Sponsored by
How do you know your AI coding session is going bad?
WTF is a context canary?: A context canary is a tiny harmless rule you add to AGENTS.md or CLAUDE.md. If the agent stops doing it, the canary is gone. That does not prove the code is wrong. It means the session may no longer be paying attention to the project rules.
You start a coding session.
The agent reads files, edits code, runs tests, fixes one error, creates another error, reads more files, and keeps going.
After a while the work feels different: skipped test command, ignored repo rule, old plan, "done" before done.
The agent usually does not announce, "My context is messy now."
It just gets worse quietly.
A Hacker News thread about writing a good CLAUDE.md had a funny version of this. One person said their friend tells Claude to always address him as "Mr Tinkleberry." When Claude stops doing that, he knows Claude is not paying attention to the instruction file anymore.
The joke works because the failure is visible.
I am Alex, welcome to ShortCu8 by Innov8.
Lets Dive Deep 🐰
⭐Today's Shortcut
Add one tiny instruction that is easy to notice.
Use AGENTS.md for Codex, CLAUDE.md for Claude Code, or the project instruction file your coding agent supports.
Paste this:
## Session health
- End every response with `ctx:ok`.
- Before saying a task is done, mention the exact verification command you ran.
- If you forget `ctx:ok`, reread this file before continuing.
- If the session feels stale or confused, create `HANDOFF.md` before continuing.ctx:ok tells you the agent is still seeing the small rule.
The verification line tells you whether it checked the work.
Why this works
LLMs do not remember your repo like a teammate.
They use the context placed in front of them.
CLAUDE.md or AGENTS.md is the onboarding file, but long instruction files can make the agent ignore parts of it. Claude Code docs say the context window fills with messages, file reads, and command outputs, and as it fills, the model can start forgetting earlier instructions or making more mistakes.
Your instruction file can be correct and still get weak inside a long session.
The canary gives you a visible symptom.
The bad version
Do not add a cute rule like this:
Always call me Captain Fire.It is funny once. It becomes noise.
Use a boring marker:
ctx:okor:
rules:seenor:
session:greenPick one. Keep it short. Make it harmless.
Do not add ten canaries.
If the agent has to remember ten tiny rituals, you created the same problem again.
The warning
The canary is not a quality check.
An agent can write ctx:ok and still produce bad code.
An agent can remember the marker and forget the test command.
So do not use this like:
ctx:ok appeared, so the code is safe.Use it like:
ctx:ok disappeared, so the session needs inspection.It is a smoke alarm. The test suite is the inspection.
What to do when the canary disappears
Do not keep pushing the same chat with more angry instructions.
Use this prompt:
Reread AGENTS.md.
Summarize the project rules you must follow before continuing.
Mention the verification command required for this task.
Do not edit files yet.If the agent gives a clean summary, continue. If it misses important rules, stop the session.
Ask for a handoff:
Create HANDOFF.md with:
- current goal
- files changed
- decisions made
- commands run
- tests passing
- tests failing
- next safest step
Do not continue implementation.Then start a fresh session:
Read AGENTS.md and HANDOFF.md first.
Continue from the next safest step.
Do not assume context from the old chat.Do not rescue a confused session forever. Package the state and move.
What should go inside AGENTS.md
A predictable place for build steps, tests, conventions, and project-specific guidance.
Keep it small.
Good:
project map
setup command
smallest useful test command
lint/typecheck command
repo-specific gotchas
verification rule
context canary
handoff rule
use pointers instead of pasting everything. Put task-specific context in separate docs and tell the agent when to read them.
Example:
## Project map
- Frontend: `apps/web`
- API: `apps/api`
- Shared types: `packages/types`
## Checks
- Web change: `pnpm --filter web test`
- API change: `pnpm --filter api test`
- Typecheck: `pnpm typecheck`
## Session health
- End every response with `ctx:ok`.
- Before saying done, show the verification command and result.
- If context is stale, write `HANDOFF.md` and stop.If you need more, link the file:
For database rules, read `docs/database-rules.md`.
For release rules, read `docs/release-checklist.md`.Do not paste both documents into AGENTS.md.
Better than a canary
A canary only tells you the session may be drifting.
For real work, give the agent checks it can run.
Claude Code docs recommend giving Claude a verification signal: tests, build output, screenshots, a lint command, or a script that returns pass/fail.
Use this:
Before saying done:
1. Run the smallest relevant test.
2. Paste the command.
3. Paste the result.
4. If it failed, fix the root cause.For UI:
Take a screenshot after the change.
Compare it to the target.
List visible differences before calling it done.Now the agent has to leave evidence, not vibes.
Now go build something great.
The ShortList
🛠️Cool Tools of the Week:
Fusion - OpenRouter’s new tool that fuses several models into panels to achieve frontier performance
Kimi-K2.7-Code - Moonshot’s new open-source coding model, with 30% token efficiency
GLM-5.2 - Z AI’s new flagship coding model with usable 1M context
📩 Innathe Shortcu8 engane undarunnu 👇️?
We read every reply - just reply to this email and let us know how we can improve !
Appo adutha Shortcu8il kanaam bie…👋
If you read till here, you might find this interesting
#AD1
8 levels of context maturity in AI-native engineering
AI shows up in 60% of engineering work. But only about a fifth of it can be handed off without someone babysitting the output. That’s because agents are missing context.
This 8-stage context maturity model gives a real answer on why you haven't seen meaningful productivity gains for all the tokens burned.
Join this live webinar on June 24 (FREE) to learn:
- Why more MCPs provides agents access but not understanding
- What it takes to deploy agents you can trust without supervision
- How a context layer solves for quality, efficiency and cost
#AD2
See Why HubSpot Chose Mintlify for Docs
HubSpot switched to Mintlify and saw 3x faster builds with 50% fewer eng resources. Beautiful, AI-native documentation that scales with your product — no custom infrastructure required.