justnnc skills
Agent skills for software work that has to be right, not just plausible. Each one makes the agent show its evidence: commands it ran, lines it read, decisions you made.
| skill | what it does |
|---|---|
| prove-by-running | Proves a script or tool works by running it on a disposable, contained copy, and ends in one verdict: runs clean, breaks, or unproven. |
| parallel-refactor | Fans one large mechanical change out across parallel agents: measure the real call sites, write a contract, keep every intermediate state valid, verify with a canary. |
| spec-audit | Audits specs against the work that just landed: built but undescribed, described but unbuilt, stale text, planned text inside a spec. |
| feature-spec | Takes a feature idea to a decided design document: what already exists, a security review of candidates, quick spikes of the risky parts, each decision recorded. |
| how-it-works | Answers how a codebase works in two checked stages: one maps where to look, the next reads those spots closely. |
| adversarial-review | Briefs a reviewer to attack a proposal before it is built: failure modes, weak evidence, simpler alternatives. |
| skill-craft | A craft checklist for writing, reviewing and testing agent skills. |
Install
In Claude Code: /plugin marketplace add justnnc/skills, then /plugin install justnnc-skills@justnnc.
Other agents: npx skills add justnnc/skills -a <agent> --skill prove-by-running parallel-refactor spec-audit feature-spec how-it-works adversarial-review skill-craft.
Examples
prove-by-running. "I wrote a script that rewrites commit dates before push. Is it safe?" The agent copies the repo into a contained scratch area with a local bare remote, runs the script, runs it again, tries the failure paths, and reports RUNS CLEAN, BREAKS or UNPROVEN with the commands and output behind it. parallel-refactor. "Move every caller of the old config API to the new one." The agent counts the real call sites first, records a canary before touching anything, writes a contract table for every case, adds a temporary shim, runs the change in waves of parallel agents, removes the shim and checks the canary again. spec-audit. "We just merged the import feature. Are the specs still right?" A read-only auditor compares each spec section with the merged diff and returns findings in four kinds, each with file and line on both sides, plus who fixes what. feature-spec. "Should we build a plugin for X, or does one exist?" The agent checks the need, searches for prior art in widening rounds, has any candidate security-reviewed before it runs, proves the risky assumptions with throwaway spikes, asks you each decision, and writes the design document. how-it-works. "How does this engine resolve collisions?" One stage maps and ranks the relevant functions and line ranges; after the top entries are checked, a second stage reads only those and returns the mechanism with formulas and citations. adversarial-review. "We're about to switch the cache to write-through." A reviewer attacks the proposal on assumptions, side effects and cheaper alternatives, and returns a verdict with ranked objections. skill-craft. "Review this SKILL.md." The agent checks it against the craft list and returns a verdict with numbered gaps, each with the line and replacement text.
Where they work best
The skills load in Claude chat, Cowork and Claude Code. Most of them direct an agent that can run commands, use git and start subagents, so they do the most in Claude Code and in Cowork sessions on your own computer. In chat they still guide the work, but steps that run commands become instructions for you.
What the plugin runs
Nothing on its own. It contains only markdown skill files: no scripts, hooks, MCP servers or network calls. Any command is run by the agent when a skill tells it to, under your usual permission settings.
The skills are new and have not yet been evaluated across models; treat them as considered practice rather than measured results. Source, examples and issues: https://github.com/justnnc/skills
MIT licensed.