thananon/9arm-skills
6 skills17k installs18.4k starsGitHub
Install
npx skills add https://github.com/thananon/9arm-skillsSkills in this repo
1Scrutinizescrutinize performs an outsider-perspective end-to-end review of plans, pull requests, diffs, design docs, and proposed code changes. It triggers on /scrutinize or when users ask to review, audit, sanity-check, or get a second opinion. The workflow runs four ordered steps without skipping: Intent restates the goal in one sentence and asks whether a simpler, smaller, or more elegant alternative exists, including doing nothing or reusing existing code. Trace follows each claimed behavior through real call paths beyond the diff, noting surprises at seams with unchanged code. Verify walks whether traced paths actually produce claimed behavior, what inputs break it, silent semantic changes, and whether tests exercise the real path or only mocks. Report outputs severity-ordered findings with what to change, why, evidence, and a minimal suggested fix, closing with ship, fix-then-ship, rework, or reject. Operating rules forbid rubber-stamps, require file and line citations, mandate one simpler-alternative pass, and ban hedging flattery.3.8kinstalls2Debug Mantradebug-mantra is an agent skill that enforces a four-step debugging discipline triggered by /debug-mantra or whenever users report bugs, failures, stack traces, or investigation requests. Agents recite a verbatim mantra block first covering reproducibility, knowing the fail path, questioning hypotheses, and treating every run as a breadcrumb, then apply the steps in order before proposing fixes. Step one demands a fast deterministic repro artifact such as a failing test or curl script, escalating flaky cases through loops and stress until failure rates are debuggable. Step two prioritizes debugger attachment, then source tracing with knob enumeration across config flags and timing, then tagged in-code instrumentation with unique prefixes like DBG-a4f2. Step three requires disproof-first hypothesis testing with three to five ranked candidates. Step four maintains a ledger of experiments cross-referenced against all prior observations. Operating rules forbid fix proposals before reliable repro, hypothesis testing before fail path narrowing, or declaring success without ledger confirmation. Developers reach for it when they need structured debug sessions instead of immediate guess-and.3.5kinstalls3Post MortemPost-mortem is a 9arm agent skill for the canonical engineering record of a fixed bug written for other engineers after debugging lands a validated fix. It refuses to draft without four inputs: reliable repro, known root cause mechanism, identified fix with PR or commit pointer, and validated fix where the original repro now passes. The mandatory sections are Summary, Root cause, Fix, and Validation with optional Symptom, Why it produced the symptom, How it was found, Why it slipped through, and Action items. Tone is engineer-to-engineer with code identifiers, function names, file paths, and commit SHAs welcome, mechanism over narrative, active voice, no hedging, and blameless gap description rather than personal blame. It distinguishes bug-fix scope from customer outage incident reports, rejects trivial typo fixes, maps to debug-mantra breadcrumb ledgers, and offers management-talk handoff for leadership versions. Output flow confirms inputs, destination such as JIRA comment or docs postmortems path, produces a draft, requires sign-off before posting, and documents validation coverage limits explicitly when only one configuration was tested.3.5kinstalls4Management Talkmanagement-talk rewrites engineer-to-engineer content for engineering-org leadership such as VPs, directors, PMs, and release managers, shaping the same facts for JIRA comments, Slack posts, async standup lines, emails, or meeting talking points. The audience reads product and framework names plus JIRA keys and PR numbers but not code identifiers such as function names, file paths, struct fields, or commit SHAs. Tone rules keep concept-level vocabulary like race condition or synchronization while translating mechanisms into plain-English cause and effect without patronizing oversimplification. Channel shapes differ: JIRA uses bolded scan blocks for status, impact, what broke, owner, and next steps; Slack leads with one bold TL;DR and two to four bullets under eighty words; standup notes front-load state in one to three sentences; emails put the TL;DR in the subject; talking points stay clause-length bullets for spoken delivery. Source material may be a JIRA ticket fetched via REST, pasted technical text, or recent conversation context. The skill prints drafts by default and posts to JIRA only after explicit approval.3.4kinstalls5Qwenchanceqwenchance is an agent skill from thananon/9arm-skills for Claude Code sessions that risk looping, over-thinking, or exhausting context on multi-step work. It addresses three failure modes: repeating steps, re-reading the same files, and second-guessing in circles; unbounded internal reasoning; and running out of context mid-task without a handoff. Agents run its checklist when users report the model is stuck, spinning, or repeating itself, or when a long implementation needs a structured handoff before the context window fills. qwenchance complements execution skills by governing session pacing and continuity rather than writing feature code directly.1.4kinstalls6Qwen Agentqwen-agent is an agent-delegation skill that routes well-scoped, low-judgment work to a cheaper Qwen model via the claude-9arm CLI instead of spending primary Claude tokens. Typical jobs include bulk renames, formatting, boilerplate generation, find-replace passes, grep-style search, log summarization, test or docstring scaffolding, and running builds, linters, or tests with pass-fail reporting. Developers reach for qwen-agent when the task is mechanical and isolated—especially after hearing use qwen, delegate this, or do this cheaply. The skill explicitly excludes architecture decisions, security-sensitive edits, debugging judgment, and anything requiring the main conversation's full context.1.4kinstalls