# Plan an AI Context Budget Before Sending Business Documents

<a id="main-content"></a>

AI &amp; Automation

October 28, 2026  ·  9 min read  ·  Victor Valentine Romo

On this page

1. [State the Answer Before Collecting Documents](<#state-the-answer-before-collecting-documents>)
2. [Count the Whole Request](<#count-the-whole-request>)
3. [Work Through a Hypothetical Budget](<#work-through-a-hypothetical-budget>)
4. [Use Counts as Preflight Evidence](<#use-counts-as-preflight-evidence>)
5. [Select Excerpts With Enough Surrounding Meaning](<#select-excerpts-with-enough-surrounding-meaning>)
6. [Split Documents Around Questions](<#split-documents-around-questions>)
7. [Use Summaries as Maps Back to Evidence](<#use-summaries-as-maps-back-to-evidence>)
8. [Return an Explicit Incomplete Result](<#return-an-explicit-incomplete-result>)
9. [Keep Capacity Separate From Acceptance](<#keep-capacity-separate-from-acceptance>)
10. [Sources](<#sources>)

A folder contains everything the team knows about a supplier review. The next AI request needs an answer to one question.

Those are different units of work.

Build a context budget around the question. List the evidence required, count the assembled request, and reserve enough room for an inspectable answer. When necessary evidence cannot fit, return an incomplete result with the gap named.

The worksheet below uses fictional documents and hypothetical token allocations. It does not describe a vendor's context limit, price, or tested performance. No business records are transmitted.

<a id="state-the-answer-before-collecting-documents"></a>

## State the Answer Before Collecting Documents

Write the reader's decision in one sentence. For this example: identify which delivery requirements a fictional supplier's written response addresses, and list unresolved requirements.

That task needs the requirements, the relevant response passages, and a way to distinguish supported answers from missing evidence. It does not automatically need every meeting transcript or every previous draft.

Define the output before assigning space. Ask for one row per requirement, a source location, a short finding, and an unresolved flag. Preserve a final section for omitted evidence and limits on the conclusion.

[This output contract](</articles/ai-structured-outputs-schema-business-validation/>) gives the request a stopping point. A general instruction to “analyze everything” leaves both the evidence boundary and the expected answer open.

Give the request an identifier. Record the question, the document set, the selected excerpts, and the intended response format together. If a colleague changes the question, assemble and count a new request.

The [RAG architecture article](</articles/rag-architecture-business-knowledge/>) addresses retrieval. This worksheet starts after candidate evidence has been found. Its job is to decide what enters this request and what the answer may claim about that material.

<a id="count-the-whole-request"></a>

## Count the Whole Request

For this article's documented configuration, use the direct Claude Messages [API](</articles/api-first-business-operations/>) with `claude-opus-5-5`, text documents, and thinking disabled. The documentation was checked on October 28, 2026. No API behavior was tested here.

Claude's [context-window documentation](<https://platform.claude.com/docs/en/build-with-claude/context-windows>) counts instructions, messages, documents, tool results, tool definitions, and generated output toward the window. Cached input still counts. Thinking and retained thinking history add model-specific accounting when enabled.

Turn that inventory into a preflight table. List each component even when its allocation is zero. [A blank cell](</articles/null-blank-zero-business-data/>) leaves the reviewer unable to distinguish an excluded component from a forgotten one.

| Request component | What to inspect before counting |
| --- | --- |
| Instructions | System instructions, task wording, acceptance conditions |
| Tool definitions | Names, descriptions, and input schemas included in the request |
| Prior turns | Messages deliberately carried into this request |
| Documents | Full passages, titles, tables, and location labels |
| Tool results | Returned content already present and expected later |
| Output allowance | Required answer length, evidence references, and uncertainty fields |
| Model accounting | Selected model, thinking mode, and applicable API behavior |

Keep the counted object identical to the planned request. A count from an earlier draft cannot establish whether a later assembly fits.

Treat text added by request templates as part of the assembly. Record version identifiers for reusable instructions and tool definitions. When either changes, recount the affected request.

<a id="work-through-a-hypothetical-budget"></a>

## Work Through a Hypothetical Budget

Suppose the fictional team sets a working ceiling of thirty-two thousand tokens for this task. This is a chosen planning allowance, not the advertised capacity of any model. A real team must verify that its allowance fits the selected service configuration.

Use the following invented allocation to expose tradeoffs:

| Component | Hypothetical tokens |
| --- | --- |
| Instructions and question | 2,000 |
| Tool definitions | 1,000 |
| Relevant prior turns | 1,000 |
| Selected document evidence | 18,000 |
| Tool results already included | 2,000 |
| Output allowance | 4,000 |
| Unallocated margin | 4,000 |
| Working ceiling | 32,000 |

The planned input totals twenty-four thousand tokens. Input, output allowance, and margin together total thirty-two thousand. The margin is a deliberate planning choice, not a provider guarantee.

Now suppose the full document evidence would require twenty-eight thousand tokens. Replacing the excerpt allocation with that material makes the plan ten thousand tokens over its ceiling.

Do not solve the overage by erasing the output allowance. The requested comparison still needs space for findings and source locations. Choose a narrower evidence set, split the task, or revise the working ceiling after checking the applicable limits.

Keep the original estimate beside the revised one. A reviewer should be able to tell whether the request became smaller because irrelevant appendices were removed or because required evidence was dropped.

<a id="use-counts-as-preflight-evidence"></a>

## Use Counts as Preflight Evidence

Claude provides a [token-counting endpoint](<https://platform.claude.com/docs/en/build-with-claude/token-counting>) for supported structured inputs. Its documentation describes counting as an estimate. Count with the selected model and record the returned input count separately from your output allowance.

The [same documentation](<https://platform.claude.com/docs/en/build-with-claude/token-counting>) identifies unsupported counting inputs, including most server tools, the MCP connector, and URL- or file-sourced image/document blocks. Base64 images and [PDFs](</articles/ocr-quality-checks-business-pdfs/>) have a supported counting path. Do not silently omit an unsupported component and label the remainder a full-request count.

For the fictional worksheet, record the method as “hypothetical allocation.” Do not label the table an API measurement. For an implemented workflow, retain the count response and the request version it describes.

Use word counts or file sizes only as rough screening signals when the selected counting method is unavailable. Label that uncertainty. Avoid inventing a universal conversion from pages or words into tokens.

After a real response, compare its reported usage with the preflight estimate. Record the difference without rewriting the original estimate. Repeated discrepancies give the workflow owner a concrete reason to inspect assembly and accounting.

Recount after adding a tool result. The original request's spare room does not establish how much remains after another passage enters the conversation.

<a id="select-excerpts-with-enough-surrounding-meaning"></a>

## Select Excerpts With Enough Surrounding Meaning

In the fictional supplier review, begin with the requirement table and the response sections that discuss those requirements. Preserve table [headings](</articles/accessible-headings-b2b-pages/>), units, exception notes, and references that affect interpretation.

Suppose section three states a delivery window and its footnote excludes custom items. Copy both together. An excerpt containing only the delivery window would remove a condition the reviewer needs.

Give each excerpt a stable label: document identifier, revision label, section or page, and row range where applicable. Keep the label attached when the excerpt moves into a prompt template.

Create an omission list beside the selected passages. Record what was left out and why. “General company background omitted because the task concerns delivery requirements” is a bounded explanation. “Unimportant pages removed” cannot be checked.

Allow the assembler to stop when a passage cannot be interpreted safely on its own. A short excerpt is useful only when it preserves the meaning required for the question.

Before sending, ask whether the answer could imply review of the complete supplier response. If it could, change the wording to identify the selected sections. The response boundary should match the evidence boundary.

<a id="split-documents-around-questions"></a>

## Split Documents Around Questions

Split a request when its evidence can be reviewed in independently meaningful groups. For the fictional example, one pass might inspect delivery timing and another packaging requirements.

Give every pass the same requirement identifiers and output fields. Retain the source locations in each result. A later comparison needs to know which question each finding answers.

Do not split solely at an arbitrary character count. A table, its legend, and its exceptions may need to travel together. Treat the meaningful unit as the section a reviewer needs to interpret the requirement.

Prepare a coverage table before combining results:

| Requirement | Planned pass | Evidence location | Coverage state |
| --- | --- | --- | --- |
| R-01 delivery timing | Delivery pass | Response section three and footnote | Included |
| R-02 packaging | Packaging pass | Response section five | Included |
| R-03 custom-item handling | Delivery pass | Appendix C | Omitted pending separate review |

All identifiers and locations in this table are fictional. The table proposes a way to track coverage; it does not assert that any supplier met a requirement.

The combining pass needs its own context budget. Count its instructions, the intermediate findings, necessary source excerpts, and output allowance. Breaking [a large task into several requests](</articles/ai-agent-workflows-business/>) does not remove the final assembly problem.

Keep unresolved rows unresolved during combination. A complete answer from the delivery pass cannot establish coverage of an appendix that never entered either request.

<a id="use-summaries-as-maps-back-to-evidence"></a>

## Use Summaries as Maps Back to Evidence

A summary offers a different tradeoff. It can orient the next pass, but it also introduces another representation of the source. Treat it as a working aid with visible limits.

Require a summary to retain requirement identifiers, key conditions, source locations, and unresolved questions. Ask for an omission note when a section is excluded from the summary's scope.

Keep direct excerpts for values that control the decision. In the fictional example, preserve the delivery window and custom-item exception verbatim in the working evidence. Summarize the background discussion separately.

If a summary says “all delivery requirements are addressed,” require the next reviewer to inspect the coverage table. A broad summary sentence cannot fill an empty requirement row.

Use the three methods deliberately. Excerpts preserve selected source wording. Splitting preserves more material across separate tasks. Summaries offer orientation with a further need to verify decisive claims against their sources.

Choose based on the evidence needed for the answer. Do not choose solely by whichever method produces the smallest request. A shorter assembly that drops the decisive exception has failed the preflight review.

<a id="return-an-explicit-incomplete-result"></a>

## Return an Explicit Incomplete Result

Claude's [stop-reason documentation](<https://platform.claude.com/docs/en/build-with-claude/handling-stop-reasons>) distinguishes `max_tokens` from `model_context_window_exceeded`. The former reaches the output cap; the latter fills the context window. Treat either truncated answer as incomplete for this worksheet.

The [context-window documentation](<https://platform.claude.com/docs/en/build-with-claude/context-windows>) separates input overflow from generation truncation. Input alone exceeding the model's window returns HTTP 400 with `invalid_request_error`. For the named model, input plus `max_tokens` exceeding the window can be accepted, then stop at the window boundary.

Check the response condition before reading a polished conclusion as a finished comparison. A rejected request produces no accepted answer. Preserve the failure instead of presenting an empty findings table as evidence of no issues.

There is another incomplete state: the response finishes, but required evidence never fit. Mark that explicitly even when the API reports a normal completion.

For the fictional review, an appropriate result reads: “Delivery timing was reviewed against section three. Custom-item handling remains unresolved because Appendix C was omitted. No whole-document conclusion is available.”

Attach the requirement identifiers affected by the omission. Name the next evidence needed. The reviewer can then request a separate pass, narrow the decision, or stop the analysis.

Avoid filling missing evidence with a plausible supplier practice. The requested work is comparison against supplied documents. A general answer about how suppliers often operate would answer a different question.

<a id="keep-capacity-separate-from-acceptance"></a>

## Keep Capacity Separate From Acceptance

The [context-window documentation](<https://platform.claude.com/docs/en/build-with-claude/context-windows>) cautions that more context does not automatically improve recall or accuracy. Treat a larger available window as capacity. Continue checking the answer against the required evidence.

For the fictional review, acceptance means every required row has a supported finding or an explicit unresolved state. The reviewer can locate each cited passage. Omissions remain visible. The conclusion stays within the reviewed material.

Store that acceptance definition with the reusable request in the team's [prompt library](</articles/prompt-libraries-b2b-teams/>). Keep the context worksheet alongside it so the next user knows what must be counted and what must be preserved.

Assign one workflow owner to approve changes to the working ceiling, output contract, and document-selection method. Those choices affect what an answer can responsibly cover. Include them in routine [LLM operations](</articles/llm-ops-for-small-teams/>) review.

Before the next run, check the assembled request, the count method, the output reserve, and the omission list together. A request is ready when its evidence boundary is clear and its planned answer fits.

Keep the missing evidence visible.

<a id="sources"></a>

## Sources

- [Claude Platform: Context windows](<https://platform.claude.com/docs/en/build-with-claude/context-windows>). Request accounting and limits of larger context. Checked 2026.10.28.
- [Claude Platform: Token counting](<https://platform.claude.com/docs/en/build-with-claude/token-counting>). Preflight counts, estimates, and unsupported input categories. Checked 2026.10.28.
- [Claude Platform: Stop reasons and fallback](<https://platform.claude.com/docs/en/build-with-claude/handling-stop-reasons>). Output-cap and context-window stop conditions. Checked 2026.10.28.
