Pick up where you left off, without explaining the whole project again.
Engineering With AI gives decisions, evidence, standards, meetings, plans, tests and learning predictable homes. Each new conversation can start from the project’s reviewed memory rather than somebody’s best recollection.
A project should remember more than its code.
The useful context isn’t the largest context window. It’s the smallest trustworthy set of evidence for the decision in front of you.
The next session shouldn’t have to guess what you agreed.
An illustrative project history, showing the difference between a conversation and reviewed project knowledge.
- Monday: proposed
“Could we let customers export their records?”
The meeting is source evidence. It isn’t permission to build an export for every role and every field.
- Tuesday: agreed
The owner narrows the scope.
Only the customer’s own records, with sensitive fields excluded. The reviewed intent records the decision and its reason.
- Thursday: tested
One permission scenario still fails.
The result belongs with the delivery evidence. A green summary mustn’t overwrite the unresolved finding.
- Next session
Start with the current decision and the failing case.
Bring the relevant source, constraint and evidence into context, rather than repeating the whole meeting transcript.
Chat history isn’t project memory.
Conversations mix ideas, decisions, corrections and abandoned routes. When all of that is fed back to an agent as one undifferentiated transcript, old assumptions can return as instructions and a late summary can quietly overwrite better evidence.
The harness keeps canonical project knowledge in a versioned SPECS structure and connects it to the evidence that supports it. Search can find the full text. Context assembly selects what’s relevant. Review decides what becomes trusted project knowledge.
This makes AI more useful without asking it to remember everything or guess which sentence mattered most.
Keep the decision, not just the conversation.
Project-local SPECS
Purpose, domain, journeys, architecture, constraints, standards, decisions, delivery work and evidence live in a predictable structure that people can inspect without opening the original AI conversation.
Mind Palace search
Index and retrieve the full text of project knowledge, identify stale or duplicated material and surface the source behind an answer.
Bounded context assembly
Assemble only the standards, decisions, source relationships and evidence needed for a task, then measure whether the context was useful rather than simply large.
Meeting evidence intake
Turn a transcript into classified proposals with provenance. A statement from a meeting can inform the project without being mistaken for an approved decision.
Evidence-to-knowledge proposals
Promote useful learning into the right canonical document through an explicit review path. Rejected and deferred proposals remain visible instead of vanishing.
A context-aware delivery companion
Ask what the project knows, what phase the work is in, which standards apply and what evidence is missing. The companion explains the route without changing delivery state or inventing approval.
Keep the unanswered questions, too.
Some of the most useful context is what you haven’t decided. In the customer-portal draft, retention and failed-download investigation still need an answer.
Recording that uncertainty gives the next conversation an honest starting point. A meeting suggestion can become a proposal; a proposal needs review before it becomes a project decision.
What does the next person or agent need to know?
Architecture and decision records retain the options, constraints and rationale behind the current implementation.
Provenance and classification keep a suggestion in a meeting separate from a reviewed decision or implemented behaviour.
Phase evidence, amendments, local error reports and retrospectives provide a usable trail of progress, correction and learning.
Task-bounded assembly chooses the relevant evidence and records omissions. It doesn’t dump the whole project into every prompt.
Let the next conversation begin with what the project has already learned.
Keep the reasoning close to the work, make changes reviewable and stop important decisions dissolving into chat history.