Use ORMs for their productivity while avoiding N+1 queries and knowing when to drop to raw SQL. Use when working with an ORM or debugging the performance problems ORMs quietly…
Authenticate and authorise access to a remote MCP server so tools act as the right user with the least necessary privilege.
Support right-to-left languages by mirroring layout, icons, and interactions while leaving numbers and code untouched.
Tie every claim in a generated answer to the retrieved passage that supports it, so users can verify and unsupported claims are visible.
Reach battery-life targets through sleep-state budgeting, duty cycling, and measured (not estimated) current draw.
Apply a code of conduct consistently, with a reporting path, proportionate responses, and records, so it protects people rather than decorating the repository.
Review translated text for accuracy, register, and fit in context, with a rubric rather than an impression.
Define measurable reliability targets from the user's perspective and use the error budget to decide between shipping and stabilising.
Close the gaps between test and production by matching versions, data shape, and configuration paths so a passing suite actually predicts prod behavior.
Choose merge or rebase by the history you want, use interactive rebase safely, and set a team policy.
Run structured unscripted testing with charters, tours, and timeboxed sessions to find bugs scripted tests never imagined.
Partition datasets by pruning-friendly keys with healthy file sizes and scheduled compaction. Use when laying out lake or warehouse tables, or fixing slow scans and small-file…
Remove features with usage evidence, migration paths, and staged communication that preserves trust. Use when retiring product surface or making the case that something should die.
Run the same task several times independently and combine the results by agreement, to damp variance on judgement calls.
Design foreign-function boundaries with explicit ownership, error translation, and contained panics/exceptions.
Use agents to triage contracts and policies, extract obligations, flag unusual terms, and prepare a lawyer's review rather than replace it.
Change prices or plan structure without breaking trust or billing, deciding who moves, when, and with what notice.
Checkpoint multi-node training runs so a save costs seconds instead of minutes and a resume reproduces the run exactly.
Follow email conventions that keep you professional and considerate: CC/BCC, reply-all, response times, and threading.
Ship over-the-air firmware with A/B slots, signed images, power-loss resilience, and staged fleet rollout.
Run external programs without ever building a shell string from untrusted input, using argument arrays that bypass the shell entirely.
Send the messages a user is waiting for, such as receipts, resets, and confirmations, with reliability and clarity that marketing messages do not need.
Write async C# that never blocks on tasks, flows cancellation everywhere, and uses ValueTask where it pays.
Block cross-site request forgery with SameSite cookies, per-session anti-CSRF tokens, and strict HTTP method discipline so a forged cross-origin request cannot act as the user.
Engineer features for tabular models (aggregations, encodings, interactions) that add signal without leaking.
Convene agents to model pricing options, argue them from customer, competitor, and margin angles, and hand a human a decision pack.
Expose data through MCP resources with stable URIs, useful listings, and sizes that fit a context window. Use when an agent needs to read data rather than perform an action.
Help users design effective organizational structures. Use when someone is thinking about team structure, deciding between functional vs.
Reason about how model performance improves with parameters, data, and compute, to allocate a training budget sensibly.
Decide between prompting, retrieval, and fine-tuning with eval-first discipline and honest data requirements. Use when someone proposes fine-tuning or a prompt has hit its ceiling.
Track every commitment from assignment to completion, chasing what is late and escalating what is stuck, so work finishes rather than fading.
Apply happens-before reasoning, volatile, and safe publication to write correct concurrent Java/Kotlin.
Verify a provider against its consumers' recorded pacts in CI so a breaking API change fails before it ships.
Serve media efficiently with caching, range requests, and access control, so playback starts fast and private files stay private.
Express organisational rules as automated checks in the pipeline rather than as documents people are meant to follow.
Use JSON Web Tokens safely by pinning the algorithm, keeping lifetimes short, and pairing them with a revocation path.
Pick an environment manager, pin dependencies with a lockfile, and produce reproducible installs that never touch system Python.
Lay out VPCs, subnets, private connectivity, and DNS so services reach each other privately and the internet only on purpose.
Scale training from one GPU to many by moving through data parallel and sharded FSDP modes, overlapping communication with compute, and reading the efficiency curve to find the…
Instrument search so queries, clicks, and abandonment tell you what to fix. Use when improving relevance without evidence, or when nobody can say whether search works.
Turn an empty result set into a useful next step instead of a dead end, and treat the query as a signal worth acting on. Use when searches return nothing and users leave.
Keep tool results small enough that an agent can hold what matters, through pagination, projection, and summarisation at the server. Use when tool output floods the context window.
Capture the current output of untested legacy code as a recorded baseline so a refactor can prove it changed nothing.
Catch fencepost bugs by writing down the boundary convention and testing the endpoints instead of the middle.
Implement retention tiers, legal holds, and deletion pipelines that actually delete, including GDPR-style erasure.
Work with an agent on a bug by supplying evidence and letting it form hypotheses, rather than asking it to guess from a description. Use when debugging with an agent's help.
Operate as a DevOps engineer who owns the delivery pipeline, keeps environments identical, and makes every deploy reversible.
Compose async JavaScript with promises and async/await correctly, propagating errors and cancelling with AbortController.
Place labeled, greppable print statements at decision points to trace the real execution, then remove them cleanly.
Predict when current capacity will be exhausted and plan the increase before it becomes an incident. Use when growth is steady enough to plan for or when scaling events surprise…
Operate as a data engineer who ships pipelines to a freshness SLA, enforces schema contracts at the source, and owns data quality end to end.
Keep an AI agent anchored to the actual goal through long or messy work. Use at the start of any multi-step task, and whenever work begins to drift, balloon, or stall.
Operate as a QA engineer who owns test strategy, automation coverage, and the release quality signal.
Manage the timing of money in and out so the business can pay what it owes when it is due. Use when growing fast, when payment terms are long, or whenever runway is under a year.
Tell a model what not to do in a way that works, by converting prohibitions into positive alternatives wherever possible.
Build multiplayer with an authoritative server, client prediction, reconciliation, and lag compensation. Use when adding networking to a game or fixing rubber-banding and desync.
Display, store, and reason about money across currencies without rounding errors or implied conversions.
Interpret what a user meant before matching, through normalisation, intent detection, and entity extraction.
Move user-facing text out of code into catalogs with stable keys and context, so translation becomes possible without touching logic.
Choose and write the right join so rows are neither lost nor multiplied, and know why a result set grew. Use when combining tables and the row count or the null columns look wrong.