AI coding decisions and constraints got lost in the chat
Decisions are usually tiny. Their absence is huge. The next AI needs to know not just what changed, but why a path was chosen.
Why this hurts.
Without decisions and constraints, the next session can undo architecture choices, re-add rejected dependencies, or touch files that were intentionally out of scope.
How ShardStitch solves it.
Find durable decision records
Pin decisions, constraints, rejected approaches, and known truths.
Attach each decision to its source
Include the note, commit, issue, or file that supports each decision; mark a source as missing if no record survives.
Expose unknown rationale
If the reason for a constraint is unavailable, state that gap rather than inventing it; route risky uncertainty for review.
Carry constraints into the next change
Tell the next tool what not to touch and what assumptions must be verified.
What the next AI receives.
- Decision summary
- Constraints
- Rejected approaches
- Do-not-touch areas
What stays out.
- Raw debate
- Unresolved guesses
- Old preferences that no longer apply
- Verbose transcript history
Recover decisions without inventing rationale
FAQ.
Can the code reveal why a previous decision was made?
Sometimes comments or review history explain it, but the implementation alone is not reliable evidence of rationale. Mark the reason unknown unless a durable source supports it.
What should I include for a decision that may have changed?
Include the decision, its source and date if available, and what evidence would invalidate it. Ask the next session to verify that condition before relying on the decision.