The instructionswere the bug.
Real replies from developers about the files their coding agents read. Every quote is verbatim and links to its author.
“ ” their wordsour read
Pain validation They felt the bug.
Our readThree developers, unprompted, describing the same failure: the code was fine, the instructions were not.
yeah my markdown file just quietly drifts until something breaks weird and i spend an hour figuring out the instructions themselves were the bug. never thought of it as a dependency problem but that's exactly what it is
Karim COn X We review every line AI writes, but barely question the instructions that control it. Wild.
sasrOn X Your codebase might be secure. The instructions guiding your AI might not be. thiss is so powerful i tried thiss 🔥
Flow SenitalOn X Technical due diligence The questions you ask before adopting a tool.
Our readSeven questions from one engineer, each more specific than the last. Interest worth earning, not a purchase.
- @agent_dot_md replied · 27 Aug@i_mika_el on X
Great question. Agent.md never overwrites your project rules. link only manages its own <!-- agentmd:start --> block. Your edits stay untouched, and modified packages won't be overwritten by update without --force.
- @agent_dot_md replied · 27 Aug@i_mika_el on X
Yes, "agentmd update --dry" writes nothing. It shows the line diff + added/removed counts for each package, including conflicts. For local edits, it shows both sides and leaves the decision to you. update only applies after you approve.
- @agent_dot_md replied · 29 Aug@i_mika_el on X
Yes. "init" scans workspaces independently, so mixed stacks work: "apps/web" can detect Next.js/React while "services/api" detects Express/Prisma/Postgres. Recommendations are unified, but the stack breakdown shows where each signal came from.
- @agent_dot_md replied · 3 Sep@i_mika_el on X
Yes, It's monorepo-aware. "init" scans each workspace independently, so "apps/web" can detect Next.js/React while "services/api" detects Express/Prisma/Postgres. Recommendations are then unified across the repo.
- @agent_dot_md replied · 3 Sep@i_mika_el on X
Three-way, content-addressed drift detection. Agent.md compares your install baseline, local file, and upstream. "update --dry" shows the diff first; CI can surface drift automatically.
No public reply yet. This one is still open.
@i_mika_el on XNo public reply yet. This one is still open.
@i_mika_el on X
Vision validation They get the bigger idea.
Our readSix people, in their own words, describing agent context as something shared, versioned and maintained.
Standards should be shared infrastructure, not personal folklore, so here's to making engineering wisdom a package you can update instead of a story you tell.
MrOziOn X Once agent instructions become reusable and versioned like dependencies, teams can improve their standards once and propagate those improvements across every project instead of maintaining another pile of slightly different rules.
Paw DogsOn X Most devs overlook how project context shapes recommendations. It's not just about detecting stacks; it's about adapting those rules to fit team dynamics and workflows. Standards are only as good as their application in real situations.
Keith TsangOn X Feels like the next layer of AI coding isn't just better models it's better understanding of the environment they're operating in. The repo itself should be part of the prompt.
SHIPITOn X Security, testing, DevOps together feel like an agent starter kit.
Fajar M RezaOn X Brilliant Agent.md eliminating setup friction lets developers focus purely on building.
Sheral TechOn X Caught in the wild
A secret, an injection or a dead path that npx activate-agentmd audit found in a real repo. Post it on X and tag @agent_dot_md. The best ones land here.
Check what your agent reads.
$ npx activate-agentmd audit owner/repo
