skills

Public agent skills for Grok, Claude, agy, kimi, opencode, etc.

All work skills ship as one plugin, prompt-manager, so teammates install a single package.

Personal tooling ships as separate plugins: scout (quota-routed codebase exploration across kilo/kimi/claude CLIs), work-scout (the same idea for Slack and ClickUp questions) and task-writing (turn a rough description into a task a stranger can pick up).

Install (Grok)

grok plugin marketplace add maxh213/skills
grok plugin install prompt-manager --trust
grok plugin install scout --trust        # optional, personal tooling
grok plugin install work-scout --trust   # optional, needs scout for its router
grok plugin install task-writing --trust # optional, personal tooling

Then /setup-prompt-manager once, then /prompt-manager-full-run.

Install (Claude Code)

claude plugin marketplace add maxh213/skills
claude plugin install prompt-manager@maxh213-skills
claude plugin install scout@maxh213-skills        # optional, personal tooling
claude plugin install work-scout@maxh213-skills   # optional, needs scout for its router
claude plugin install task-writing@maxh213-skills # optional, personal tooling

Restart Claude Code afterwards — plugin skills are enumerated at startup.

Alternatively, copy each folder under any plugin's skills/ directory into ~/.claude/skills/.

Layout

plugins/prompt-manager/
  plugin.json
  references/            # credential catalog + worktree isolation
  scripts/               # credentials.py, worktree.py
  skills/
    setup-prompt-manager/
    sprint-tasks/
    task-context/
    idea-to-prompt/
    prompt-manager/
    prompt-manager-full-run/
    task-to-clickup/     # TASK-<slug>.md -> current sprint (assigned) or PM TRIAGE (unassigned)
    guide-me/            # task-context dive, one plain-English paragraph out

plugins/scout/
  plugin.json
  skills/
    scout/
      SKILL.md           # routing rationale, brief discipline, backend recipes
      route.py           # live quota probes + DECISION=<backend> picker

plugins/work-scout/
  plugin.json
  skills/
    work-scout/
      SKILL.md           # when to scout vs look, the two-pass loop, reading the report
      scripts/
        wscout.py        # read-only Slack + ClickUp helper the scout calls
        run.py           # route -> pass 1 -> critique pass 2 -> report.md
        brief-template.md

plugins/task-writing/
  plugin.json
  skills/
    write-task/
      SKILL.md           # description in, TASK-<slug>.md out
      CHEAT-SHEET.md     # section anatomy, per-type shapes, 60-second check
    grill-task/
      SKILL.md           # grilling -> write-task

Secrets are read from ~/.config/workstation/config.toml at runtime (mode 0600). Nothing is embedded.

prompt-manager-full-run accepts --skip-thoughts to skip the extra feedback step between task-context and idea-to-prompt.

scout needs at least one of the kilo, kimi, or claude CLIs installed and logged in; see plugins/scout/README.md for how routing works.

work-scout answers "what did X show me / what did I miss / what's the history of GLOBAL-123" from Slack and ClickUp without the raw messages ever entering the main session: a cheap model, routed by scout's route.py, searches through a read-only helper, is asked twice, and hands back a cited report file. Needs [slack] and [clickup] tokens in ~/.config/workstation/config.toml; see plugins/work-scout/README.md.

task-writing writes nothing the description didn't say: a gap that changes the work is asked about first, and the rest is left out. Tickets never carry open questions. /grill-task uses the grilling skill from mattpocock-skills when it's installed, and runs the interview itself when it isn't.

/task-to-clickup TASK-<slug>.md files a task file into ClickUp: the current sprint (assigned to you, status backlog) or the PM TRIAGE queue (unassigned). It asks for sprint points and MoSCoW, proposes Field of work, Project name, Project category and a one-line Definition of Done from the file, and creates the task only after you confirm. The triage list id is clickup.triage_list in the credential catalog.

/guide-me GLOBAL-XXXXX runs the whole task-context dive and answers in one plain-English paragraph: what is wrong, where (one file link), the fix with its one caution, and whether anyone has started it. Shaped by the i-have-adhd rules, then cut further: no steps, no estimates, nothing the reader will not read.