Briefing an agent
A six-part brief that tells an agent the goal, the inputs, the output, the limits and how to check the work.
You can write a six-part brief and ask for a plan before Codex does the work.
You can write a six-part brief, ask Codex for a plan with no changes, correct the plan, and then ask it to prove the work meets the brief.
You can specify a task with an explicit goal, inputs, output, constraints and acceptance criteria, and use a plan-then-act loop with executable checks as the oracle.
Why a brief matters
An agent does what you say. When you leave a gap, it fills the gap with a guess.
Think of a new colleague on day one. If you say only "sort out the suppliers", they must guess what you mean. A clear note tells them what you want. They do not need to guess.
A brief is that clear note. It has six short parts.
An agent does what you wrote and fills every gap with a guess. A vague request such as "review the quotes" can come back as a summary, a ranking, or an email draft.
A brief closes the gaps. It states what you want, what the agent may use, what the result looks like, and how you will both know it is right.
A brief takes a short time to write. Redoing an answer that guessed wrong takes longer.
A natural-language request under-specifies the task. The model resolves the ambiguity by sampling, so identical vague prompts yield different interpretations. Output variance is partly an ambiguity problem, and explicit constraints narrow the distribution of acceptable outputs.
Instruction-tuned models are trained to follow stated instructions (Ouyang et al., 2022, "Training language models to follow instructions with human feedback"). A stated instruction is therefore the main lever you control. An unstated one is a guess.
The brief below is a checklist for a specification: an objective, an input set, an output format, constraints, a termination condition and a verification method.
| Part | Question it answers | Example from procurement |
|---|---|---|
| Goal | What do you want, and for whom? | Compare three laptop quotes for a finance manager |
| Inputs | Which files and facts may Codex use? | The RFQ and the three quotes in the procurement folder |
| Output | What should the result look like, and where does it go? | A CSV of scores and a one-page recommendation in outputs/ |
| Constraints | What must Codex not do? | Do not change the source files. Use only facts in the files |
| Done when | How will you know the work is finished? | Every quote has a score, and the method is shown |
| How to check | What evidence will Codex show you? | The scoring table with the page or line for each figure |
Goal: ... Inputs: ... Output: ... Constraints: ... Done when: ... How to check: ...
Copy this into the message box. Replace each dot line with your own words.
Copy the template and replace each line. A short answer per line is enough. Keep the six labels so you can see at a glance what is missing.
Keep the labels. They make omissions visible and give the model a stable structure to attend to. Order matters less than completeness.
Plan first, then act
Ask for a plan with no changes
Send your brief. Add one line at the end: "Show me a numbered plan first. Do not change any files yet."
Codex replies with a plan and the questions it has.
New thread
- Compare the laptop quotes
Brief: compare the three laptop quotes using the RFQ criteria. Show me a numbered plan first. Do not change any files yet.
Plan: 1. Read the RFQ and list the criteria and weights. 2. Read each quote. 3. Score each quote and show the method. 4. Write outputs/quote-scores.csv. I have not changed any files.
Codex shows a plan and waits. No files have changed. Correct the plan
Read each step. Fix anything wrong, for example a missing input or a wrong output folder.
Reply with the correction and ask for the updated plan.
Approve the plan and let Codex act
When the plan is right, reply: "Go ahead with this plan."
Codex still asks for approval before each file change. Read each card.
Ask it to check its work
Ask Codex to check its output against your "done when" line and to show evidence.
Look at the evidence yourself. Do not accept "all done" alone.
Check your work against the Done when line. Show the evidence for each figure, with the file and the page or line. List any assumption you made.
Codex shows its proof. You read the proof.
Evidence means a file name and a page or line you can open. If Codex cannot point to one, treat the figure as unconfirmed.
A model that checks its own output shares the blind spots that produced it. Prefer evidence you can verify independently: a script you can re-run, totals that reconcile, a cited line you can open.
Key terms
- Brief
A short note with six parts that tells an agent what to do.
A six-part task description.
A task specification: goal, inputs, output, constraints, termination condition and verification method.
- Plan
A list of steps the agent will take, shown before it starts.
The steps Codex proposes before it acts, which you can correct.
An intermediate decomposition of the task, produced before any action so that a human can review it.
- Acceptance criteria
The tests that tell you the work is done and right.
The conditions the result must meet before you accept it. They are the Done when line of the brief.
Verifiable conditions that define completion. They work best when a script or a reconciliation can test them.
Check yourself
Name the six parts of a brief.
Goal, inputs, output, constraints, done when, and how to check.
Goal, Inputs, Output, Constraints, Done when and How to check.
Goal, Inputs, Output, Constraints, Done when, How to check: an objective, an input set, a format, limits, a stop condition and a verification method.
Why ask for a plan before Codex changes any file?
You can fix a wrong plan in seconds. A wrong result takes longer to fix.
A wrong plan is cheap to correct. A wrong set of changes costs review time and may need undoing.
Reviewing a plan exposes misreadings before they are encoded as actions. It moves human judgement to the cheapest point in the loop.
Codex says the task is complete. What do you ask next?
Ask it to show proof, such as the file and line for each figure.
Ask it to check the work against the Done when line and show evidence you can open.
Ask for evidence tied to the acceptance criteria. A claim of completion is not a verification, and agents can report success when they have not met the goal.
OtterBlock is not affiliated with OpenAI. Codex and ChatGPT are OpenAI products.