Skip to content
AI Training

Changing the brief while the agent works

In this lesson we give an agent a brief with a gap in it and watch what the agent does with the gap. Then we answer the agent’s question in two places, first in the chat and then in the brief. A second run after each answer shows which one survives. After that we change a limit in the brief between two runs, and we stop a run whose first result shows the goal itself was wrong. The task is small, and you can run it as often as you like.

You need an agent that can read a folder of files, such as Claude Cowork or a chat tool with file access. If your tool reads only what you paste, paste the four files below into the chat with their names above them. Every step of the lesson then works the same way. Every exchange shown here is illustrative. The author wrote it to show the pattern, and your agent gives a different text.

The task and its gap

Section titled “The task and its gap”

Make a folder with these files. They belong to Brindlewick Instruments, the fictional company from the earlier lessons, and the task is the calibration team’s weekly status update for the operations director.

notes.md:

Week 38, calibration team
- Finished the recalibration of the 20 torque wrenches for the Delft site (Jonas)
- Started the thermocouple batch for the food lab, 12 of 40 done (Mira)
- Waiting on the 1 kg reference weight from the supplier, due 26 Sep (Mira)
- Storeroom B stock count done, two calipers over, three probes short (Sam)
- Next week: finish the thermocouple batch, start the spirit levels for the survey team

templates/short-update.md:

Subject: Calibration team, week <number>
One paragraph of at most 120 words: what was finished, what is in
progress, what is blocked and by whom, and what comes next week.

templates/full-update.md:

Subject: Calibration team, week <number>
## Done
## In progress
## Blocked
## Next week
One bullet per item under each heading. Keep each bullet to one line.

brief.md:

Write this week's status update for the operations director. Take the
items from notes.md and use the template in templates/. Use only the
notes as your source, and do not add work that is not in them. Done when
every item in the notes appears exactly once, each under the heading that
fits it, and the update follows the template.

The brief has the four parts from the lesson on delegating: a goal, the context to use, limits, and done-criteria. It also has a gap, and you can see it before the agent does. The folder holds two templates and the brief says “the template”. Predict what happens when you send it. Then open a fresh session in your agent, point it at the folder, and send this.

Prompt (illustrative, not a recorded transcript)

Do the task in brief.md.

Response

Before I write the update: templates/ holds two files, short-update.md (one paragraph, at most 120 words) and full-update.md (four headings with bullets). The brief says “the template” and doesn’t say which. Which one should I use?

Some agents ask, as this one does. Others pick one, write the update and mention the pick at the end, or don’t mention it at all. Either way the gap is the same gap. When the agent asks, the gap is now your question. When the agent guesses, the gap is filled with a default you didn’t choose. Both tell you something the brief left out, and that is the thing you have to add.

Answer in the chat or in the brief

Section titled “Answer in the chat or in the brief”

You have the answer. You want the full update, with the four headings. The question is where the answer goes. Try both places, one after the other.

First, answer in the chat. Type the answer as a reply and let the agent finish.

Prompt (illustrative, not a recorded transcript)

Use full-update.md.

Response
Subject: Calibration team, week 38


## Done
- Finished the recalibration of the 20 torque wrenches for the Delft site
- Storeroom B stock count done, two calipers over, three probes short


## In progress
- Thermocouple batch for the food lab, 12 of 40 done


## Blocked
- 1 kg reference weight from the supplier, due 26 Sep


## Next week
- Finish the thermocouple batch
- Start the spirit levels for the survey team

Check it against the done-criteria. The notes hold five lines and the update has six bullets, because the last line of the notes holds two items. Every item appears once, under a heading that fits, and the update follows the template. The names in parentheses are gone, and the template’s one-line limit is a fair reason for that. The result passes. Now close this session, open a fresh one on the same folder, and send “Do the task in brief.md” again. The agent asks the same question, or picks a template again, and the pick is the agent’s default and not your answer. The answer you typed was part of the chat, and a new session starts from the files. The chat is over, and nothing in the folder says which template you meant. The same thing happens next week when a colleague runs the brief, or when you run it in a session that doesn’t have this conversation.

Now put the answer in the second place. Edit brief.md and change “use the template in templates/” to “use the template in templates/full-update.md”. Open a fresh session, send the same one line, and this time there is no question. The agent writes the update, and it does so every time anyone runs this brief from now on.

This is the same fact you met when iterating on a prompt: a correction holds only where the next run reads it [1]. In a chat, the next run reads the prompt. In a delegated task, the next run reads the brief and the files next to it. A reply in the chat fixes this run. An edit to the brief fixes every run.

Checkpoint · choice

Where does the answer go?

Section titled “Where does the answer go?”

The agent asks which of two templates to use. You know the answer, and the short template is used for another weekly message. Where does the answer belong?

Changing a limit between runs

Section titled “Changing a limit between runs”

The full template says “keep each bullet to one line”. The first update came back without the names, and the director wants to know who is responsible for each item. The notes have the names, and a one-line bullet barely fits them. You want the bullets to name the person and be allowed a second line. This is a change to a limit, and how you state it matters.

Type “add who is responsible to each bullet” and the agent now has two instructions: the template’s one-line limit and your new one. It tries to satisfy both. You get a bullet that makes room for the name by dropping another fact, or a name squeezed in as initials. Nothing told it the earlier limit no longer applies, so it holds on to it. State the change as a change instead, in the chat for this run and in the file for every run after it.

Prompt (illustrative, not a recorded transcript)

Replace the limit “keep each bullet to one line” with this one: each bullet names the person responsible and may run to two lines. Rewrite the update with that change and nothing else changed.

Response
Subject: Calibration team, week 38


## Done
- Finished the recalibration of the 20 torque wrenches for the Delft site.
  Responsible: Jonas.
- Storeroom B stock count done, two calipers over, three probes short.
  Responsible: Sam.


## In progress
- Thermocouple batch for the food lab, 12 of 40 done. Responsible: Mira.


## Blocked
- 1 kg reference weight from the supplier, due 26 Sep. Responsible: Mira.


## Next week
- Finish the thermocouple batch. Responsible: not in the notes.
- Start the spirit levels for the survey team. Responsible: not in the notes.

The exchange above is illustrative. In your own result, look for a name on every bullet and no bullet cut short to make room. Where the notes give no name, as on both next-week lines, a plain “not in the notes” is right and a guessed name is a failure, even where the guess is easy. Then make the same change in templates/full-update.md. The next run gets the new limit without being told. The word “replace” did the work here. It names the old instruction and retires it, where “also” or “add” would have left it standing.

Checkpoint · repair

State the change as a change

Section titled “State the change as a change”

The template the agent has just used says “keep each bullet to one line”. A colleague sends the message below and gets bullets with initials squeezed onto the end of each line. Rewrite the message so the agent drops the old limit.

When the first result shows the goal was wrong

Section titled “When the first result shows the goal was wrong”

Sometimes the first result tells you something worse than a missing name. It tells you the goal was wrong. Suppose the update comes back with the torque wrenches under “Done”, and that line reminds you of what the director said last week: she signed those off in a meeting and wants the update to cover the decisions she has to take. A list of what happened was never the document she asked for. The update the brief asks for is the wrong document. Every criterion in the brief can pass, and the result is still no use.

Stopping is cheap because the work so far was the agent’s, and the agent starts again from the brief without complaint. The expensive part is the review you would give a result that can pass every criterion and still miss the point. So a first result gets checked against the done-criteria and then against the reason you wrote them. Does meeting them still get you what you wanted? When that answer is no, the brief has the wrong goal, and the edit goes there.

Checkpoint · scenario

Stop or finish?

Section titled “Stop or finish?”

The agent is halfway through the update, and the “Done” section shows you the director wanted the decisions she has to take, and the brief asks for a list of work done. What do you do?

The loop

Section titled “The loop”

Every step in this lesson was the same move. A result came back, you read it against the brief, and what you learned went into the brief before the next run. The agent’s question became a line in the brief, the changed limit became a replacement in the template, and the wrong goal became a rewrite of the done-criteria before the run went on. A result comes back, the brief changes, and the run goes again [2]. The chat is where you steer this run, and the files are where the steering is kept.

Exercise

Make the folder from this lesson and run brief.md as written, so the agent has to choose between the two templates. Answer its question once in the chat and once by editing the brief, and after each answer start a fresh session and run the brief again. Write down, in two or three lines, what the second run did in each case. It takes ten minutes and shows you which of your answers the next run can see, which is what protects next week’s run from this week’s question. A good result is two lines that say “asked again” for the chat answer and “no question” for the edited brief. If your agent picked a template instead of asking, its pick is the question, and you answer it the same two ways. Which other answers have you given an agent in the chat this week that belong in a file?

Stretch: Run the original brief with 'the template' left ambiguous in three fresh sessions and note whether the agent asks or picks each time, and which template it picks.

Recap

  1. An agent that asks a question and an agent that guesses are showing you the same gap in the brief.
  2. An answer in the chat fixes this run. An answer in the brief fixes every later run of this brief, because a new session starts from the files and the earlier chat is not among them [1].
  3. State a changed limit as a replacement, and name the old limit. An agent holds the earlier instruction and tries to satisfy both.
  4. When a first result shows the goal itself was wrong, stop the run and rewrite the goal and the done-criteria. A stopped run costs one run, and a finished wrong result costs the run, the review, and the next run.
  5. A result comes back, the brief changes, and the run goes again. Steer this run in the chat, and keep the steering in the files [2].

You can now

  • Adjusts the brief when the work reveals new information

References

Section titled “References”
  1. Anthropic. AI Fluency: Framework and foundations. Claude Academy. Course. Academy ai-fluency-framework-foundations
  2. Anthropic. Introduction to Claude Cowork. Claude Academy. Course. Academy introduction-to-claude-cowork