Agent Skills

npx skills add Mark-Life/agent-skills -g

The skills I run every day: the parts of my judgment I got tired of retyping into every session. Each is plain markdown, loaded only when the work matches.

The command above uses the skills CLI and installs globally, prompting for which skills and which agents. Add --skill <name> for just one, or --all to take everything without prompts.

Coding Workflows

Taste and process for the build loop: what to make, how to start it, how to make it answerable once it is running, how to write it up.

• product — scope the ambition to what agents made possible, and name the primitive others compose on.
• new-project — scaffolds a repo from my Next.js monorepo template and runs the bootstrap.
• observability — one wide event per request or job, high-cardinality fields, span and metric design, OTel export, tail sampling.
• pr-issue — titles in Vim's area: summary style, then labelled sections in template order.
• media-cdn — a private object store where a screenshot, GIF or video flips to a public CDN URL, and back.

product is the one that runs before any code exists: it decides what is worth building. It makes the agent name the primitive others will compose on and leave the rest to the caller, so an API, an SDK or a spec comes out with a surface someone can build on instead of a pile of features.

observability owns the instrumentation decisions for TypeScript services. A request, job or run emits one canonical record carrying every dimension you would want to group by later, instead of log lines scattered through the code, so a production question gets answered by a query rather than by a guess.

A PR body carries Problem:, Solution:, Security Impact:, Testing:; an issue carries Problem:, Reproduction:, Expected vs actual:, Proposed solution:. Nothing else: no changelog, no checklist, no command output that was not really produced. media-cdn is what makes that write-up visual when words are not enough: upload private by default, publish when the image has to be embedded, unpublish when it should not be on the open internet.

Communication

One skill per direction of writing. agent-to-human governs what the agent says to you: it reasons at any length in private, then answers first, cuts to the decision, and shows only receipts. It is the highest-leverage skill here, because it stops the model narrating what it did and makes it say the thing you needed. human-to-agent governs what you say to the agent: prompt the positive, cut no-ops, give checkable completion criteria. Load it whenever you are writing something an agent will later execute as instruction — a skill, a CLAUDE.md or AGENTS.md, a set of agent rules, a prompt. writing-for-readers governs what either of you publishes: an article, a blog post, a course lesson, a landing or docs page. It names nine tells that make prose read as machine-written — the aphorism pair, the thesis opener, negate-then-correct — and carries the rejected and approved version of each from one real rewrite.

The line to draw is between agent-to-human and writing-for-readers, and it is who came to whom. The first is output handed back to a person who asked for it, so it can assume the question and answer it. The second is prose a reader found on their own, half-skimming on a phone, still deciding whether to keep going.

Inspect & Debug Agents

These answer one question at two scales: what is actually in the agent's context, and what is it costing you.

• session-report — HTML report of a session's context window: token budget, thinking cost, dumb-zone cutoff, full history.
• memory-view — secret-redacted HTML explorer for Claude Code's per-project auto-memory. Read-only, no curation.
• context-doctor — audits the context auto-loaded into every session and ranks the cuts: prune, gate behind a trigger, or route elsewhere.
• harness-doctor — mines the last N days of transcripts across every project into five fact tables and returns a ranked fix list, each fix carrying the number behind it.

That auto-loaded context is tool and MCP definitions, connectors, plugins, skills, subagents, memory and rules files. You pay for it on every turn, before you have typed anything, so context-doctor cross-checks it against what you actually invoke.

context-doctor is the single-session view; harness-doctor is the fleet view. It reads the last 30 days of bash calls, errors, tokens, human corrections and subagent fan-out across every project on the machine, because the waste is invisible from inside any one session — a command that costs 9 seconds is nothing until you see it ran 1,400 times.

session-report and memory-view are the one-shot path: nothing to install, one command, one report answering one question. Peektrace is the same data as a live pane — browse every session, and edit the memory instead of only reading it.

MIT licensed.