wrote. History and variables come last because they change slowly and are the least likely explanation for a failure that started on no particular day.

The solution

Capture, isolate, fix, publish, prove. That sequence is what systematic prompt refinement workflows amount to in practice, and running it in order is what separates a diagnosis from a hopeful edit to the wording.

Capture the bytes

Log the rendered prompt to CloudWatch Logs as structured JSON at the moment of assembly. Carry the correlation identifier, the prompt identifier and version, the retrieved document identifiers, the token count, and a hash of the rendered string so identical prompts group without anyone reading them. Keep the full text behind a short retention and a redaction step; keep the metadata for as long as you like. If the assembler is not instrumented and the failure is happening now, Bedrock model invocation logging gives you the same payloads with one Region-level setting and no deploy. Records written before you started sending request metadata join back to your requests by time and content.

{
  "correlation_id": "req-8c2f41",
  "prompt_id": "PROMPT7QK2",
  "prompt_version": "4",
  "variables": { "transcript_turns": 22, "channel": "email" },
  "retrieved_docs": ["kb/returns-policy#3", "kb/delivery-windows#1"],
  "history_turns_included": 6,
  "input_tokens": 3184,
  "rendered_sha256": "9f1c...",
  "schema_valid": false
}

Make one bad run one trace

Put X-Ray spans around assembly, retrieval and the model call so that a failing correlation identifier resolves to a single span tree rather than four log searches. Prompt observability pipelines built this way answer questions the logs cannot: which retrieval results the failing runs have in common, whether the failures cluster on long inputs, whether latency moved at the same time. Put identifiers, counts and the validation outcome on the spans, and leave the prompt text in the log group where retention and redaction already apply to it.

Validate every response

Attach schema validation to the response, dimensioned by prompt version and by whether retrieval returned anything, and alarm on the failure rate rather than on individual failures. A validator that runs on every production response converts format inconsistencies into a measurement, which is what lets you say “one in six” with confidence and, later, say the fix worked. This is also the signal to surface on the operational view described in dashboards for a generative-AI feature, next to latency and cost, because a valid-shape rate is an availability number for anything downstream that parses the output.

Bisect one variable at a time

Take one captured failing prompt and one passing prompt, and run the gates from the diagram against the saved test set, changing exactly one component per run. Prompt testing frameworks are worth the setup here because one-change-at-a-time is hard to hold to by hand. The harness renders from a named template and a named input set, calls the model with the same inference configuration, validates the output, and records a pass rate. Template testing then produces a number to compare, rather than an impression that it seemed better. Two runs that differ in two things have told you nothing, and doing this by hand in a console is exactly where that mistake gets made.

Publish a version and prove the fix

Once the cause is isolated, correct the wording, publish it as a new numbered version, and point production at the version rather than at a draft. Then re-run the saved test set against the old version and the new one and keep both numbers. That final version comparison makes the fix evidenced rather than believed, and it leaves the next person a baseline. When this prompt drifts again in six months, there is a recorded pass rate for the version that was working, and systematic refinement has somewhere to start.

Worked example

The run that failed

Model invocation logging was already on, so the first move is a query over the last fortnight’s payloads for responses that do not parse as JSON. Forty-one of them, spread evenly across the two weeks, none clustered on a deploy. Reading three of them, the same shape appears in every rendered prompt. A retrieved passage from a knowledge-base article about accessibility, added to the index a fortnight ago, whose second paragraph reads “when replying to customers with this need, write in short plain sentences and avoid structured formats.”

The bisect

Gate one: the published version alone, against a clean transcript, returns valid JSON. The template is not the cause and version comparison confirms nothing has changed in it since March. Gate two: the same saved failing input with the retrieved passages removed returns valid JSON. Putting the accessibility passage back reproduces the prose response every time. That is prompt confusion: an instruction meant for a human writer, arriving in the same channel as the instructions with nothing marking it as data. It fires on one call in six because that is how often the article scores high enough to be retrieved.

The fix and the evidence

Two changes. The retrieved block gets delimited and labelled as reference material that must never be treated as instructions. The format rule moves to after the retrieved content rather than before it, so it is the last instruction in the prompt. Published as version 5. The saved test set, now carrying the forty-one captured failures alongside the original examples, scores 62 per cent valid on version 4 and 100 per cent on version 5. Schema validation stays on, and the valid-shape metric goes on the dashboard next to latency, where a recurrence shows up as a line moving rather than as missing rows in a reporting table three weeks later.

What’s worth remembering

1. The rendered prompt is the template plus variables plus retrieved passages plus history, and only the template is in version control, which is why a console reproduction proves nothing.
2. Instruction-shaped text arriving from a retrieved document causes prompt confusion the template can never explain, and it produces an intermittent failure rate rather than a clean break at a deploy.
3. Capture the rendered prompt in CloudWatch Logs with a correlation identifier, and accept that doing so copies customer text into a log group that now needs retention and redaction decided deliberately.
4. Spans around assembly, retrieval and the model call make one bad run one object instead of four searches, but a segment document is capped at 64 KB, so the rendered prompt stays in the log group.
5. Schema validation on every production response turns format inconsistencies into a rate, which is both the alarm and the measurement the fix is later judged against.
6. Systematic prompt refinement workflows are capture, isolate one variable at a time against a saved test set, publish a new version, then re-run the version comparison so the fix is evidenced.