nadavnaveh-skills

Plugin nadavnaveh-skills from marketplace nadavnaveh in nadavnaveh/skills. Commands read from the repo's manifest — 2 days ago.

Skills (104)

• access-levels

  Sets how much the agent may do on its own across five levels (read only, sandbox edits, ask first, safe auto-run, production), starts at the lowest level that can do the job, and moves up one level only when the owner says so. Fixes "I gave the agent full access on day one, and it changed things I n

• agent-failures

  Watches a task for the five ways agents fail (Drift, Sure and wrong, Loops, Forgets, Overreach), catches each one at the step where it starts, and steers back with one move per failure. Fixes "the agent went wrong, and I only noticed at the end." Use at the start of any task you hand to an agent, wh

• agent-guard

  Sets three guards before an agent works near your files. Delete commands wait for your approval, the agent works only inside the project folder, and what matters is saved in git and in a separate backup. Fixes "I asked the agent to tidy a folder, and years of my files were gone." Use before an agent

• agent-hooks

  Moves one rule the agent keeps breaking from AGENTS.md or CLAUDE.md into a git pre-commit hook that checks the staged diff, refuses the commit and prints the exact fix in words. Fixes "I wrote the rule in AGENTS.md, and the agent keeps breaking it." Use when the agent breaks the same written rule tw

• agent-limits

  Maps which steps your agent may take alone, from its real config, and writes one LIMITS.md: each action, whether it can be undone, its tier, the stop outside the agent, and whether that stop was tested, plus a halt plan. Fixes "I can't say which steps my agent takes alone, or what would stop it." Us

• agent-loop

  Runs every task as the agent loop (plan the next step, act with one tool call, observe the result, check that it worked, then stop or go another round) and never skips the check. Fixes "the agent made the change and never checked if it worked." Use on any task that changes something (fix a bug, add

• agent-patterns

  Picks the right way to run several agents on one job (Parallel, Debate, Pair, Pipeline or Manager), says why, and runs it in Claude Code with subagents. Fixes "I opened five agents and got five messes." Use when a task is big enough to split, when you want a second opinion before deciding, when work

• agent-wiki

  Keeps a wiki folder the agent writes to as it reads. It makes one page per topic, gives every claim a link to its source, edits the right page when a new source arrives, keeps an index.md that lists every page, and runs a weekly review that deletes what went stale. Fixes "The agent read it last week

• ai-in-business

  Maps one business process step by step, finds the step done most often, builds a small agent for that step only, measures the hours before and after, and gives the agent a named owner with a weekly check. Fixes "We added AI to the business, and nobody can say what it saved." Use when someone wants t

• answer-checks

  Checks an answer before you act on it, using five checks. It opens the source, makes the agent argue against its own answer, asks a fresh session that never saw the answer, runs the method on a case where the result is already known, and sets how hard to check by what an error would cost. Fixes "It

• ask-first

  Before the agent writes code on a vague task, it asks until the goal, the why and what done looks like are clear, asks every decision it would otherwise guess with a suggested answer, and writes your answers as one definition with a yes/no check. Fixes "I asked for one line of work, and it made seve

• automate-questions

  Asks five questions before a task is automated (does it repeat, can you check it, what if it fails, who owns it, is the data ready) and builds the job only when every one has an answer. Fixes "I automated it, and now I spend more time fixing it than doing it." Use before you set up a scheduled job,

• batch-judge

  Speeds up an AI judge that reads the same text many times. It keeps every check in one prompt with one JSON answer, pastes the files in, and cuts the repeat reads only after old and new give the same verdicts on examples you already know. Fixes "my AI judge is slow, and it keeps reading the same tex

• big-archive

  Splits a big archive (hundreds or thousands of pages) into numbered parts, gives one part to each subagent, reads all parts at the same time, makes every finding cite its page, merges the lists into one with duplicates gone, and spot-checks two cited pages at random. Fixes "I gave it a thousand page

• brief-parts

  Writes a five-part brief before the agent starts: a one-line goal, what done looks like, what must not change, the key files, and the test that proves it works. Fixes "I explained the task five times before the agent got it right." Use before you hand any task to an agent or a subagent, when a reque

• bug-triage

  Checks a list of bug findings and keeps only the real ones. It reproduces each one, has a second agent try to disprove it, merges findings with one cause, writes what a user sees when it breaks, and starts each fix with a failing test. Fixes "the agent found ten bugs, and most of them were not real.

• build-an-eval

  Builds a small test from ten of your own past tasks, writes the right answer for each one before any run, and scores the agent pass or fail out of ten every time you change something. Fixes "I changed the prompt, and I can't tell if it got better or worse." Use when you are about to change a prompt,

• calls-to-keep

  Marks the five calls a person keeps when an agent works alone (the goal, any step that cannot be undone, the merge into main, spending money or using keys, and the final look before release) and makes the agent stop and ask at each one. Fixes "the agent made a call that should have been mine." Use w

• check-ways

  Checks an agent's work in five ways before you trust it. It runs the tests, has a second agent review the work, sets a gate the work must pass before it counts as done, reads the diff line by line, and tries the result the way a user would. Fixes "I trusted the agent's work, and it broke the first t

• checkpoints

  Commits a clean point before a long job starts, keeps a progress.md file that says what is done, what runs now and what is next, and works on a branch with one commit per step. If the job crashes, it starts again at the step that crashed. If a step breaks things, one git command takes the work back

• claude-md-rules

  Writes and trims CLAUDE.md, the file your agent reads at the start of every session. It keeps only specific lines, gives each rule its reason, writes down the exact build, test and lint commands, says what to leave alone, and adds a rule when you correct the agent on the same thing twice. Fixes "I k

• cloud-offload

  Moves your agents' heavy work off your computer into Claude Code cloud sessions on Anthropic's machines. Fixes "my laptop freezes with several agents running." Use when your computer slows down with many agents, before a long build or test run, or when you want work to keep going after you close the

• code-map

  Maps old code before anyone changes it. It finds where every request starts, follows the data step by step, counts who calls the function you want to change, marks the risky parts and checks what is tested. Fixes "I changed one small thing in old code, and something far away broke." Use it before yo

• command-guard

  Adds a hook that blocks the dangerous shell commands on its list (rm -rf, force push, git reset --hard, DROP TABLE) before they run in Claude Code, sends the reason back to the agent, and tests it on four commands. Fixes "I told the agent to be careful, and that was the only thing stopping it." Use

• compact-safely

  Saves your decisions and exact numbers to a file before /compact, tells /compact what to keep, then reloads the rules and asks about the goal before the work goes on. Fixes "I compacted, and it forgot what we decided." Use when a session is long and you are about to run /compact, when Claude Code wa

• context-rules

  Shows how the context window fills and keeps it small. It compacts early with a keep list, sends big reads to a subagent, keeps the opening the same so it is cached, and counts the room each tool call takes. Fixes "the session got slow and forgetful, and I don't know why." Use when a session has run

• corrected-twice

  Stops patching the output after the second correction on the same thing. It writes the correction as a rule in CLAUDE.md, finds the step in the plan that went wrong, redoes that step on one file with a check, and starts a fresh session with the new rule loaded. Fixes "I keep correcting the same thin

• count-check

  Makes your agent end every count with its rule, three random rows and how many rows the rule could not read, and stop if more than 10% went unread. Fixes "the count was wrong, and it looked right." Use whenever the agent gives a count or total from a sheet, a CRM, a log or a folder, and before you a

• data-leaks

  Checks the five places data slips out of an agent's work (logs, screen captures, prompts to tools, pages it reads, and commits), closes each one, and reports each as closed, open or not checked. Fixes "My key ended up somewhere I never meant to send it." Use before you give the agent a key or passwo

• day-gate

  Puts all of a day's packages in one list, gets one yes from you for the whole day, then runs each approved item at its slot without asking and holds anything you did not approve. Fixes "my agent waited for my approval, and the post missed its slot." Use when agents post, send or publish on a schedul

• disprove-findings

  Hands an agent's bug list to a fresh agent without the first one's history, has it try to prove each finding is not a bug by opening the code, and reports each one as not a bug, almost certainly fine or still standing, with the evidence. It fixes nothing. Fixes "the agent found bugs, and I started f

• dogfood-first

  Runs a new skill, prompt or agent setup on your own real tasks for this week before anyone else gets it. It logs one line for each thing it got wrong, fixes only the miss that keeps coming back, runs the same tasks again to count the misses, and shares the tested version only after it has held up. F

• done-test

  Writes one to three yes/no checks before a task starts, then runs each one at the end and reports pass or fail with evidence. Fixes "the agent said it was done, and it wasn't." Use at the start of any task that builds, changes or sends something (add a feature, fix a bug, make a page, write a docume

• draft-only-mail

  Sets email access to draft-only. The agent reads mail and writes drafts, every send tool or route is denied in Claude Code's permission settings, and the block is tested. Fixes "the agent sent an email in my name without asking." Use when you connect email to an agent, before it works in your inbox,

• drift-check

  Checks a working agent session for five signs of drift (the same fix twice, edits to files you never named, padding, a lost decision, and "done" before any check), shows the evidence for each one it finds, and steers it back with one move per sign. Fixes "it started well, and somewhere along the way

• fail-closed

  Makes every guard block the action when it breaks, instead of letting it through. Fixes "the safety check broke, and the agent went ahead anyway." Use when you write or change a hook, report a test result, wait for a yes before a deploy or push, set which commands may run, or find a key or setting m

• failed-reads

  Makes every automated report list its sources, mark each one read or failed with the real error, keep the failed ones in a file for the next run, and end with one line naming what it could not read. Fixes "My agent said nothing new, and it just couldn't read the source." Use when you build or run a

• faster-same

  The text you paste into your agent when it offers to make code faster, so it keeps the original, compares old and new on at least 1,000 inputs, times both and adds a switch back, and you own the check. Fixes "The agent offered to make my code faster, and I won't read every line." Use when an agent o

• first-message

  Turns a loose ask into a first message with the files, the exact change and a done test, and keeps one task per conversation. Fixes "My allowance runs out long before my work does." Use before you start a change in Claude Code, when you are about to send a short general ask like "add a reset button"

• fix-review

  Reads the change an agent made to turn a failing check green, asks whether it solves the problem or only moves the number, and adds a check that catches the shortcut. Fixes "the agent fixed the failing check, it's green, and the work got worse." Use whenever an agent says a failing test, gate, lint

• handoff-note

  Writes a four-part handoff note before a session closes, checks each claim against the code and the repo, marks the rest as guesses, then tests the note on a fresh agent. Fixes "The session ended, and everything it understood was gone." Use before you close a session or run /clear, when you stop in

• handoff-rules

  Writes a HANDOFF.md when one agent stops and the next one takes over. The file records what was done, what is still open, what was tested and how, and who holds the task now. The next agent reads it before it acts. Fixes "the new agent started over, redid finished work and trusted a test nobody ran.

• hard-problems

  Attacks a problem nobody has solved yet. It writes many guesses down, runs one cheap test per guess, logs every try (failures included), keeps only the guesses that survived and proves the answer with a result someone else can rerun. Fixes "The agent tried one idea after another, and we are still st

• harness-starter

  Installs the ready kit for your own harness in a Claude Code project, a hook before every command you can't undo, a queue for heavy jobs, and tests for that code itself. Fixes "I want my own harness, and all I have is rules in a prompt." Use when you want a harness and don't know where to start, whe

• hidden-text-check

  Checks a document someone else wrote for text a person cannot see, before an agent reviews it. It reads a plain text copy, has the agent list every instruction inside without doing any, and runs a finder for white, tiny or hidden text in Word files. Fixes "the AI read a line in the document that I c

• hooks-to-add

  Adds five Claude Code hooks, small commands that run by themselves at the right moment. They run the tests after each edit, block risky commands before they run, format every file the agent writes, ping you when the agent needs you, and log every command it runs. Then it tests each one. Fixes "I kee

• injection-guards

  Sets five guards so text the agent reads cannot turn into actions. The agent reads pages as data and not as orders, it cannot open your keys, it runs only the commands you named, it asks before it sends anything, and it works inside a sandbox. Fixes "a web page told my agent what to do, and it did i

• inventory-first

  Before any job with several agents, lists what already exists (folders, finished outputs, tickets) with dates and counts, shows you the list and asks what is new, then starts agents only for the new items you approved. Fixes "I sent an agent to do work that was already done." Use before you start mo

• keys-and-agents

  Keeps keys, tokens and passwords out of the chat and out of git, scoped to the task, read-only where it can be, and replaced the day it leaks. Fixes "I ran the agent in a sandbox, and it could still see my keys." Use when a task needs an API key, token, password or database URL, or when one shows up

• live-check

  Opens the live address itself, reads which version is really running there, and walks one real user path start to end before anyone says production is restored. Fixes "the agent said production was restored, and users still got errors." Use after any deploy, rollback, hotfix or restore, whenever a s

• lock-on-failure

  Lists every hook in your Claude Code settings, sorts the guards from the hooks that only report, adds onFailure "block" to each guard after your yes, then makes each guard fail once on purpose in a throwaway project and checks that the action stopped. Fixes "my guard crashed, and the command ran any

• long-videos

  Works through an hour of video cheaply. It turns the video into a timed transcript first, searches the words for the minute that matters, looks at one frame every 10 seconds, cuts a short clip for a close look, and puts a timestamp on every claim. Fixes "I gave it an hour of video, and it cost a for

• mcp-parts

  Connects an agent to your tools through MCP in five parts (the server, the call, the result, permissions and many agents). Fixes "I connected my tools to the agent, and I don't know what it can reach or what it did with them." Use when you add an MCP server, when the agent used a tool you did not ex

• memory-queue

  Sends every heavy job (builds, tests, renders) through one small queue script that caps how many run at once and waits until at least thirty percent of memory is free before each one starts. Fixes "I run five agents at once, and my Mac runs out of memory, so jobs crash or crawl." Use when you run se

• memory-sources

  Sets up the five places an agent reads before it starts (the CLAUDE.md rules file, decision notes, ticket history, a progress file and a folder of good examples) and makes it read them first. Fixes "every new session starts from zero, and it tries what we already ruled out." Use at the start of a pr

• memory-trim

  Measures the agent's memory index (MEMORY.md) against what Claude Code loads, lists the rules past the cut, and rebuilds it as one short line per rule, most important first. Fixes "the agent ignores a rule I wrote, as if it never saw it." Use when the agent breaks a rule that is in its memory, when

• night-alarm

  Lists what can stop a night run (free disk, credits or quota, permissions) with each current value and the line at which it matters, writes a check, schedules it every thirty minutes so it wakes the agent in charge, and tests it once by tripping a line. Fixes "the run stopped at night, and the repor

• objector

  Adds an objector to the agent's decisions. Before a decision that costs money or is hard to undo, a fresh subagent that did not make the decision argues that it is wrong, point by point. The first agent answers each point and changes the decision only for a fact it can check. Fixes "The agent made a

• old-code-judge

  Makes the old code the judge of a rewrite. It saves the old code's exact output on real inputs, runs the new code on the same inputs, compares them byte for byte and fixes the new code until there are 0 differences. Fixes "the agent rewrote my code, and I can't prove it still does the same thing." U

• one-task-gate

  Builds a gate for one task you give your agent often. It writes what good means as 5 to 8 yes-or-no questions before the work, then a separate checker with no access to the conversation answers each one from the result alone, quoting the line that proves it. Fixes "my agent checks its own work, and

• past-green-tests

  Checks the design of the code after the tests pass. It looks at which layer imports which, how big each file is, where the same logic is copied and which names say nothing, then gives a person a short list to check the shape. Fixes "the tests were green, and the code was still a mess." Use after the

• phone-rules

  Sets the rules for running an agent from your phone: reads run without asking, every change waits for your yes, reports come in three short lines, a notification arrives when the job ends, and nothing ships until you are at your desk. Fixes "I left the agent running, and I couldn't tell from my phon

• plan-first

  Decides whether a task needs a plan before any change, writes that plan and waits for your yes. It maps a new codebase, agrees the goal, numbers every file, marks the steps that are hard to undo, and splits a big refactor into steps you approve one at a time. Fixes "It started changing things before

• prompt-upgrades

  Upgrades a request before the agent starts on it. It has you talk instead of typing one line, asks who the answer is for, offers three options to pick from, agrees the shape of the answer, and asks for the real source instead of your summary. Fixes "I typed one line and got a generic answer back." U

• prompt-ways

  Adds five small habits to a request before the agent starts: one good example, who it is for, a plan first, a check it can run, and a list of what it could not confirm. Fixes "the answer was close, and I still had to fix it myself." Use when you write a request for any task (a piece of writing, a co

• quote-check

  Checks every quote in a draft against its source. The writing agent pastes the source's sentence and the link, a different agent opens the link and searches those exact words, and each quote is reported as exact, reworded or unverified, with the line found. Fixes "The agent put it in quote marks, an

• read-filing

  Reads any filing, annual report or earnings release with three questions (what is in the loss, what operations earn, what the company is committed to pay) and writes a one-page reading where every number is quoted with its location. Fixes "A company I depend on posted a scary loss, and I can't tell

• redact-first

  Checks a screen recording, demo or screenshot for the five things that leak (email addresses, API keys, real names, open tabs and notifications), clears them before you record, and blurs or replaces anything left before you post. Fixes "my demo showed my inbox." Use before you record a demo, take a

• rejections-to-tests

  Turns every result you reject into one sentence on what was wrong and why, turns that sentence into a yes/no question someone else could answer, and has the agent run all the questions on every new result before you look. Fixes "I keep rejecting the same mistake, and the agent keeps making it." Use

• report-endings

  Makes every report an agent writes from its sources end the same way. It lists what it read, names each source that failed and why, shows only what changed since the last run, lists what it skipped and why, and puts the one decision that needs you at the top. Fixes "The report said nothing new, and

• review-habits

  Reviews code an agent wrote with five habits: keep the change small, read the diff, run it yourself, try empty, huge and odd inputs, and list what no test covers. Fixes "I approved the agent's code after a quick look, and it broke later." Use before you accept, commit or merge any change an agent ma

• rubric-numbers

  Rewrites every check you give an AI judge as a number with a unit and a pass/fail line, makes the judge name the failing element and its value, and measures in code what code can measure. Fixes "I asked the AI to check the same thing twice, and it gave me opposite answers." Use when you ask a model

• rule-count

  Runs one small real task three times with your CLAUDE.md and three times without it, in a throwaway copy of the repo, and counts from the logs which lines changed what the agent did. Fixes "I wrote the rules, and I can't tell which ones it follows." Use when you add or rewrite lines in CLAUDE.md, wh

• safe-delete

  Puts three habits in front of any script that deletes files. It pushes the work to a remote first, runs a dry run that only counts and lists, and rehearses the delete on a copy before the real folder is touched. Fixes "I let the agent clean up, and it deleted my project." Use before an agent writes

• save-points

  Commits a save point before the agent starts work and before you close the laptop, accepts the agent's pending edits in Cursor, and runs git status when you open the project, so emptied files come back with git restore. Fixes "I opened my project, and my files were empty." Use before you hand an age

• second-ear

  Turns every AI-generated voice take back into text, compares it with the script word by word, regenerates any take with missing, extra or repeated words, and sends the spots the check cannot hear to a person. Fixes "the AI voice skipped a line, and nobody noticed until it was posted." Use whenever a

• second-pass

  Adds the second pass to every coding task. After the agent writes code, it lists the odd cases the request implies, writes a check for each, runs those checks and the project's tests, fixes and reruns for up to three rounds, then reports what was checked, what passed and what was not checked. Fixes

• server-lockdown

  Takes a rented server (VPS) off the public internet by closing its two doors, following @levelsio's routine. Fixes "My server is online, and I don't know which doors are open." Use when you have a VPS with a site or app on it, when you set up a new VPS, or before you trust an SSH key alone. Also on

• ship-check

  Compares what you approved with what's about to ship, file by file, and stops the send on any difference. Fixes "I approved the right file, and yesterday's copy was about to go out." Use when you approve files for a later send, when a post, email, upload or release sends from a staging or copy folde

• ship-steps

  Takes a change from a prompt to a safe merge in five steps. It agrees the plan, builds in a branch, runs the tests, puts the diff in front of you for a second look, and merges only after you say yes. Fixes "The agent changed main, and I never saw what went in." Use when you ask for a feature, a fix

• skill-button
• speedup-guard
• stack-compare
• stage-clock
• stale-memory
• stamp-check
• step-cache
• stop-rule
• stop-signals
• stranger-test
• subagent-jobs
• taste-skill
• teachability-check
• team-roles
• tests-as-spec
• tests-to-run
• to-gantt
• token-savers
• update-sweep
• user-path-test
• vendor-claims
• what-do-i-want
• write-skills
• write-the-check

Similar plugins