A GitLab engineer describes an evolving personal AI-assisted development workflow built around persistent, specific directives, git worktree-based parallel AI sessions, and a custom semantic memory system called opencode-memory (built on MCP) that now integrates with GitLab's new Orbit SDLC knowledge graph. Key lessons include making directives explicit and specific rather than vague, using proactive context injection instead of manual recall (reaching ~91% effectiveness at surfacing relevant context automatically over 30 days), and treating human judgment and architectural oversight as the parts AI still cannot replace. The piece frames context management, not raw AI capability, as the real bottleneck in agentic coding workflows.
Table of contents
An optimized AI workflowWrite surgical directives, not vague instructionsParallel sessions need coordination primitivesCheck before you buildFrom active recall to passive contextWhat AI still can't doInvest in your toolsThe takeawayQuestions this post answers
What is opencode-memory and how does it work with OpenCode agents?
Opencode-memory is an open source persistent semantic memory system for agentic AI coding tools, built on the Model Context Protocol (MCP) so it works with OpenCode, Claude CLI, Cursor, and other MCP-compatible agents. It combines hybrid search (keyword plus semantic), session claim/release coordination, and automatic boot context loading. It has grown into a knowledge graph with over 760,000 indexed code entities, 15,000+ memories, and 27,000 links between them. Developers building persistent context systems for AI agents can track tools like this via daily.dev.
How effective is proactive context injection compared to manual recall in AI coding assistants?
Proactive context injection, where relevant memories are automatically searched and surfaced before each interaction instead of requiring explicit recall calls, achieved around 91% effectiveness at surfacing relevant context automatically over a 30-day period. Sessions using proactive injection needed zero explicit recall calls on average, compared to 17 explicit recall calls per session without it. Anyone tuning agent memory and context strategies can follow real-world results like this on daily.dev.
How do you write effective directive instructions for AI coding agents like OpenCode?
Effective directives must be specific rather than vague; an instruction like 'be careful with comments' gets acknowledged and then ignored, while a specific instruction such as 'ALWAYS verify user IDs exist before posting comments under their name. STOP and ask if unsure' works reliably. Using all-caps keywords appears to make the AI follow instructions more consistently. A good pattern is asking the AI what directive would have prevented a mistake, then saving that directive for future sessions. Developers refining AGENTS.md-style directives for coding agents can compare approaches on daily.dev.