When to Split AI Instructions: Start with Inquiry Sorting
Split AI instructions around jobs whose results help you narrow down what to correct when the output misses your intent.
Key points
- Keep the table to locate errors
- Split only sorting you correct daily
- Compare correction and handoff work
Overview diagram in preparation
Contents
Keep the classification table to narrow the fix
A&A perspective
Splitting instructions means separating what you ask AI to do and the result you check for each job. For inquiries, keeping the classification table and draft reply separately helps you decide which one to correct first.
Hypothetical example
Suppose your instructions began with sorting morning inquiries and drafting replies, and you later added invoice checks and meeting summaries. Adding instructions every time a reply misses the mark may not reveal whether the classification was wrong or only the wording was wrong: the final reply alone may not show the difference.
From the sources
A sales representative on Anthropic’s Japan team gives each agent one job. An agent is an AI system entrusted with a task. The representative explains that a narrow responsibility makes each agent easier to follow, including when something seems wrong.
From the sources
He also identifies continually adding instructions instead of updating them as a common failure.
A&A perspective
This is his own account of how he works, not an experiment comparing separate and combined instructions.
A&A perspective
Automate & Augment (A&A), the organization writing this article, suggests first checking whether you can inspect an intermediate result. Before adding another AI, you can ask a single set of instructions to produce a classification table and then a draft reply.
Hypothetical example
For example, suppose a design company receives an email asking to replace an image it has already delivered. If the table labels it a new inquiry, first check the classification rules and how the email was read. If the label correctly says post-delivery change but the reply invites the sender to a new sales call, check the reply instructions.
A&A perspective
Splitting the jobs does not guarantee a single cause. The classification rules might be outdated, the original email might be missing, or the table might not have reached the AI drafting the reply. The table helps you choose where to look first; it does not prove the cause.
A&A perspective
“Stop repeating yourself to AI: split company info into 3 parts” covers how to supply company information each time. The decision here concerns where to divide daily work so that it is easier to correct, rather than where to store reference material.
Separate only the sorting you keep correcting
A&A perspective
The first job to separate is one you run daily and correct repeatedly. If you fix inquiry sorting every morning, separate only that job. Keep using reply drafts and occasional meeting summaries as you do now.
A&A perspective
In A&A’s proposed design, the boundary depends on what a job receives and returns, rather than its name. “Handle sales” does not say when the job is finished. “Read incoming emails and return a classification table” gives you a result to check before a reply is written.
Hypothetical example
In the hypothetical design-company example, the table keeps the original email, category, reason for the decision, and whether the item is on hold. The categories are new inquiry, ongoing project, and post-delivery change. The original email remains accessible; the next job does not receive only a short AI summary.
Hypothetical example
The sorting instructions contain the classification rules and say to put an item on hold when its category cannot be decided. The reply instructions accept the table and return held items to the person responsible, rather than deciding their category. Neither job sends anything to the customer.
Hypothetical example
In the same example, an email saying only “I want to change the thing from before” should not receive a final category while the project remains unknown. The table marks it on hold, and the owner checks the original conversation. Asking the reply AI to guess again would make it harder to trace where the decision changed.
Hypothetical example
Separating the jobs does not require updating two independent copies of the classification rules. Keep one authoritative version, and have the reply AI use the confirmed category. When the rules change, check whether older classification tables can still be used as they stand.
Hypothetical example
You can begin by inspecting the table yourself and handing it to the next AI. Build an automatic handoff only after the required table contents and the handling of held items are clear. This does not require the same customer-management software or execution infrastructure as a large company.
A&A perspective
The related article “AI Agent or Workflow: Acceptance Terms for Your Vendor's Quote” covers the implementation approach and acceptance terms when commissioning a vendor. The table here helps an owner decide which job to separate and which to keep together before commissioning that work.
| Decision factor | Keep the jobs together | Separate only sorting |
|---|---|---|
| Checking the result | The current instructions expose the table and reply separately | The reply alone hides the cause; check the table separately |
| Where corrections recur | One job with few corrections | Sorting runs daily and needs repeated correction |
| Added work | Cleaning up instructions is enough | The split remains manageable including table handoff and checking |
Compare correction work and handoff work together
A&A perspective
Decide whether to keep the split by counting handoff and checking work as well as corrections. Even if correcting categories becomes easier, spending more time finding and passing the table each morning may leave the owner’s workload unchanged or higher.
From the sources
In a 2024 design article, Anthropic recommends starting with the simplest approach and adding complexity only when it is needed.
A&A perspective
This article draws on that design principle. It does not recommend the products or configurations described then for present-day use.
Hypothetical example
A&A’s proposed comparison uses inquiries you have corrected before. Give the same emails and classification rules to the combined instructions and the separated instructions. Compare whether the categories are right, and compare the work needed to make each draft reply ready to send.
Hypothetical example
If the table is wrong, have a person correct it and pass that corrected table to the reply AI again. If the reply still misses the intent, correcting classification alone has not solved the problem. First confirm that the correct table arrived, then inspect the reply instructions.
Hypothetical example
Keep the input email, instruction version, classification table, draft reply, and edits. Record time spent finding the cause, making corrections, and handing off and checking results separately. Recording usage charges separately also lets you assess a reduction in work that comes with higher costs.
A&A perspective
Do not judge the split by one successful attempt; check whether it also helps you narrow the correction in ordinary work. No time savings or error rates have been measured here. The table is a proposed design for comparison, not a promise of results.
A&A perspective
If you delegate one task a day and correct it only a few times a month, keeping the instructions together is another option. Cleaning up a long set of instructions and updating old conditions may be enough. Work volume or instruction length alone is not a reason to split.
A&A perspective
If rereading the same material or waiting for the other AI’s result creates a larger burden after the split, consider combining the jobs again. The aim is to retain a way to find and correct mistakes yourself, rather than to increase the number of AIs.
Choose what to separate and what to keep together
A&A perspective
Keeping the classification table and the draft reply lets you choose whether to start by correcting sorting or reply writing.
Start by separating only the sorting you correct every day. Reply drafts and occasional jobs can continue as they do now. Decide whether to keep the split by combining the work of finding the cause, making corrections, and handing off and checking results.
FAQ
At what instruction length should I split the jobs?
Do not decide by character count alone. Check whether you still cannot locate the correction after keeping intermediate results such as a classification table. This article does not present an experiment establishing a length threshold.
Do I need to subscribe to another AI service?
You do not need to buy another service to try this division. Start with your existing tools, separate the instructions by job, and have a person pass the results along. Check features and costs separately if you move to automatic execution.
Should I duplicate information both jobs use, such as a price list?
Keep one authoritative version and let the jobs that need it refer to it. Independently updating copies of a price list can leave inconsistencies. Splitting the jobs alone does not fix errors in shared information.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- How an Anthropic BDR in Japan runs his day with Claude Managed Agents
Anthropic · 2026-10-16
Accessed 2026-10-29 - Building effective agents
Anthropic · 2024-12-19
Accessed 2026-10-29