Context Handoff

Context Handoff is a free browser game for software testers about writing the instructions file an AI coding agent reads: you decide which notes belong in a durable agents.md, which die with the current run, and which should never be written down, then author that file - choosing headings, tightening the wording and resolving conflicts with a nested AGENTS.md - and finally watch an agent read it line by line and do visible damage wherever a line was weak.

AI coding agents start every session with no memory of the last one. Whatever you want tomorrow's agent to know, you have to write down tonight, in a file it reads before it starts work. That file is called agents.md, and this game is about what goes in it. Every piece of context you hold is true right now: some of it will still be true next month, some of it stops being true tonight, and some of it should never have been written down. Context Handoff makes you decide which is which, then makes you write the file, then hands that file to an agent that does not remember you and never asks what you meant. Context Handoff is supported by ShiftSync, the Tricentis community for quality engineering, and the glossary inside the game links on to further reading from ShiftSync, Testing Titbits and Ministry of Testing.

Open Context Handoff on Testing Titbits Playground · 5-7 min · AI Testing · Free, no sign-up

How to play Context Handoff

  1. A three-note warm-up teaches what an agents.md file is and what does not belong in one.
  2. A team charter states four things the team holds true: how often it ships, what its runner can already reach, what counts as evidence, and what it does with an unmeasured hunch.
  3. Act I: nine notes arrive one at a time. Send each to agents.md, to this run only, or to nowhere at all.
  4. More than one lane is often defensible. The charter makes the call answerable without making it automatic.
  5. Act II: the notes that belong in the file become a file you author. Choose the heading each line sits under.
  6. Rewrite the rambling lines. A meter prices every line you keep, on every run, forever.
  7. A nested tests/AGENTS.md contradicts the root file three times. Decide which one the agent obeys.
  8. Act III: the agent reads your file line by line. Where a line was weak you watch the damage land, then patch it and run again.
  9. One last beat fires on a line you wrote well. The agent honoured it on turn two and had stopped by turn forty, and only one of the three remedies actually holds. Scored out of 100.

What you will learn

Frequently asked questions

What is context engineering?

Context engineering is the practice of deliberately choosing what information an AI model receives - the instructions, documents, examples and tools - rather than relying on prompt wording alone. Deciding where each piece of that context lives and how long it should live is part of it.

What is an agents.md file?

agents.md is a plain-text instructions file that AI coding agents read when they start work in a repository. It carries the durable context a session needs - conventions, environments, known workarounds - so it does not have to be re-explained every run. Anything in it is read on every future run, whether or not it is still true.

What should go in an agents.md file?

Context that is true on every run and cheap to state once: environments, the conventions your team always follows, prohibitions, and known traps in the product. Not anything with a date, a version or a ticket number in it, and never a credential, a token or real customer data.

Which file wins, agents.md or a nested AGENTS.md?

The file nearest the work usually wins for its own scope, because a tests/AGENTS.md was written by the people who own the tests. The exception is a prohibition. A nested file can narrow a rule from the root file, but it must not be allowed to lift one, since the point of writing a prohibition down is that it does not have local exceptions.

What is stale context?

Stale context is a note that was accurate when it was written and is not any more - last sprint feature flags, an expired token, a version number. It is more dangerous than missing context because the model uses it confidently and nothing flags it as out of date.

How much does an agents.md file cost to keep?

Every line is re-read on every run, so the cost is a rate rather than a one-off. A file read a hundred times a week costs a hundred times the tokens of a file read once, which is why a line that says nothing is not free.

What is a skill file, and when should a procedure move out of agents.md?

A skill file is a separate document an agent opens when it needs it, rather than reading on every run. A multi-step procedure - a release checklist, a data reset runbook - is the clearest case: it is durable, so it does not belong in the session, but it is only needed on the runs that perform it, so keeping all of its steps in agents.md charges every other run for them. Leave one line in agents.md pointing at the file and keep the steps in the file. That line has to point at something the agent can actually open: a pointer to a procedure nobody wrote down is the same length as one that works and worth nothing.

Why do AI agents stop following their instructions file?

Because the file is only obeyed while it is still in the context the model is attending to. An agents.md read at the start of a long run competes with everything that arrives after it, so a rule honoured on turn two can be quietly gone by turn forty, and nothing errors when it goes - the agent simply does what it would have done if the line had never been written. Restating the rule works, and then wears off again. The remedy that holds is structural: keep the file short enough that the rule is still being read at the end of the run, rather than relying on the agent to remember it.

Who supports Context Handoff?

ShiftSync, a Tricentis community of around 80,000 testers and quality engineers, supports Context Handoff. The game is free to play, needs no account, and the ShiftSync link on the result screen goes to their community. Testing Titbits builds and maintains the game.

Is this game suitable for a team session?

Yes. It runs about five to seven minutes, several of the calls are deliberately arguable, and a workshop seed pins everyone to the same nine notes so the answers are worth comparing afterwards. Share the seed link from the result screen.

More free games and infographics for testers

Created by Rahul Parwal · TestingTitbits.com