What information reaches the model?
A language model generates an answer from the information supplied to it and patterns learned during training. Context is the material available for a particular request: your instructions, messages and any source material included by the application. A file sitting on your laptop is not automatically part of that context. An assistant needs the file attached, pasted or read through a tool before its contents can inform the answer.
A tool is a function the surrounding application can run, such as reading a document or searching the web. The application adds the returned text to the model’s context. If a search returns a policy page, the model can use that page in its next answer. If the tool returns “file not found,” the context contains an error rather than the policy. A plausible answer after that error still lacks the missing evidence.
Training adjusts the model’s parameters, numbers that encode patterns learned from training data. Supplying a document gives the current request information without establishing that the document became permanent training data. A product can separately retain chat history or save memory for later requests. Check its settings and data policy when deciding what material to share.
The model processes text in tokens, units that can contain a word, part of a word or punctuation. Token counts depend on the tokenizer and language. Generated text can be fluent even when its claims are wrong, because producing a plausible continuation and checking a fact are different operations. For a claim about a particular policy, compare the answer with the applicable policy text.
Answer a question from a supplied policy
Suppose a teammate asks whether a member can currently create an export. Include the applicable rule so the assistant has evidence about this workspace. In this fictional example, the supplied policy is current:
Can a member create an export in a Team workspace?
Use the supplied current policy.
Team export policy, effective 29 October 2026:
Only workspace owners can create exports.The answer follows directly from the supplied role restriction: “A member cannot create an export; creation requires an owner.” If the policy is absent and a reading tool returns “file not found,” the model has no current rule to use. Recover the source or report that the answer remains unverified. A confident recollection cannot supply the missing policy.
Create a study explanation from selected material
Context selection also helps you create learning material. Suppose your notes contain the rule “a linear scan checks each of n items once.” Asking for a broad explanation of algorithms could introduce concepts you have not studied. Supply the rule and ask for a small explanation built around it:
I am learning linear scans. Use this rule from my notes:
"A linear scan checks each of n items once."
Show how to find the largest value in [3, 8, 2, 6].
Explain what is remembered after each item.
Use ordinary language before introducing notation.
End with one question I can answer without seeing the example.A useful explanation can trace the remembered largest value through the input. Starting with 3, reading 8 updates the largest to 8. Reading 2 and then 6 leaves it at 8. Remembering the largest value seen so far is enough because a smaller earlier value cannot become the largest after another item is read.
Read 3 -> largest so far is 3
Read 8 -> largest so far is 8
Read 2 -> largest so far is 8
Read 6 -> largest so far is 8
Answer: 8
Recall question: What should the remembered value mean
after each step of a scan for the smallest item?If the first answer uses “loop invariant” without explaining it, ask: “Explain that term using the value remembered after each item in this example.” An invariant is a statement that remains true at the specified point in every iteration. Here it says that the remembered value is the largest of the items already read. You can then adapt it to a minimum scan by remembering the smallest value seen so far.
Compare the generated trace with the four supplied numbers, then answer the recall question yourself. For another topic, replace the source rule and the input together. Giving an assistant a narrow source and an example helps it produce an explanation you can inspect and use, while your own follow-up answer checks what you understood.
Choose context within a token budget
A context window is the model’s limit on how many tokens it can work with in one request. Input consumes capacity, and the answer needs capacity too. Some models also use reasoning tokens within their limits. Check the current model documentation and actual usage fields; a product’s file-size limit and a model’s token limit can differ.
Choose material by asking what uncertainty it resolves. For an export authorization bug, the requirement tells the assistant who should be allowed, the implementation shows the current rule, and the failing test shows an observed mismatch. Unrelated old logs answer none of those questions. Leaving them out protects the limited space available for relevant evidence.
The lab uses fixed teaching counts rather than a real tokenizer. Its capacity is 8,000 tokens, with at least 1,000 reserved for the answer. Select the four relevant inputs, then add their sizes:
Instructions 600
Current task 400
Relevant implementation 2,800
Failing test and output 1,200
-----
Input total 5,000
Answer reserve 1,000
-----
Planned total 6,000 / 8,000
Remaining capacity 2,000The 4,500-token unrelated log would raise the planned total to 10,500, beyond this exercise’s capacity. A larger window could fit it, but its relevance would still need a reason. When useful evidence itself is too large, read the sections connected to the question, keep source locations and retrieve more when the next uncertainty requires it. A summary can omit an exception, so return to the original passage before making a claim that depends on that exception.
Continue a long task from a working record
A long conversation can outgrow its available context or continue in a new session. Write a small working record while the evidence is available. It should preserve enough information to recover the task, including the distinction between a proposed edit and a checked result.
Goal: only workspace owners may create exports.
Source: docs/team-export-policy.md, effective 29 October 2026.
Observed: owner export succeeds; member export also succeeds incorrectly.
Files: exports.js and exports.test.js.
Change so far: none.
Next: reproduce the member case and inspect its authorization check.
Release scope: local patch only.This record tells the next session where to look and which question remains open. “Exports fixed” would erase the observed defect and imply a result that has not occurred. After an edit, replace “Change so far” with the actual change and record the command and outcome of the check. Keep the original files and logs available because a continuation record is a selected account of them.
In the lab, you choose context and an output reserve using the fixed budget. The assessment changes the task to a current-policy answer, where an old policy archive is irrelevant to the requested current rule. Apply the relevance question again instead of copying the same selected boxes.
Try the skill
Fixed local fixture. No model or external tool is called.
Build context for a local code fix using its implementation and failing test. The simplified capacity is 8,000 tokens. Reserve at least 1,000 output tokens. Include exactly the relevant material.