makerskills: A Documentation-First Skill Pack for Solo Operators

The operator who has no process for decisions

The README frames the target audience narrowly: founders and indie operators. That is a different user from the one most agent tooling addresses. A solo operator does not need a code review bot or a deployment pipeline. They need a way to stop re-litigating the same decision every three weeks, a place to put research so it is not lost, and a repeatable shape for the content they publish. makerskills answers each of those with a separate skill, and the routing table in the README maps intent to skill in one line per row: think through a decision with a real fork goes to decide, break through a wall goes to unstuck, capture and query a personal knowledge base goes to second-brain. The problem it solves is not capability. Models can already reason about a decision. The problem is that the reasoning evaporates, and the README's answer is that each skill writes structured output to disk, with decide specifically archiving the result and a revisit date. That archival step is the actual product.

Each skill is a workflow document, not a program

The README is explicit that the plugin is documentation-first and that the automation is a side effect. The mechanism follows from that: a skill is a SKILL.md file under skills/<name>/, and you can either invoke it as a slash command such as /decide or open the file and run the steps by hand. The README gives the exact path for the first one: skills/decide/SKILL.md. This matters for how you evaluate the project. There is no daemon, no index, no server. The data flow in the documented first-touch experience is structured input, structured output, written to disk. The agent host supplies the model and the tool access; makerskills supplies the procedure and the stopping conditions. The README also notes that a skill's SKILL.md description carries trigger phrases and disambiguation from adjacent skills, which is how the host decides which skill to load when you are vague about intent. If you are choosing between decide and business-brainstorm, that description is the tiebreaker, not the routing table.

Install paths and the config directory

Two install routes are documented. Inside Claude Code, the marketplace commands are /plugin marketplace add coreyhaines31/makerskills followed by /plugin install makerskills@makerskills. For local development the README suggests cloning to ~/code/makerskills and symlinking that directory to ~/.claude/plugins/makerskills. After installing, the getting-started path sets one environment variable in ~/.zshenv: export MAKERSKILLS_CONFIG="$HOME/.config/makerskills". The README points to INSTALL.md for env vars, personal-config setup and runtime dependencies, which means the full dependency list is not in the README itself and you should read that file before assuming a clean install. Skills differ in what they need. /decide is described as requiring no config. /paste is described as a pure utility that needs no config file but may prompt you to install pbcopy-adjacent dependencies on first invocation. The domain skill is the heaviest: it lists Vercel CLI, whois per TLD, the Domainr API, the Namecheap API, rdap.org and agent-browser for USPTO trademark screening, plus aftermarket sites. That is an ensemble of external services, and the README does not claim they are bundled.

Where the pack is thin: credentials, platforms and blast radius

Three limitations are visible from the material alone. First, the skills that reach outside your machine depend on credentials and third-party APIs the repository does not provide. domain is the clearest case: an 11-step hunt across Domainr, Namecheap, whois and rdap.org will degrade to whatever subset you have authenticated, and the README does not describe a fallback path when one of those is unavailable. Second, the tooling is macOS-flavoured in places. The README's example for reading a skill is open ~/code/makerskills/skills/decide/SKILL.md, and /paste is described as reading the clipboard and copying back cleaned output, with pbcopy-adjacent dependencies named. On Linux or Windows those specific utilities are not the native ones, and the README does not document substitutes. Third, the blast radius is your own files. Because output is written to disk and decide archives with a revisit date, a misconfigured MAKERSKILLS_CONFIG or a wrong working directory writes into the wrong tree. There is no staging area described. If you need a skill to run inside an enforced pipeline with schema validation and rollback, this is the wrong shape of tool: it is a set of procedures an agent follows, and procedures can be skipped.

How makerskills differs from a general research or PKM tool

The obvious comparison is a note-taking system with an AI plugin, or a general-purpose research agent. The difference is in where the structure lives. A note-taking tool stores what you give it and leaves the method to you. A general research agent optimises for retrieving an answer. makerskills encodes the method as the artefact: decide carries the 37signals framework with 38 questions plus house additions such as Q39 opportunity cost, triaged down to six to eight questions per decision, and the triage is the point. A generic agent asked to help you decide will ask whatever it thinks of. A skill with a fixed question bank and a triage rule produces a comparable record each time, which is what makes the archive worth revisiting. The same pattern shows up in business-brainstorm, which pressure-tests an idea on nine dimensions and is documented as composing with deep-research and domain. That composition is a design choice worth noting: skills are meant to be chained, not used in isolation, and the README's own adoption advice is to pick the one skill that maps to a workflow you already do by hand.

The -ify trifecta and what it implies about maintenance

Three of the 21 skills exist to produce more skills. skillify has three modes: CREATE (from-chat, from-video, from-dump, from-scratch), ADAPT (port an external skill with a license check and attribution), and UPDATE (improve existing skills with cross-skill propagation and semver discipline). toolify is an interactive wizard for wiring an integration, API or MCP server into a Next.js or Rails project, covering auth, env vars, a client wrapper, webhook handling and a smoke test. loopify sits on top of ScheduleWakeup, CronCreate and /loop, and its documented job is judgment: pick the pattern (dynamic, cron or one-shot), tune the delay for cache windows, and enforce idempotency and bail-out conditions. That last set of concerns is the honest part of the design. Recurring agent tasks fail by running twice, running forever, or running at the wrong time relative to a cache window, and loopify names all three. The maintenance implication is that the pack is not a fixed product. The release history shows versioned additions rather than a stable surface: v1.3.0 added unstuck, v1.4.0 added Q39 opportunity cost to decide, and v1.5.0 added maker-council alongside a domain v0.2.0. Per-skill versioning means a skill can move independently of the plugin, so pinning the plugin version does not pin every skill's behaviour.

Licence, upkeep and who should walk away

The repository is MIT-licensed, which permits reuse and modification with the licence and copyright notice retained. That matters more here than for a typical library, because skillify's ADAPT mode is documented as porting external skills with a license check and attribution, so the pack is designed to absorb other people's work into your tree. MIT on this repo does not extend to whatever you port in, and the README does not claim it does. On upkeep: the cost is not compute, it is reading. Each skill you adopt is a document you should skim before trusting it, and the README's own five-minute path ends by telling you to pick one skill that maps to a workflow you already run manually, which is a reasonable filter. The people who should walk away are those without an Agent Skills host (the README names Claude Code, Codex and Cursor), those who want a single tool rather than 21 documents to triage, and those on non-macOS machines who are unwilling to translate the clipboard and file-open steps. The people who should stay are operators already keeping decisions and research in files, for whom the archive-with-revisit-date pattern in decide is the difference between a decision and a recurring argument.

Editorial conclusion

Adopt makerskills if you run a small operation and already keep decisions, research and content in files on disk, and if you are willing to read a SKILL.md before invoking it. Skip it if you need an enforced pipeline, a hosted service, or anything that works without an Agent Skills host. Before installing, open skills/domain/SKILL.md and check whether the Domainr, Namecheap and whois tooling it names is something you can actually authenticate, because that skill's usefulness depends on credentials the repo does not supply.

Official sources

1. coreyhaines31/makerskills on GitHub
2. License: MIT
3. Project website
4. README
5. Releases