Vibeleaderboard
← Course overviewModule 04 · Give the agent the right information · Beta
Lesson 4.5

Module contents

Give the agent the right information

Lesson 5 of 5

When should an old instruction go?

Two lines of the garden file are still unsorted: a workaround written for a model that has been replaced, and a rule about examples copied from advice for a different job. Both were right for something once. This lesson gives four questions for deciding what to do with an instruction like that, draws on two vendors' current guidance about auditing instructions when a model changes, and closes the module with a plan you carry into the rest of the course and a transfer task in a different setting.

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

An instruction written for one model

Line 4 of the garden file says: always list every file in a folder before reading any, because the assistant used to miss files. That is a in the making: an instruction whose purpose was real when it was written and whose condition may no longer hold. The condition was a particular model's habit. The model has changed. The line is still loaded for every garden task, and every task pays for it with a listing operation that may no longer be needed, or that may still be, since nobody has checked.

Both vendors’ guidance supports checking instructions when models change. OpenAI's guidance page for its GPT-6 Astra model, read on 29 September 2026, strongly recommends auditing skills and other files the model can read for instructions that could influence its behavior, and warns that unclear or conflicting guidance in a skill file may cause that model to pause and block work early. The same page states its own precedence: a person's explicit instructions come before a skill's. 4 That is guidance for one named model on one date, and it will need re-reading when the model changes. Thariq Shihipar's July 2026 post makes the matching claim for Anthropic's models: the constraints removed from Claude Code's system prompt had been written for older models that needed them, and the newer models did as well without them on the company's evaluations. 9 The site's Intel collection also holds a practitioner's reading of the Astra guidance, collected from a public post whose original was not retrieved for this module; it recommends shorter skill descriptions and explicit completion criteria, and treats a new model as an occasion to clean out old instructions. Read it beside the official page as one person's interpretation of that guidance, not as a second, independent result.

None of this says that line 4 should go. It says that a line written for one model is a line whose condition has to be re-checked when the model changes, and that the checking is yours to do.

Four questions before changing a line

The audit both vendors recommend comes down to four questions about the line, and they are the same four the carry-forward artifact at the end of this lesson asks you to write down.

What was its purpose? If nobody remembers, find out before changing anything: read the older briefs or records it came from, or ask whoever added it. A line whose purpose is forgotten may encode a safety, access or project requirement that still holds, and "never contact suppliers" is such a line whatever model is in use. Where does it apply? Name the model, the task and the situation the line was written for, then ask whether that condition still holds. Line 4's condition was one model's habit, and that model is gone. What effect should the change have? Say what you expect to see differently if the line is removed or rewritten: for line 4, fewer listing operations, with the same files still read. How will you compare? Decide what you will run with and without the line, on the model you now use, and what you will look at. For line 4 that is the quote task both ways, checking the for which files were read.

The fourth question is where the vendor reports stop and your work starts. Anthropic measured its own product on its own evaluations; OpenAI describes what its guidance recommends. Neither has measured the garden file, and a claim that some instruction helps a particular model becomes something to rely on only after it has been observed on the kind of task you actually run. Two runs of one task is a small comparison, and a small comparison beats none: it turns a belief about an old model into an observation about the current one, for those runs. It does not show that the line causes the difference in general; if the two runs differ in ways that could come from ordinary variation between runs, repeat the comparison or vary the task before drawing a conclusion. That is a judgment about how much the change matters, not a quota, and not every small edit needs an experiment.

The examples rule, checked against a different job

Line 8 says to include three worked examples in anything written for the committee. Run it through the four questions and the first one already shows a problem. The line came from advice about prompting: Anthropic's September 2025 account recommends that a developer writing instructions for a model include a few varied, representative examples of the behavior wanted, rather than trying to list every case. 8 That advice is about what goes into the model's context to show it a format. Line 8 applies it to what the model writes for people to read. The committee did not ask for worked examples in a price comparison, so the line's purpose, as copied, does not match the job it is now attached to.

The second question is where it gets interesting, because the answer is not "delete". Examples in a prompt may still help the current model when a task's format is unusual, and no one has checked that either. The July 2026 post discusses examples in tool instructions: its author reports that these examples constrained exploration and describes replacing them with a clearer tool interface. That observation concerns a particular way of describing tools to particular models. It does not test whether showing the committee’s desired table layout helps with this writing task. Treating it as a reason to remove every example would repeat line 8’s mistake: applying advice without checking the job it addresses. The expected effect and the comparison follow from the purpose you settle on. If the purpose is to show the model the committee's table layout, the line becomes one sample row in the skill's instructions, and the comparison is whether the layout comes out right with and without it. If the purpose was never the committee's, the line goes, and the comparison is a committee member reading the next report. Either way the rule is not "always add examples" and not "never"; it is a line with a stated purpose, a place it applies, and a way to find out whether it is doing anything.

The information plan

Everything in this module reduces to a short document written before a task, which this course calls an : for each thing the task needs, where it will be kept so that it is present when needed, is current, and survives a handoff; and, for any instruction you intend to change, its purpose, where it applies, the effect you expect and how you will compare. Here is the authored plan for the next fencing task.

Task facts (in the brief): the committee needs the table by Wednesday evening; installed prices
  with delivery listed separately; Parkside's quote still outstanding.
Evidence: the two retrieved quote files in Documents/Garden/Suppliers; the Parkside URL
  in the brief remains an outstanding source until its quote is obtained.
Conventions (garden instruction file): folder locations; installed-price rule; no contact with
  suppliers; ask before changing an existing file.
Procedure (skill review-supplier-quote): the four review steps.
Must survive any handoff: the Wednesday deadline and the Committee folder. Both are in the brief,
  which is kept as a file the harness reads at the start.

Proposed change: remove line 4 (list every file before reading).
  Purpose: an earlier model skipped files. Applies to: that model; not checked on the current one.
  Expected effect: fewer listing operations; every quote still read.
  Comparison: run the quote task with and without the line on the current model; compare the
  action records for files read and operations used.

The plan is short because the work of the module was sorting. It carries forward into module 5, where the question changes from what the agent is told to what it can do, and into module 8, where the survival questions apply to workflows that run longer than one session.

Transfer checkpoint: a school newsletter

Try the module's distinctions on a case that is not the garden. Everything here is authored.

A parent volunteer produces a monthly newsletter for a primary school with an assistant running in a harness. Drafting happens in one session; a second session, started the next day, lays the draft out and saves the file for the office. The volunteer keeps this instruction file in the newsletter folder, and the harness loads it for every newsletter task:

NEWSLETTER NOTES
1  Drafts go in Newsletter/Drafts; the final file goes in Newsletter/ToOffice as newsletter-YYYY-MM.pdf.
2  Never include a child's surname.
3  The October issue must mention the harvest assembly on the 17th.
4  Always write three sample headlines before each article, because the old assistant's first
   headline was always dull.
5  The office wants the file by the last Friday of the month.
6  To check an event date: compare the date in the draft with the school calendar file; if they
   differ, consult the head's email; flag any date you cannot confirm.

In the drafting session the volunteer also says, in conversation: "This month the office wants it a day early, Thursday, because Friday is an inset day." The drafting session is compacted before the draft is finished, and the layout session starts the next day. Write your own answers to four questions before reading the guide. First, sort the six lines by the module's four kinds and say which belong in this file. Second, which facts must survive the handoff to the layout session, and why each one? Third, take line 4 through the four audit questions, without deciding in advance that it should go. Fourth, which line is a request the harness does not enforce, and what could check it, with what limit?

Self-review guide. A complete answer sorts lines 1, 2 and 5 as conventions that belong in the file, noting that line 5 is a standing default with this month as an exception; line 3 as a task fact for one issue, true of October and noise in November, better kept in that issue's brief; line 6 as a procedure that could be a skill, loaded when dates are being checked; and line 4 as an instruction written for an earlier model. For the handoff, it names this month's Thursday deadline as the fact most at risk: the layout session depends on it, line 5 in the file says Friday, and the exception was said only in the drafting session's conversation, which the layout session's context does not contain; so it moves the Thursday exception into the brief or a note the layout session reads; it recognizes that the folder and file-name rules will return through the instruction file, and that the draft itself is evidence the layout session can read from disk. For line 4, it states the purpose, a dull first headline from a previous model; the applicability, that model, unchecked on the current one; the expected effect of removing it, shorter drafts with headlines still acceptable; and a comparison, two articles drafted with and without the line on the current model and read by the volunteer, which is what turns the old observation into a current one. A strong answer adds that the surname rule on line 2 is a request to the model and that no filename makes it enforced. One concrete check is a release step before anything goes to the office folder: a person reads the draft for names, or a program compares the text against the class list and blocks the save on a match. Either has a limit worth stating. A person can miss a name, a list comparison misses a surname that is not on the list or is spelled differently, and a check attached to one folder does nothing about a file saved somewhere else unless writes elsewhere are denied. An answer that names the four kinds without saying why each line belongs where it does is not yet complete.

Carry the module into the next

Instructions describe what to do. They do not set what can be attempted or executed. The fencing skill can describe reading a quote in four careful steps, and the read happens only because a tool exists for it and the harness's rules allow it, returning a result in a form the model can use. The capabilities a harness offers and the permissions it enforces set what can be attempted and what runs; the model, guided by its instructions, chooses among them. The next module is about those capabilities: how a harness is connected to the operations it offers, how results come back, and what changes when you add a tool of your own.

Carry this forward

  • An instruction written for one model has a condition that must be re-checked when the model changes. The cited vendors' guidance says to audit instruction files and skills when a model changes; the checking is yours to do.
  • Before changing a line, state its purpose, where it applies, the effect you expect and how you will compare. A forgotten purpose is a reason to investigate first, since the line may encode a requirement that still holds.
  • A tip copied from one job to another can be right for neither. Vendor results are evidence about their products on their evaluations; a small comparison on your own task and model makes an instruction's effect an observation for those runs, to be repeated or varied when outcomes differ for other reasons.
  • An information plan records where each thing a task needs will be kept and what must survive a handoff, together with any proposed instruction change and its comparison. Instructions describe what to do; capabilities and enforced permissions set what can be attempted and executed, and the model chooses among them.

Lesson check

The organizer reads that a vendor removed most of its product's system prompt for a new model and decides to delete line 4, the list-before-reading workaround, and line 9, "never contact suppliers", from the garden file in the same pass. Which statement follows from 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.

You have reached the end of module 4. The module quiz below checks how these ideas fit together before module 5 turns from what the agent is told to what it can attempt and execute.

Module 4 quiz: put the ideas together

1. "The fencing comparison is needed for the committee meeting on Thursday" sits in a file loaded for every garden task. Six months later the assistant, asked to compare pond pump quotes, mentions a Thursday deadline. What does the module say about that line?
2. The organizer's brief omits that the committee wants delivery listed separately. The organizer switches to a more capable model and gives it the same brief and the same tools. What does the lesson say can and cannot change?
3. A colleague cites the report that Anthropic removed over 80 percent of a product's system prompt with no measured loss and proposes cutting the garden file by the same fraction. Which response matches the module?
4. Lines 6 and 7 of the garden file contradict each other about asking before writes. The organizer adds a line: "Where these notes disagree, the later line wins." What has changed?
5. An authored harness loads the project instruction file at the start of a run and does not reload it automatically. The organizer later adds “Never contact suppliers” to that file. The agent has not read it again and now proposes an email. What should the organizer check?
6. A harness lists a skill named send-committee-report whose steps end with emailing the finished file. The harness has no email tool, no tool that runs programs and no network access, so no available capability or account can send. The model selects the skill and follows the steps. What happens at the last step?
7. A session has spent most of its context on a failed attempt to fetch the Parkside quote, but it also holds two facts the organizer said only in conversation: the Committee folder and the Wednesday deadline. The organizer wants to start the layout work. Which choice matches the lesson's method?
8. A friend reads a 2025 recommendation to include a few representative examples in prompts and adds "always include three examples" to every instruction file they keep. Which statement follows from the 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.

Sources · Reviewed 2026-09-29
  1. 4. OpenAI: GPT-6 Astra guidance
  2. 9. Thariq Shihipar (claude.dev): The new rules of context engineering for Claude 5 generation models
  3. 5. VibeLeaderboard Intel: Auditing Skills and AGENTS.md for GPT-6 Astra
  4. 8. Anthropic: Effective context engineering for AI agents (September 2025)

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.