Analyze code changes on the current branch and recommend updates to AGENTS.md files that have become stale. Use this skill before creating a PR, when the user asks to check AGENTS.md freshness, or when preparing an AGENTS.md update after a PR merges to main. Trigger whenever the user mentions syncing, updating, or checking AGENTS.md files, or when they're about to create a PR that touches code structure (new modules, renamed directories, changed commands, modified architecture).
Analyze code changes on the current branch and recommend updates to AGENTS.md files that have become stale. Use this skill before creating a PR, when the user asks to check AGENTS.md freshness, or when preparing an AGENTS.md update after a PR merges to main. Trigger whenever the user mentions syncing, updating, or checking AGENTS.md files, or when they're about to create a PR that touches code structure (new modules, renamed directories, changed commands, modified architecture).
AGENTS.md Sync
Detect when code changes make durable AGENTS.md guidance stale and produce a report with specific recommendations. Treat AGENTS.md files as operating manuals for future agents, not changelogs for recent diffs.
When to use
Before creating a PR — check if the branch's changes warrant AGENTS.md updates
After a PR merges to main — audit whether merged changes left AGENTS.md files outdated
On explicit request — "are the AGENTS.md files up to date?"
Workflow
1. Determine the diff
Figure out what changed relative to the base branch. The approach depends on context:
After a merge to main: the user should specify which PR or commit range to audit. Use the merge commit or PR number to get the diff:
gh pr diff <number>
Explicit check: if the user just says "check AGENTS.md", diff against main
2. Map changes to AGENTS.md territories
Each AGENTS.md file "owns" the directory it lives in and all subdirectories (unless a child directory has its own AGENTS.md). Build a list of which AGENTS.md files are relevant by walking up from each changed file to find the nearest AGENTS.md.
Example: a change to is owned by (if it exists), otherwise .
src/prefect/server/api/flows.py
src/prefect/server/AGENTS.md
src/prefect/AGENTS.md
Only analyze AGENTS.md files that have at least one changed file in their territory. Also include parent AGENTS.md files if structural changes (new directories, moved files) affect the directory tree they document.
3. Check for missing AGENTS.md files
Before analyzing existing AGENTS.md files, check whether any changed directories should have their own AGENTS.md but don't. New AGENTS.md files should be rare. Recommend one only when the directory has both:
A real ownership boundary that a future agent needs to recognize before editing
Hidden conventions, non-obvious contracts, or recurring pitfalls that cannot be learned quickly from reading the code
Look at the directories containing changed files. For each one that lacks an AGENTS.md, check:
Ownership boundary: Does this directory represent a responsibility shift with different contracts than its parent?
Hidden knowledge: Are there non-obvious patterns, invariants, conventions, or failure modes that a developer would not learn quickly from code?
Durability: Will this guidance still matter after the triggering PR is old news?
Hierarchy fit: Would this belong better in an existing parent AGENTS.md instead?
Directory size and sibling AGENTS.md files can be weak signals, but they are never sufficient on their own. Do not recommend a new AGENTS.md just because a directory has 5+ files, gained a new module, or has sibling directories with AGENTS.md files.
When the case is not high-confidence, do not create the file. Capture the possible update as a recommendation instead, with the missing evidence or question that would make it actionable.
When recommending a new AGENTS.md, use the Explore agent to read the directory's code and draft the file using this template:
# <ModuleName><One-sentencepurposestatement.>## Purpose & Scope
What this module does. Explicit responsibility boundaries.
What it does NOT do (prevents scope creep).
## Entry Points & Contracts
Key APIs, functions, classes that external consumers use.
Input/output contracts. Invariants that must hold.
## Usage Patterns
Canonical examples of correct usage.
"If you need to do X, here's how."
## Anti-Patterns
What NOT to do. Common mistakes. Each should be a real mistake, not hypothetical.
## Pitfalls
Non-obvious gotchas. Implicit assumptions.
"You'd think X, but actually Y because Z."
The initial content must be concise and durable. Do not include [ASK] placeholders, file lists, changelog context, or speculative guidance. If the hidden knowledge cannot be stated confidently, write a recommendation instead of creating the file.
4. Analyze each relevant AGENTS.md
AGENTS.md files form a hierarchy — when an agent loads any file, all ancestor files load too, creating a T-shaped view (broad context at the top, specific detail where the agent works). This has two implications for the sync check:
Least common ancestor rule: Shared knowledge should live at the shallowest AGENTS.md that covers all paths needing it. If the same fact appears in multiple sibling AGENTS.md files, flag it — it should move to their parent. Conversely, if a parent AGENTS.md contains detail that only applies to one child directory, suggest moving it down.
Upward propagation: When a child AGENTS.md changes (or a new one is created), check whether the parent needs updating — either to add a cross-reference to the new child, or to re-summarize what the child directory does. Parent nodes should summarize their children's responsibilities, not duplicate their details.
For each AGENTS.md in scope, read it fully, then compare its claims against the diff and the current source code. Look for these categories of issues:
Accuracy — claims in AGENTS.md that don't match the actual code, regardless of whether the diff caused the discrepancy:
Described behaviors that don't match what the code actually does (e.g., "renews at 75% of duration" when it actually renews immediately then sleeps 75%)
Function signatures or parameters listed that are wrong or incomplete
Responsibilities attributed to the wrong layer or module
Error handling or failure modes described incorrectly
For claims that are verifiable (specific behaviors, function signatures, which module does what), read the source file and confirm. Don't trust that existing AGENTS.md content was ever correct — it may have been wrong from the start.
Signal density — content that duplicates what's discoverable from code rather than surfacing hidden knowledge:
Full function signatures that just repeat the code (the contracts and gotchas around those functions are the high-signal parts)
Trivial usage examples that match what's in docstrings or is obvious from the API
File-by-file directory listings that ls would show — the layering concept or architectural invariant matters, not the enumeration
Content that a developer would learn in 30 seconds of reading the source
AGENTS.md should capture what's not visible in the code: invariants, hidden contracts, non-obvious failure modes, things that look one way but behave another.
A behavior discovered in one changed function is not AGENTS-worthy by default.
Promote it only if it exposes a subsystem-level editing rule, a cross-file
contract, or a recurring failure mode future agents need before choosing
which implementation to read. Otherwise prefer a code comment, docstring,
test name, or no documentation change.
If the same guidance is already present in nearby code comments, docstrings, or tests, AGENTS.md should usually not repeat it.
Missing invariants — implicit contracts the code relies on that aren't documented:
Parallel implementations that must stay in sync (e.g., sync and async versions of the same module)
Ordering requirements or sequencing constraints
Singleton behaviors or shared mutable state
Cleanup responsibilities and what happens when they're skipped
"This looks stateless but actually depends on X"
Structural drift — the AGENTS.md describes an ownership boundary, durable package layout, or module responsibility that no longer matches reality:
Directories or packages added/removed/renamed in a way that changes responsibility boundaries
New architectural layers or cross-file contracts not mentioned
Removed components still listed as active responsibilities
Command drift — documented commands that may no longer work or are incomplete:
Convention drift — code patterns or conventions described in AGENTS.md that the diff contradicts:
New patterns established by the changes that aren't documented
Existing conventions the changes deprecate or replace
Description drift — prose descriptions that no longer accurately characterize the code:
Component descriptions that miss new responsibilities
Architecture notes that don't reflect refactoring
Hierarchy drift — content living at the wrong level in the AGENTS.md tree:
Knowledge duplicated across sibling AGENTS.md files that should be in their parent (least common ancestor)
Parent-level detail that only applies to one child directory
Parent summaries of child directories that no longer match after changes
For each finding, check whether it's actually stale — read the current state of the filesystem (not just the diff) to confirm. A diff showing a new file, helper, flag, or one-off fix doesn't mean AGENTS.md is wrong if AGENTS.md intentionally omits that level of detail.
Only recommend or apply updates for high-confidence, lasting guidance:
Architectural boundaries
Non-obvious invariants
Cross-file contracts
Recurring failure modes
Commands or conventions that future agents are likely to need
Do not document:
One-off implementation details or temporary bug context
Facts obvious from reading names, imports, or nearby code
File lists or every new module
Changelog-style summaries of the triggering diff
Advice that only matters for the current PR
5. Produce the report
Output a concise report grouped by AGENTS.md file. For each file, list findings with:
What's stale — the specific line or section
Why — which change caused it
Suggested fix — the concrete text to add, remove, or change
Format:
## AGENTS.md Sync Report### New AGENTS.md recommended: `src/prefect/utilities/`
This directory now owns a distinct retry protocol used by multiple callers, and contributors must preserve its ordering invariant. Suggested outline:
- Purpose and ownership boundary
- Ordering invariant that callers rely on
- Failure mode to test when changing it
### `src/prefect/server/AGENTS.md`1.**[Accuracy]** Lease renewal described as "renews at 75% of duration" but code renews immediately then sleeps 75%
- Caused by: incorrect from initial creation
- Suggested fix: "renews immediately on entry, then sleeps for 75% of `lease_duration` between renewals"
2.**[Signal density]** Full function signatures duplicate the code — remove parameter lists, keep contracts
- Suggested fix: Replace `concurrency(names, occupy=1, ...)` with `concurrency()` and describe the non-obvious behaviors instead
3.**[Missing invariant]** Sync and async implementations must stay in lockstep but this isn't documented
- Suggested fix: Add "Any behavior change to `_asyncio.py` must be mirrored in `_sync.py`"
4.**[Structural]** New directory `src/prefect/server/events/` not listed in directory structure
- Caused by: `src/prefect/server/events/` added in this branch
- Suggested fix: Add `events/` to the directory listing with description
### `./AGENTS.md` (root)
No changes needed.
If nothing is stale, say so clearly — don't manufacture findings.
6. Offer next steps
After presenting the report, offer:
Apply fixes — make the suggested edits directly
Create a follow-up PR — if this is a post-merge audit, offer to create a branch and PR with just the AGENTS.md updates
Skip — if the changes are minor or the user disagrees with the suggestions
Reactive mode
Outside of diff-driven analysis, if during normal development you discover missing context that should be in an AGENTS.md, propose an addition. Signs of missing context:
You spent nontrivial time reconstructing a durable invariant that could be a one-liner in AGENTS.md
You hit a non-obvious error because of an undocumented constraint
You discovered a cross-system dependency that isn't mentioned anywhere
A code review caught something that AGENTS.md should have prevented
Propose additions like this:
I discovered that [finding]. This isn't in [path]/AGENTS.md.
Proposed addition to the [section] section:
> [concrete text to add]
Should I add this?
Quality checks for new or updated AGENTS.md content
Compression: Can any section be shorter without losing information?
Deduplication: Does anything repeat what a parent AGENTS.md already says? If so, remove it and rely on the hierarchy.
Specificity: Replace vague statements with concrete ones. Not "be careful with auth" but "place logic after the require_* call".
Target length: 20-80 lines for new files. If longer, compress or write recommendations instead.
Important guidelines
Be conservative — only flag things that are genuinely wrong or missing. AGENTS.md files are intentionally concise and don't need to document every file.
Don't suggest adding content that's obvious from the code itself. AGENTS.md captures non-obvious patterns, gotchas, and high-level structure. Equally, flag existing content that merely restates the code — full function signatures, trivial examples, and file enumerations are low-signal and should be trimmed or replaced with the hidden knowledge around them.
Don't trust that existing AGENTS.md content was ever correct. Verify factual claims (behaviors, signatures, responsibility attribution) against the source code, not just against the diff. Content can be wrong from day one.
Don't suggest stylistic rewrites. Focus on factual accuracy and signal density.
When in doubt about whether something is stale, check the filesystem to confirm rather than guessing from the diff alone.
Respect the existing level of detail in each AGENTS.md. If a file only lists top-level directories, don't suggest adding individual files.
Look for undocumented invariants — parallel implementations that must stay in sync, ordering constraints, singleton behaviors, cleanup responsibilities. These are the highest-value additions because they're invisible in code and cause the most breakage when violated.
