How we give an AI assistant current project context
An internal implementation note on retrieving current objectives, decisions, and evidence without treating an old chat summary as live truth.
An assistant can sound well informed while working from yesterday’s understanding of a project. Long chat histories make this harder: the original request, a later correction, and a completed change can all be present, but they do not have equal relevance to the next action.
This is a note about an internal Bransford Media workflow for retrieving business context. It describes an approach we use; it does not claim that an assistant can operate without oversight or that a saved summary proves current reality.
Keep current work in one owning system
Our current business tasks live in Twenty. A task provides a place for its objective, status, relevant evidence, decisions, and remaining work. Website content has a different owner: the website CMS. A project map can point to those systems without becoming another copy of every record.
That ownership distinction matters. A useful retrieval tool should help an assistant find the authoritative record, not quietly create a competing tracker because it is convenient in the current chat.
Retrieve the small amount of context needed now
The initial context should explain what is being attempted, what has already happened, what remains, and which decisions constrain the next step. It should link to evidence rather than copy every historical message into the prompt.
When more detail is needed, the assistant can search and retrieve the relevant history. Exact history remains available. The summary helps navigation; it is not a replacement for the underlying record.
Check whether the handoff is stale
A checkpoint is useful only if the reader knows what it covered. Our context workflow compares the checkpoint with the task’s later history so that newly added decisions can be identified.
An unchanged task still does not prove that an external system is unchanged. Before editing a website, for example, inspect the current CMS record and live release. Before continuing implementation, inspect the current source checkout. That prevents a previously accurate handoff from becoming permission to overwrite newer work.
Separate evidence from authority
A historical note can explain why a decision was made. It does not automatically authorize a new publication, message, purchase, or permission change. The assistant needs the current request and the applicable operating boundaries as well as the project context.
Likewise, a tool response that says “saved” establishes a save, not a deployed result or an approved design. Keeping those states explicit makes a handoff more useful to the next person or assistant.
Make the next handoff easier
At the end of substantial work, preserve a concise checkpoint with the verified result, links to evidence, remaining work, and any material limitation. Keep it in the existing task rather than opening another status document.
This pattern is useful when work spans people, tools, and multiple sessions. Start by making ownership and retrieval clear. Add automation when the retrieval process is dependable enough to support it.
Our marketing systems work covers the underlying records and handoffs. AI implementation covers how an assistant uses those sources, what it may do, and how the result is checked.