harshnag
- 11 skills
- 0 followers
- 10 hours ago last updated
- ▌ Agent Handover · harshnag bundleHow to preserve task state and end a session without losing its reasoning — checkpoints, context-pressure recovery, and durable handovers rather than arbitrary token cutoffs. Use when assessing context usage or a restart rule, when compaction or missing constraints threaten continuity, when asked to wrap up or hand off, when creating a successor, when finishing work, or when writing, updating or reading a HANDOVER file.
- ▌ Checking Claims · harshnag bundleRules for deciding whether a claim is true before acting on it, especially claims about tooling, searches, censuses, CI, and test gates. Use when about to state how tooling works, interpreting an empty or partial result, quoting a count, searching for absence, choosing a corpus or denominator, reading a linter or reviewer finding, asserting a test catches something, or reviewing an inherited fact, cost, discouragement, or instruction.
- ▌ Measured Changes · harshnag bundlePractices for changing systems whose behaviour must be measured rather than judged, and for building gates that remain capable of catching failures. Use when tuning values, scores, ranking, or performance; designing a benchmark, corpus, baseline, control, denominator, or sample; adding or auditing a gate; interpreting a snapshot failure; setting payload, latency, verification, CI, or deploy budgets; or deciding whether a check belongs in a fast or slow gate.
- ▌ Shipping Changes · harshnag bundleSafely carry reviewed work from pull request through merge and deployment — pinning the exact head that passed, monitoring the exact post-merge run, applying migrations before an automatic deploy, verifying deployed identity, separating artifact trust domains, and interpreting skipped, queued, absent, or cancelled work honestly. Use when opening, updating, merging, or auto-merging a pull request; deploying to production; changing a schema; diagnosing a green merge that did not ship; designing deployment automation; or claiming which commit is live.
- ▌ Agent Concurrency · harshnag bundleProtocol for multiple AI agents working on one repository without destroying each other's work — isolated trees, visible claims, reviewed pull requests, safe landing, and session retirement. Use when running agents in parallel, delegating or creating a successor, claiming shared work, inspecting another ref, archiving a session or worktree, opening or merging a pull request, reviewing or approving work, setting branch protection, enabling scheduled agents, or being asked to finish pending work.
- ▌ Coordinating Agents · harshnagCoordinate a fleet of AI agents without treating stale session metadata, partial censuses, or delivered messages as ground truth. Use when orchestrating multiple workers, dispatching sub-agents or project sessions, deciding whether a worker is alive, issuing holds or corrections, monitoring a queue, archiving sessions, relaying findings between agents, or running scheduled and heartbeat agents.
- ▌ External Data Claims · harshnag bundleConsuming somebody else's dataset without inheriting their assumptions — publishing the selection predicate beside every count, hard-failing on unrecognised categories, keeping nothing-observed, too-few, withheld and a measured zero as distinct claims, and never letting a derived number travel without what produced it. Use when ingesting a third-party or public dataset, mapping source categories to your own, joining datasets, computing a score or index from data you did not collect, choosing thresholds or bands, deciding how to display missing data, or writing a figure into a document, deck or interface.
- ▌ Resource Safe Tooling · harshnag bundleDesign and run build, test, benchmark, concurrency, and process-control tooling without endangering a shared host or mistaking coordination for containment. Use when adding parallel workers, stress or load tests, repeated mutation loops, CPU-heavy CI, machine admission checks, process cleanup or reaping, background servers, PID or process-group signalling, concurrency tests, or any command that may consume most of a shared or low-resource machine.
- ▌ Durable Project Memory · harshnag bundleTurn engineering work into durable project memory without bloating every agent prompt — deciding what belongs in a reasoning document versus a repeatable skill, recording root causes and external research, correcting stale claims, and routing large documentation corpora. Use when closing an investigation, writing a design or research note, deciding whether to create a skill, preserving findings from web or vendor research, correcting advice already sent to another agent, compressing long documentation, or designing how agents find the right part of a large documentation set.
- ▌ Verify In The Real Thing · harshnag bundleWhy a passing test suite is not evidence that a built thing works, and what to do instead — driving the real artifact in the real environment, opening built artefacts rather than the documents about them, and handing back the checks you could not run. Use when about to call work done, when a suite passes and nothing has been opened, when reviewing or building a user interface, when writing a check for a rendering or interaction bug, when reading a README about an artefact, or when sandboxed or firewalled out of the environment the thing actually runs in.
- ▌ Project Agent Instructions · harshnag bundleWriting and maintaining the AGENTS.md that governs a repository — what belongs in it, what belongs in a linked document instead, how to route a large reasoning corpus, how to record inherited rules, and how to detect stale or missing injected instructions. Use when setting up agent instructions, editing or reviewing an AGENTS.md or CLAUDE.md, designing an agent read-in order, noticing instructions disagree with the repository or vanished after a workspace change, deciding whether a rule belongs in instructions or a doc, or onboarding agents onto a project.