Keeping context fresh on long tasks

A session that started brilliantly can turn sloppy two hours later. Nothing broke. The window just filled with old file versions, dead-end attempts and instructions you have since changed your mind about.

The fix is not a bigger window. It is a habit: checkpoint, reset, resume.

Spot a stale session

Flip the cards. If you recognize two or more in your current session, it is time to reset.

🃏 Symptoms of stale context

Tap a card to flip it

It reintroduces a bug it fixed an hour agoAn older version of the file is still in context and competing with the current one. Reset and let it re-read.It follows an instruction you already reversedBoth the old and new instruction are in the window. Put the current rule in writing and start fresh.It keeps retrying the same failing approachThe dead ends are anchoring it. A clean session with a note on what didn’t work breaks the loop.Answers get vaguer and slowerThe window is crowded with low-value output. Trim, or reset from a summary.It forgets a constraint from the start of the sessionEarly details get buried under hours of tool output. Move durable rules into a file it re-reads.You’re about to start a different task in the same sessionDon’t. Unrelated history is pure noise for the new task. One task, one session.

Checkpoint, reset, resume

1. 🛠️WorkOne focused chunk of the task
2. ↓→
3. 📝CheckpointCommit, then write a handoff note
4. ↓→
5. 🧹ResetClose the session, open a fresh one
6. ↓→
7. ▶️ResumeStart from the note and the code

↺ Repeat at every natural milestone

The key idea: state lives in the repo, not in the chat. A commit captures what was done. A short note captures what the code can’t: what’s left, what failed, what was decided.

A handoff note that works

Ask the agent to write it before you reset. Then read it: you’re the one who knows if it’s accurate.

Before we stop, write docs/progress/csv-export.md with:

- Goal: one sentence
- Done: what is finished and committed (with commit hashes)
- Next: the next 2-3 concrete steps, in order
- Decisions: choices made and why (so we don't revisit them)
- Dead ends: approaches tried that failed, and why
- Open questions: things I (the human) need to answer

Keep it under 40 lines. No code, only pointers to files.

Then the new session starts with one line: “Read docs/progress/csv-export.md and continue with the next step.”

The same handoff note is what lets you split work across several agents in parallel, which we get to in Module 5.

Before you move on

✅ Key takeaways

0 / 4 completed

• I can recognize the symptoms of a stale session and reset instead of pushing on.
• I keep one task per session and commit at each milestone.
• I have the agent write a handoff note, and I review it before resetting.
• Durable rules and decisions live in files, not in the chat history.

↺ Reset