Vibeleaderboard
← Course overviewModule 04 · Give the agent the right information · Beta
Lesson 4.2

Module contents

Give the agent the right information

Lesson 2 of 5

Why does a fact in the context still get missed?

The previous lesson dealt with information that never reached the model. This one deals with information that did. A fact can be in the context window and still not shape the reply, and two instructions can both be present and disagree. The lesson uses an authored instruction file that has grown for a year to show both problems, then draws on two vendor reports, a year apart, about what happens when instructions are cut back.

Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.

Present is not the same as used

Module 1 described the as bounded. A larger window can hold more of the garden archive, but capacity and correct use are different questions. The quotes might be present beside old prices, an obsolete brief and the current request. The model still has to identify which information applies.

How much a given model loses as its context fills, and on which tasks, is a measured property of that model, and this course has not measured it. Nor is attention a fixed pool that every added line drains by the same amount: the learned values module 1 described, together with the instructions in the window, do weight some material more than the rest, and a clear request can direct the model to the quotes even in a long context. What the course can say follows from module 1: the model generates from the whole window, weighting is learned and not guaranteed, and so evidence that is present can still be overlooked, or read alongside something misleading. A larger window raises how much can be present. It does not by itself settle whether any particular fact will be used well, and Anthropic's recommendation to keep context relevant is one vendor's answer to that, for the models and tasks it works with.

An instruction file that has grown for a year

Here is an authored file the organizer has kept for the garden project. Suppose the harness loads it at the start of every garden task. Each line was added for a reason at the time.

GARDEN PROJECT NOTES FOR THE ASSISTANT
 1  The Garden folder is Documents/Garden. Anything for the committee goes in Garden/Committee.
 2  Compare supplier prices installed, including delivery, not the headline figure.
 3  The fencing comparison is needed for the committee meeting on Thursday.
 4  Always list every file in a folder before reading any, because the assistant used to miss files.
 5  To review a quote: check the length or quantity against the brief, then the price basis, then
    delivery time, then guarantee. Flag anything the quote leaves out.
 6  Always ask before writing any file.
 7  Write results straight to the folder without asking; I don't want to be interrupted.
 8  Always include three worked examples when writing anything for the committee.
 9  Never contact suppliers.
10  The pond pump quotes are in Garden/Suppliers/Pump.

Before reading on, sort each line into the previous lesson's four kinds, and mark any line you would not want loaded for a task about pruning the apple trees.

Lines 1, 2 and 9 are conventions and belong here, though line 9 is an in module 3's first layer, and the harness's rules still decide whether an email tool exists. Lines 3 and 10 are task facts from earlier tasks. Line 3 was true of one meeting last spring, and as written it carries no date and no note of which meeting, so a model reading it during a fencing task in a different month has nothing to tell it that the Thursday has passed. A dated line would at least give the model that chance. Line 10 is harmless until the pump quotes move. Line 5 is a procedure, useful for quotes and noise for pruning. Line 4 is a workaround written for a model that has since been replaced, which lesson 5 returns to. Line 8 is a stylistic rule copied from advice somewhere and applied to everything. And lines 6 and 7 contradict each other: both are present for every task, and the file gives the model no way to choose.

When two instructions disagree

Lines 6 and 7 are an : two instructions in the same context that cannot both be followed. The documentation for Claude Code, read on 29 September 2026, states the consequence plainly for that product's own instruction files: if two rules contradict each other, the model may pick one arbitrarily, and it advises reviewing the files periodically to remove outdated or conflicting instructions. 3 A July 2026 post by Thariq Shihipar, a member of Anthropic's technical staff who works on that product, gives an example from its own system prompt: one passage asked the model to leave documentation as appropriate, another said not to add comments, and the model spent effort working out which to obey before it could act. 9

Notice what a conflict is not. In this example the harness has no rule that resolves the two sentences; they are instructions in the model’s context. There are three ways to resolve it. An explicit precedence line can work: "where these notes disagree, the later line wins" gives the model a rule to apply, and some harnesses build such an order in, and the product’s higher-priority instructions still apply. Making the two lines apply to different situations so that they no longer meet is often easier to maintain: ask before overwriting an existing file, write new files freely. And deleting the line that no longer says what you want is simplest of all, once you know which line that is. When the organizer cannot remember why line 6 was added, that is a reason to find out before touching it, because a line like that may encode a safety or access requirement that still holds. Lesson 5 gives the questions to ask.

Less, chosen for a reason

The same July 2026 post reports what happened when Anthropic cut instructions rather than adding them. Its author writes that the team removed over 80 percent of Claude Code's system prompt for models such as Claude Opus 5 and Claude Fable 5 with no measurable loss on the company's coding evaluations, and attributes the removed material to constraints written for older models that had needed them. 9 Read the scope with the claim. It is one vendor's measurement of its own product on its own evaluations, for the models named in that July report; the post itself says older models required stricter constraints. It is not a percentage to aim for in the garden file, and cutting a line whose purpose you have not identified is not what the post describes.

Put the two reports side by side and the direction of the question transfers, not a rule. The 2025 account advised curating a few varied, representative examples rather than listing every edge case, and asking of each addition whether it makes the wanted outcome more likely. 8 The 2026 post applies a similar test to accumulated instructions and finds, for the models it names, that much less was needed than had been written. Neither says that shorter is always better; a fact the task needs is not improved by leaving it out. For the garden file the test is that each of the ten lines has to earn its place with a purpose you can state and a situation it applies to. Lines 1, 2 and 9 pass. Lines 3 and 10 fail because their situation has passed or may have. Line 5 passes for quotes and fails for everything else, which is a placement problem, and the next lesson gives it a place.

Carry this forward

  • A fact in the context is not thereby used well. Learned weighting and instructions can direct the model to relevant material, but not reliably, and how much a model loses in a long context depends on the model and the task. A larger window raises what can be present; one vendor's advice, for its models, is to keep context relevant.
  • An instruction file accumulates task facts, workarounds and copied advice unless each line is sorted by purpose. A task fact from an earlier task is a misleading cue for tasks it does not concern, and one with no date or scope gives the model nothing to recognize it by.
  • Two instructions that cannot both be followed are resolved by the model somehow, and one product's documentation says arbitrarily. A stated precedence can resolve it; scoping the two lines to different situations, or removing the obsolete one, is often easier to keep right. Find out why a line was added before changing it.
  • One vendor reports removing most of a product's system prompt for its newest models with no measured loss on its own evaluations. What transfers is the test, not the percentage: each instruction needs a stated purpose and a situation it applies to, and a needed fact is never improved by omission.

Lesson check

The organizer switches to a model whose context window holds ten times as much and pastes the entire garden archive, including the three quotes, into the request. Which statement matches this lesson?

Choose an answer for each question.

This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.

The garden file needs sorting, and two of its lines need somewhere else to live. The next lesson looks at the places real harnesses provide for conventions and procedures, and at why a file's name alone does nothing.

Sources · Reviewed 2026-09-29
  1. 8. Anthropic: Effective context engineering for AI agents (September 2025)
  2. 3. Claude Code documentation: How Claude remembers your project
  3. 9. Thariq Shihipar (claude.dev): The new rules of context engineering for Claude 5 generation models

Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.