The workflow gap in Claude Code projects
Claude Code can complete a task quickly, but project context is often spread between terminal output, chat logs, and commit messages. That makes the next session slower and less predictable.
What “project memory” should include
- Decisions with rationale, not only the chosen outcome.
- Work logs with concrete file references.
- Pending tasks that survive session boundaries.
- A refreshed summary that a new agent can read in minutes.
A practical AGENTS.md pattern for Claude Code
Keep the instructions simple: read context first, write memory records after implementation. This keeps behavior consistent without making prompts long.
# Before coding
sidecar context --format markdown
# After coding
sidecar worklog record --done "<what changed>" --files src/a.ts,src/b.ts --by agent
# If behavior/design changed
sidecar decision record --title "<decision>" --summary "<why>" --by agent
# If follow-up exists
sidecar task add "<follow-up>" --priority medium --by agent
# Always refresh summary
sidecar summary refreshWhy this improves Claude Code output quality
- Less repeated investigation across sessions.
- Clearer intent behind changes when reviewing PRs.
- Better continuity between human and agent contributors.
- Lower risk that small but important follow-up work gets dropped.
Start simple
If you already run Claude Code in a repository, start by adding Sidecar and a 5-line instruction block. You can expand from there. The Agent Workflow guide shows a complete pattern.
FAQ
Does this only work for Claude Code?
No. The same Sidecar workflow also works with Codex, Cursor, and other agent tools. This page focuses on Claude Code usage patterns.
What should I put in AGENTS.md for Claude Code?
Keep it short: read sidecar context before work, record worklog and decision entries after work, add follow-up tasks, then refresh summary.
Will this replace my issue tracker?
Not usually. Sidecar is best for repo-local implementation memory and follow-ups discovered during coding.