I Got Tired of Re-explaining My Project to AI Agents Every Single Session

Let me paint a picture you might recognize.

You spend a solid session with an AI coding agent. It understands your project deeply — the architecture, the decisions you made, the weird edge case you worked around. It's flowing. You're building fast.

Then you close the laptop.

Next day, you open a new session and... blank slate. You're back to square one, re-explaining everything. Which branch. What phase. What was left unfinished. What you already tried.

Every. Single. Time.

I got tired of it. So I built something small to fix it.

---

The Core Problem: Agents Don't Talk to Each Other

When you're working with AI agents across multiple sessions (or multiple agents on the same project), there's no shared memory by default. Each session is an island.

What you actually need is a coordination hub — a living document that any agent can read on startup and write to before it leaves. Something that answers:

• What phase are we in?
• What tasks are done, in-progress, or pending?
• What did the last agent accomplish?
• Are there any blockers?

That's exactly what I built.

---

Introducing AGENT_SKILLS

AGENT_SKILLS is a small open-source repo with two plug-and-play skills for opencode:

🔧 Skill 1: init-agent-sync

This skill creates an AGENT_SYNC.md file in your project root — a structured coordination hub with four sections:

• AI Instructions — tells every agent exactly how to behave (read this first, update tasks while working, write a handoff summary before leaving)
• Current Context — active branch, current phase, master reference file
• Session Task Board — granular short-term tasks with [ ], [~], and [x] states
• Handoff Log — a timestamped log for agent-to-agent communication

Run it once at the start of a project and you're set.

mkdir -p .opencode/skills/init-agent-sync
curl -o .opencode/skills/init-agent-sync/SKILL.md \
  https://raw.githubusercontent.com/Hadi99K/AGENT_SKILLS/main/init-agent-sync/SKILL.md

Then just tell your agent: "Run the init-agent-sync skill."

---

🧹 Skill 2: wrap-phase

At the end of a development phase, this skill:

1. Reads AGENT_SYNC.md to understand everything that was done
2. Writes a clean ## Phase Summary section documenting accomplishments
3. Clears all completed [x] tasks from the board
4. Purges old handoff log entries (keeps only the most recent one)
5. Updates the phase status to complete ✅
6. Auto-commits and pushes — no confirmation needed

Just say: "Wrap up the phase."

mkdir -p .opencode/skills/wrap-phase
curl -o .opencode/skills/wrap-phase/SKILL.md \
  https://raw.githubusercontent.com/Hadi99K/AGENT_SKILLS/main/wrap-phase/SKILL.md

---

The Loop This Creates

Once you have both skills, your workflow becomes a clean cycle:

init-agent-sync
      ↓
   [build stuff]
      ↓
  wrap-phase
      ↓
   [new phase]
      ↓
   repeat...

Every new session, your agent reads AGENT_SYNC.md first. It knows what was built. It knows what's pending. It picks up exactly where you left off — no re-explaining, no repeated work, no lost context.

---

Why I Kept It Simple

I deliberately kept this as two small Markdown files rather than a complex tool or plugin. A few reasons:

• No dependencies — just a file your agent reads and writes
• Version-controlled — AGENT_SYNC.md lives in your repo, so you have full history
• Agent-agnostic — any agent that can read and write files can use this pattern
• Transparent — you can always open the file and see exactly what's happening

---

Try It Out

The repo is open-source (MIT licensed) and lives here:
👉 github.com/Hadi99K/AGENT_SKILLS

It currently targets opencode, but the underlying pattern works with any AI coding assistant. If you adapt it for Claude Code, Cursor, or anything else — I'd love to see a PR.

If this solves the same problem for you that it solved for me, drop a ⭐ on the repo. And if you have ideas for more skills — contributing is wide open.

---

What coordination problems are you running into with your AI agents? Would love to hear in the comments.