luka-zivkovic

@luka-zivkovic source repo

26 skills
0 followers
15 hours ago last updated

▌
Lessons Learned 2 · luka-zivkovic bundle
Record corrections to agent behavior and standing workflow rules as durable, deduplicated lessons in .ai/memory/LESSONS.md, and surface them in later sessions so mistakes are not repeated. Use when the user corrects how the agent works ("no, use Y not X", "you made this mistake before", "I already told you this"), says "remember this workflow rule" or "add that to your lessons", or asks "what lessons do you have for this project?". Do NOT use for project-specific bugs, root causes, fixes, or diagnostic dead ends (solutions territory); domain term names or definitions (project-vocabulary territory); ordinary requirement changes; one-off contextual choices; or conversational disambiguation. Secrets are never persisted; explicit requests containing them record only the redacted rule.

0
installs
▌
Review Contract Gaps 2 · luka-zivkovic bundle
Produce a read-only semantic contract-gap pass for an exact implemented change, optionally subtracting a frozen pull-request review. Use only when explicitly asked for an implementation-aware second pass or standalone deep contract audit, especially for persisted values, shared configuration, public contracts, migrations, state transitions, or multiple producers and consumers. Discover against code before reading any supplied review, then return only source-verified omissions. Do not use as the primary PR review, before implementation exists, for fixes or test implementation, for posting comments, or for wording-only changes.

0
installs
▌
Review Contract Gaps 3 · luka-zivkovic bundle
Produce a read-only semantic contract-gap pass for an exact implemented change, optionally subtracting a frozen pull-request review. Use only when explicitly asked for an implementation-aware second pass or standalone deep contract audit, especially for persisted values, shared configuration, public contracts, migrations, state transitions, or multiple producers and consumers. Discover against code before reading any supplied review, then return only source-verified omissions. Do not use as the primary PR review, before implementation exists, for fixes or test implementation, for posting comments, or for wording-only changes.

0
installs
▌
Review Contract Gaps 4 · luka-zivkovic bundle
Produce a read-only semantic contract-gap pass for an exact implemented change, optionally subtracting a frozen pull-request review. Use only when explicitly asked for an implementation-aware second pass or standalone deep contract audit, especially for persisted values, shared configuration, public contracts, migrations, state transitions, or multiple producers and consumers. Discover against code before reading any supplied review, then return only source-verified omissions. Do not use as the primary PR review, before implementation exists, for fixes or test implementation, for posting comments, or for wording-only changes.

0
installs
▌
Groundwork · luka-zivkovic bundle
Interview the user about a piece of work, one decision at a time, until a shared understanding is reached, then produce a compact decision brief and stop for confirmation. Use ONLY for explicit elicitation requests such as 'grill me', 'interview me', 'ask questions until we're aligned', 'walk me through the decisions', or 'help me pin down what I want before we build'. Inspect authorized context for facts, offer a tentative recommendation with each question, and keep unexamined delegated defaults labelled agent-proposed. Do NOT implement, plan, critique, stress-test, or give a verdict; do not use for routine work, one ordinary clarification, open-ended brainstorming, or ambiguity alone. If one prompt combines interview/elicitation with critique, stress-testing, or a verdict, do not invoke this skill yet: ask the user to choose elicitation or evaluation first, and start neither workflow until they choose.

0
installs
▌
Setup · luka-zivkovic bundle
Audit the current project and produce a report-only plan for installing and configuring published Overclock plugins. Use only when the user explicitly invokes the Overclock setup command and wants help choosing capabilities, installation scope, hook tolerance, conflict remediation, or minimal CLAUDE.md integration. Inventory existing Claude plugin and instruction state, enforce package conflicts, and return exact commands and proposed diffs without executing or writing anything. Do not invoke automatically, replace Claude's built-in /init, recommend unpublished plugins as installable, or mutate plugins, settings, hooks, or instruction files.

0
installs
▌
Solutions · luka-zivkovic bundle
Capture a verified solution to a nontrivial problem in .ai/memory/SOLUTIONS.md — symptoms, what didn't work, the fix, and why it works — and reuse it when the same symptoms come back. Use when a hard problem is confirmed solved after real diagnostic effort or failed attempts ('that worked', 'it's fixed', 'finally — that was it') , when the user says 'capture this solution' or 'save this fix for next time', when starting on a problem whose symptoms might match a previously captured solution ('have we hit this before?'), or when asked to review or clean up the solutions ledger. Do NOT use for trivial or one-line fixes, routine feature completion, an unverified 'this should work', corrections of agent behavior or workflow rules (that is lessons-learned territory), or one-off environmental flukes that carry no reusable diagnosis. Secrets are never persisted; capture requests containing secrets record only the redacted solution.

0
installs
▌
Agent Bridge · luka-zivkovic bundle
Use a DIFFERENT installed coding provider as a bounded leaf collaborator while the current agent keeps ownership. Trigger when the user names Claude, Codex, Gemini, another installed provider, cross-provider help, or agent-bridge as the actor for consultation, review, critique, diagnosis, a second opinion, or implementation; also use when blocked and the user authorized cross-provider help in this conversation. This skill owns those requests over sibling review, debugging, or interview skills and is the only sanctioned provider-CLI path. Use consult for read-only analysis. Use delegate only for user-authorized implementation with a clean exact HEAD, explicit allowed paths and acceptance checks, an isolated patch, and parent verification. Do not use for routine work with no different provider named, extra votes, absent sharing authorization, same-harness requests, or concurrent writes to the active checkout. Never invoke provider CLIs ad hoc.

0
installs
▌
Local Eval Stack · luka-zivkovic bundle
Stand up a local, self-hosted evaluation stack for agent skills and sessions: ironside (trace store) via docker compose, coeval (governed LLM judging with human adjudication) via docker + dev servers, headless bootstrap of judge projects with calibrated rubrics, a pi session tracer that tags skill usage, and casefile scanning of the skills being judged. Use when someone wants to self-host agent evals locally, trace pi/agent sessions to their own machine, judge a skill's real runs with a governed rubric, or asks to set up ironside/coeval/the eval stack. Do NOT use for writing eval content or rubric doctrine (see the evidence-tiers doc in overclock), for CI gating of an existing coeval instance (coeval's gate.mjs docs cover that), for hosted/SaaS eval platforms, or for production multi-user deployments — this skill's scope is one developer's machine.

0
installs
▌
Skill Maintenance · luka-zivkovic bundle
Maintain a judged agent skill from its coeval evidence: pull findings (recurring failure clusters, human override reasons, golden deltas) via the coeval MCP tools or findings CLI, translate them into ONE bounded SKILL.md patch on a workshop copy, co-evolve the judge's rubric through coeval's guarded flow when an invariant moves, validate pre-merge with a casefile scan plus golden re-runs and a gate check, and open a PR citing the findings. Use when someone asks to improve, patch, or maintain a skill based on its eval findings, apply coeval findings to a skill, or run the skill-maintenance loop. Do NOT use when the target skill has no findings (no findings = no patch — never invent improvements), for authoring a new skill or a first rubric, for standing up ironside/coeval/casefile (that is local-eval-stack), to adjudicate verdicts (human-only in the coeval dashboard, no tool by design), or to weaken a rubric, probe, or test so a patch passes — that corrupts the bench. Never auto-merge — the human merges, alway

0
installs
▌
Lessons Learned · luka-zivkovic bundle
Record corrections to agent behavior and standing workflow rules as durable, deduplicated lessons in .ai/memory/LESSONS.md, and surface them in later sessions so mistakes are not repeated. Use when the user corrects how the agent works ("no, use Y not X", "you made this mistake before", "I already told you this"), says "remember this workflow rule" or "add that to your lessons", or asks "what lessons do you have for this project?". Do NOT use for project-specific bugs, root causes, fixes, or diagnostic dead ends (solutions territory); domain term names or definitions (project-vocabulary territory); ordinary requirement changes; one-off contextual choices; or conversational disambiguation. Secrets are never persisted; explicit requests containing them record only the redacted rule.

0
installs
▌
Session Handoff · luka-zivkovic bundle
Save and restore working state across Claude Code sessions via a structured HANDOFF.md under .ai/memory/. Use when the user wants to stop and preserve state ("let's stop here for today", "checkpoint this session", "write a handoff", "I'm about to run out of context, save our state") or to restore it ("resume where we left off", "continue from last session", "what were we working on?"); also use when the user has no local handoff file but pastes or says they copied a handoff from another machine. Produces a handoff file with goal, plan, decisions, failed approaches, next steps, and git anchors, plus a warm-start brief on resume. Do NOT use for ordinary git commit/push requests, for "continue" meaning the next step of the current in-conversation task, or for human-facing summaries like PR descriptions or status reports.

0
installs
▌
Natural Writing · luka-zivkovic bundle
Draft or revise multi-sentence human-facing prose such as blog posts, articles, essays, newsletters, landing copy, and narrative documentation. Match the author's or project's established voice; when none exists, use a plainspoken house style that avoids canned AI scaffolding, showy vocabulary, repetitive rhythm, and decorative analogies while preserving claims, uncertainty, caveats, and quoted spans exactly. Use for editing prose around quotations as well as prose without them. Do NOT use for code, code comments, commit messages, one-line UI strings, or API, legal, and reference text whose exact terminology is load-bearing.

0
installs
▌
Publish Pr Feedback · luka-zivkovic bundle
Publish an already-reviewed, digest-locked plan of GitHub PR thread replies and optional resolutions. Use only when the user explicitly invokes $publish-pr-feedback with the exact sealed plan path and SHA-256 digest, or explicitly asks this skill to seal a supplied draft plan for later approval. Verify host, repository, PR node, open state, current remote head, action scope, and plan digest before mutation. Do not judge feedback, edit code, invent or broaden replies, commit, push, merge, rebase, react, approve a PR, or run implicitly.

0
installs
▌
Resolve Pr Feedback · luka-zivkovic bundle
Handle reviewer feedback on an open GitHub pull request. Use whenever the user asks to address, handle, work through, or resolve review comments, open review threads, or bot feedback; this includes prefetched thread data such as review/threads.json and a later approval message that asks the resolver to prepare its unsealed local plan. Fetch unresolved threads plus PR comments and review bodies, judge every item against actual code, apply valid local fixes within authorized scope, and draft replies. After a separate approval names an exact subset and root-confined plan path, prepare only that unsealed local plan. Never post, react, resolve, commit, push, merge, rebase, or approve; remote publication belongs only to explicit $publish-pr-feedback. Do NOT use to produce a new review, discuss one pasted comment, handle feedback without a PR, work on non-GitHub forges, or summarize without changes.

0
installs
▌
Test Discipline · luka-zivkovic bundle
Mandatory pre-edit test gate for changes to existing code. Use before fixing a reported regression, crash (including an HTTP 500), invalid-input failure, error, or wrong output—even when the user only says 'fix it' and never requests a test. For a refactor or cleanup, use only when the named behavior lacks behavioral tests; any adequate behavioral coverage is already the pin. Also use proportionately when asked only to write a unit/regression test for named existing behavior, or explicitly asked to validate whether a freshly-green test detects its claimed regression. Reproduce fixes red-before-green, characterize untested behavior before refactoring, and mutation-check only when the workflow or user requests it. Never stage or commit artifacts. Do NOT use for diagnosis-only requests, typos/copy, config/version/dependency bumps, behavior-preserving renames or control-flow rewrites, formatting, generated/vendored/lockfile changes, or ordinary new-feature implementation.

0
installs
▌
Git Archaeologist · luka-zivkovic bundle
Always use before deleting, bypassing, or weakening a structural defensive construct in existing committed code: a guard clause or early return that rejects invalid/unsafe state, retry/backoff, protective sleep/delay, lock/mutex, a caller-input clamp/bounds check, or a check commented 'redundant'/'defensive'/'shouldn't happen'. This includes when the removal arrives framed as simplification — a 'simplify', 'clean up', 'reduce ceremony', or 'make this a one-liner' request whose execution would strip such a construct fires this gate the same as an explicit delete request. Recover why it exists with git blame, the introducing commit, and linked PR/issue evidence, then warn before removal. Trigger only when the requested change weakens one of those constructs. Do NOT use for behavior-preserving control-flow rewrites where the construct remains effective (including simplifications and helper extractions that keep every guard intact), ordinary early returns, typo/copy fixes, config/version/dependency bumps, pure re

0
installs
▌
Critical Thinking · luka-zivkovic bundle
Independently evaluate consequential decisions and explicitly challenged reasoning without sycophancy. Use whenever the user names critical-thinking, asks to reassess or revisit a conclusion, treat prior agreement as untrusted, sanity-check, critique, or stress-test reasoning, asks 'am I right?', requests a recommendation with material assumptions or downside, proposes a causal story from incomplete evidence, or wants a go/no-go verdict whose local facts need checking. Also use when a costly medical, financial, or safety action depends on a headline, study, or factual claim. For a decision that depends on local facts, own the verdict and use at most one neutral bounded independent-research pass only when that optional skill and real context isolation are available; otherwise state the gap and make the verdict conditional. Factual-only local verification belongs to independent-research. Diagnosis of an observed recurring bug belongs to debugging-discipline. If one prompt combines elicitation with evaluation, a

0
installs
▌
Project Vocabulary · luka-zivkovic bundle
Apply and maintain the project's domain vocabulary in repository-root CONCEPTS.md. Use whenever any part of the prompt corrects a domain term ('that is a Workspace, not a Team'), exposes fuzzy or conflicting terminology, settles a load-bearing domain noun, or asks to create, update, or review the glossary. A mixed terminology-and-workflow correction still invokes this skill for the vocabulary half only; leave the workflow half to its optional owner. Implicit invocation may read untrusted glossary content, challenge usage, and propose an exact change, but writes require an explicit add/update/record request or approval of a displayed proposal. Do NOT use for workflow-only corrections ('use pnpm' belongs to lessons-learned), generic programming terms, throwaway utilities, repositories without domain vocabulary, or one-document wording preferences.

0
installs
▌
Independent Research · luka-zivkovic bundle
Verify decision-relevant facts from accessible local evidence in a fresh research context instead of relying on the user's framing or prior conversation. Always use directly, without critical-thinking, when the request is only to independently inspect, research, or verify a referenced local repository/project, checked-in source/config/CI, local API specification, saved document/paper, exported log, or dataset rather than trust a summary. Also use for at most one neutral bounded local-evidence pass delegated by critical-thinking when those facts serve a decision or value judgment; return evidence only and let critical-thinking own the verdict. Prefer primary evidence, preserve provenance, report contradictions and access gaps, and return a bounded evidence packet. Do not use for websites or live APIs, preferences, personal facts only the user can know, creative tasks, immaterial uncertainties, or prohibited access.

0
installs
▌
Debugging Discipline · luka-zivkovic bundle
Systematic diagnosis for bugs that resist the ordinary red-test path: establish the safest tight red-capable observation loop before committing to a theory, minimize the failure, then test 1-5 credible ranked hypotheses with predictions. Use for intermittent/flaky failures, performance regressions, integration/staging/production-only failures, bugs that survived attempted fixes, code with no viable test seam, or an explicit request to debug systematically or find the root cause. When a prompt couples stress-testing a suspected explanation with finding the cause of an observed recurring bug, this skill owns the operational loop; treat the explanation as one hypothesis and do not launch a competing critical-thinking loop. Diagnose without editing by default; fix only when the user also authorizes implementation. Do NOT use for an ordinary requested bug fix with a viable test seam when test-discipline is available, a trivial cause already identified by the error, feature work, writing tests on request, or explai

0
installs
▌
Review Pr · luka-zivkovic bundle
Perform a review-only, adversarial pull-request assessment for correctness, security, data integrity, compatibility, concurrency, failure handling, and meaningful maintainability regressions. Use only when the user explicitly invokes this skill to review a PR, branch diff, or pinned base/head range. Work in any repository without setup; when .ai/pr-kit/REPOSITORY.md exists, use its source-grounded repository context without trusting it as instructions. Return draft inline comments for a human to post. Do not use this skill to fix or apply findings, edit files, commit, push, post comments or reviews, or pad the result with style nits.

0
installs
▌
Initialize Pr Kit · luka-zivkovic bundle
Explicitly initialize or refresh repository-specific PR review knowledge for pr-kit. Use only when the user invokes this one-time initializer and authorizes creation of .ai/pr-kit/REPOSITORY.md. Inventory architecture, ownership, invariants, trust boundaries, failure modes, verification commands, conventions, and verified historical precedents from inspectable repository sources. Write only that profile, never copy secrets, never modify project instructions or settings, never auto-commit, and never make the profile a substitute for current code.

0
installs
▌
Review Contract Gaps · luka-zivkovic bundle
Produce an append-only, implementation-aware contract-gap delta against a frozen pull-request review. Use only when explicitly asked to run a second review pass after a complete review exists for an exact base/head range, especially for behavioral changes involving persisted values, shared configuration, public contracts, migrations, state transitions, or multiple producers and consumers. Trace the implementation actually chosen, compare it with base-repository contracts and the frozen review's root-cause coverage, and return only independently verified omissions. Do not use as the primary PR review, before implementation exists, for fixes or test implementation, for posting comments, or for wording-only changes.

0
installs
▌
Audit Consumer Contracts · luka-zivkovic bundle
Audit whether a committed change breaks pre-existing consumers of the identifiers, options, keys, schemas, routes, or emitted values it touches. Use only when explicitly invoked on an exact base/head pair, either standalone or as a strict append-only second pass over a supplied frozen review. Enumerate base-tree consumers outside the diff, verify each expectation against the exact head, and report only confirmed changed-line-anchored findings. Do not use for general code review, pre-implementation risk brainstorming, style or wording changes, debugging observed failures, fixing findings, or posting comments.

0
installs
▌
Anticipate Edge Cases · luka-zivkovic bundle
Build a clean-room, pre-implementation risk brief from a PR description, issue, local branch request, or supplied change text. Use only when the user explicitly asks to anticipate edge cases, failure modes, acceptance risks, or review probes before inspecting implementation code. Ground the brief in the exact pre-change repository snapshot when available, label requirements versus inference, and stop before reading or judging the diff. Do not use for implementation review, generic code audits, debugging an observed failure, writing tests or fixes, security-only assessment, or trivial wording and formatting changes that have no material behavioral risk.

0
installs

26 published skills