Artificial intelligence works from the information available to it. Focused jobs, useful instructions, and saved records help it keep the important things in view. That is why each A.N. Creation bot has its own job, and why a bot hands a contained job to a temporary helper (the analyst) on a written brief.
Understand memory · The Christmas-card example · Why smaller files help
Working memory (context window) and saved records are different
Imagine the information on your desk and the records in your filing cabinet. You can own a large cabinet, but you cannot work with every page spread out in front of you at once.
A bot can work with only so much information at a time (its context window). The material available for one response (its context) is its working memory for that moment. That can include instructions, parts of the conversation, retrieved memories, files, and tool results. A stored chat or file is not necessarily all in view when the assistant answers.
Some products keep memories or summaries between sessions. That helps continuity, but it does not mean every earlier detail will be recalled correctly. For an important decision, have the assistant read the current project record.
Grok’s documentation describes retained role context and shared files. OpenAI’s explanation of conversation state distinguishes saved conversations from the information used for a particular response. The useful habit is the same: save what matters and point the next task at it.
Why not ask one bot to do everything?
One capable assistant can help with many subjects. The problem grows when one conversation accumulates every draft, calculation, supplier search, budget, and side question. Relevant instructions compete with a lot of detail that belongs to another job.
Think of a company’s departments. The product side concentrates on the product and the person using it; the business side on the company, the money and the customers. Each has substantial material to consider. Giving them separate bots keeps their instructions and checklists easier to follow and review. Inside each, the work is still sorted by subject, such as design or engineering; those are roles the bot works in, not more bots.
These roles guide an existing computer intelligence through instructions and references. Naming a bot is not the same as training a new model, and a separate bot is not automatically more accurate. Good information, clear questions, and review still matter.
See why the work is split between bots. Each bot works in the role the job needs (design, engineering, legal, funding) without adding a bot.
Your staff assistants (persistent bots)
Your project lives in a private folder wherever you keep it (a cloud folder like Google Drive is best), with one copy of the bots’ instructions beside it in sops/. Each bot works from the version it has until you tell it to update, and keeps its plan there current. See what’s in your project folder.
Your bots are the continuing staff. You can call them back up, with their jobs, saved instructions, and available project history. Their job continues after today’s task ends. The analyst is the temporary helper: each run does one job and ends, and the next run starts from what was saved. They still need to read the relevant records; returning to the same assistant does not mean it recalls every past conversation perfectly.
Keep the current concept, requirements, decisions, sources, and open questions in the project folder. Give each assistant the relevant records and ask it to check their dates before working.
A handoff should say what was asked, what was found, where the result is saved, what is uncertain, and what needs a decision. The receiving assistant reads that result. Sharing a folder does not mean every bot automatically knows everything another bot has done.
If a job becomes too large, split it into smaller questions with named owners. Keep one assistant responsible for reviewing and combining the results. More helpers are useful only when the work divides sensibly.
An intern bot (temporary subagent): the Christmas-card example
In a plant, the contained job might be surveying the water flow on twenty molds, or pulling a year of scrap data into one table. The example below uses something everyone knows.
Suppose an HR assistant (persistent bot) looks after ongoing HR work. Now you want Christmas cards or personal holiday emails for customers, vendors, and employees. Preparing lists, checking names, and drafting all those messages could fill the HR conversation with a large seasonal project.
HR can write a clear instruction for an intern bot (temporary subagent): a helper called in to do that one job. HR gives the helper a defined assignment (task delegation). HR keeps responsibility for the work while the intern handles the detail.
- HR writes the brief. It names the approved contact lists, the kind of greeting, how to personalize it, the deadline, and where the drafts should go.
- The intern prepares the work. It checks for duplicates and missing information, drafts the greetings, and records anything HR needs to resolve.
- HR reviews the handoff. It checks the drafts and the exception list. You approve the final messages before they are sent.
- The office keeps the useful records. Save the reviewed drafts, the completion record, and an improved instruction sheet for next year. Once those are saved, the intern’s assignment ends.
Next year, HR reads the saved instructions, checks the current contact lists and preferences, and briefs a new intern. It can reuse the method without depending on last year’s temporary helper remembering the work.
The intern leaves the active team when the job is done. Treat its working memory as temporary and unavailable to the next helper unless you explicitly pass information on. Some platforms retain a transcript or let you resume a session; ending the assignment does not necessarily erase that stored history. The dependable memory is the reviewed work you saved.
HR’s ongoing role and records remain. Delegating keeps the seasonal detail out of its main conversation; the brief, review, and saved result give it what it needs to supervise and repeat the job.
Try writing the intern’s instructions
This is the kind of brief HR could prepare. The same structure works for a supplier comparison, a document review, or another contained task.
Example intern brief
You are the temporary helper for this year's Christmas greetings.
Use only the approved customer, vendor, and employee lists I provide.
Prepare one draft greeting per recipient. Personalize only from supplied facts.
Flag duplicate contacts, missing details, and uncertain preferences.
Save the drafts and an exception list in the agreed project folder.
Do not send anything. Return the file locations and a short summary to HR.
HR reviews the work, and the person approves the messages before sending.
Before finishing, suggest improvements to the saved instructions for next year.
Copy
In A.N. Creation, a bot writes this kind of brief and runs it as a separate Grok Build step (the analyst), then checks what comes back. If your assistant cannot launch another bot, you can open a separate conversation and pass it the brief yourself. Return its saved result to the supervising assistant for review.
OpenAI’s subagent guide describes delegating work and returning summaries to a main conversation. We call that temporary helper an intern bot (temporary subagent). The available controls depend on your assistant.
Why a smaller instruction file helps
The human page explains the role and the reason behind it. The assistant’s short file contains the instructions it needs to follow. It leaves out menus and presentation details so the important rules are easier to find.
How much page markup reaches a bot depends on the tool that reads it. A plain text instruction file gives you a clear, inspectable version either way. Concise relevant input leaves more room for the task; its effect on cost depends on your provider and plan.
Open the Project Manager Bot page, then its instruction file, standing.txt. The page explains the method; the file gives the office its working instructions. Both stay public and are updated together.
- Keep the lasting rules in a short file: the assistant’s role, boundaries, and where the project records live.
- Point the assistant at that file and ask it to confirm that it could read it. If it cannot open the link, provide the text directly.
- Put today’s task in the conversation, with the relevant project records.
- Before the task ends, save the result, decisions, and next step where the continuing assistant can find them.
For a job you intend to continue in the same session, keeping a resumable session can also help. That is a separate choice from using an intern that finishes one task and leaves. Both need saved results.
Pick your bot Read the procedures
How a bot keeps a record between turns is in the hotbotspot tool library.
Everything here is a starting point. Work with artificial intelligence always needs tuning, so test it and dial it in for your own work, data and model before you rely on it.