Claude Skills by junielton

Use when a change touches Harnu's agent-facing surface and the self-awareness doc (docs/harnu-features.md) must be updated to match — i.e. you added or changed an MCP verb in src/main/mcp/tool-catalog.ts, changed the shape of an ACK a verb returns, changed grant/confirm semantics, or added a UI affordance the agent should proactively offer the user. Also use when the CI "Self-awareness doc updated" gate fails, or when someone asks how to keep harnu-features.md current. This skill codifies the...

Drive the Harnu fleet as a Conductor — see every running session/worktree and create new sessions, split terminals, and git worktrees via the `harnu` MCP tools. Use when the operator asks you to orchestrate, manage, or set up work across multiple Harnu sessions/folders ("spin up a session for X", "make a worktree and start a session there", "what's my fleet doing", "fan this out across worktrees", "act as the conductor"). Only works inside Harnu with the Control server (MCP) enabled.

Use before pushing a branch or opening/updating a PR in this repo, when you need to know whether GitHub Actions will go green — it runs the same gates as the CI `verify` job locally (format:check, i18n parity, the CHANGELOG/self-awareness/user-docs contract gates, typecheck, lint, test:coverage). Also use when CI is unavailable (billing block, offline, rate limits) and a merge decision still has to be made, or when you want to fix every gate failure in one pass instead of one push per red che...

Drive the Harnu fleet through the `harnu` MCP verbs — see every session/worktree and create sessions, terminals and worktrees. Use when the operator asks you to orchestrate, manage or set up work across several Harnu sessions or folders — "spin up a session for X", "make a worktree and start a session there", "what is my fleet doing", "fan this out across worktrees", "act as the conductor". Only works inside Harnu with the Control server (MCP) enabled. Do NOT use for planning a delivery end t...

Verify that a finished unit satisfies its acceptance criteria one at a time, from evidence rather than the executor's own report, with a rubric per deliverable kind (code, UI, research, decision), record the verdict on the Mission step with mission_verify_step, then produce the Delivery Report. Use at the end of a dispatched unit or a whole fan-out wave — "check whether it really delivered", "go through the acceptance criteria", "which ACs do I still have to test by hand", "write the delivery...

Read-only stall detector for deliveries that are already dispatched — reads every active Mission through mission_list/mission_get and raises ONE notification naming the exact next action when one is stale, blocked on the operator, or holding a unit that is finished but parked (pushed with no PR, a card the board is lying about, an executor that vanished). Built to run unattended as a Harnu Scheduler worker in `observe` mode, on the repo an orchestration is running in, so a stalled unit is fou...

Transform a rough idea, voice-note-style draft, or vague request into a production-quality prompt for Claude models (Haiku, Sonnet, Opus, Fable 5). Use this skill whenever the user shares a messy prompt draft and asks to improve/refine it, asks "help me write a prompt for X", asks which model or technique fits a task, or mentions prompt engineering, prompt refinement, or converting an idea into a prompt — even if they don't say the word "prompt" but are clearly describing a task they want to ...

Run one coordination tick over dispatched work: read the Mission with mission_get, act on what is unblocked, record it with the mission_* verbs, report one line. Use "mission set <objective>" to declare or re-scope your mission (a structured Harnu Mission), a bare invocation to run a single tick, and "mission done" to ask the operator to close it. Trigger when an orchestration session needs a recurring heartbeat over workers, worktrees and approvals — "run a tick", "how is the mission going",...

Decompose an objective into board cards, dispatch one Harnu session per unit, monitor until each opens a PR, verify every acceptance criterion independently, and hand back a Delivery Report. Use when the operator hands over an objective too large for one session and wants it delivered rather than discussed — "orchestrate this", "break this down and dispatch it", "fan this out across worktrees", "deliver this epic" — or when a Harnu board already holds several ready cards that need to be plann...

Speak a short, ear-friendly summary of a result out loud — through Harnu's `speak` verb when the session runs inside Harnu, or a text-to-speech command already on the machine when it does not. Use when the operator asks to hear something instead of reading it — "read that out loud", "read it to me", "say that out loud", "tell me what happened", "I'm away from the screen, speak it" — with no argument (the last result produced in this conversation), a file path (that file), or quoted text (that...

The shared final-report format for any multi-step run: a flat rail of executed steps (✓/-/!/✗ with concrete outcomes), a gutter ledger for what is owed (`you` — always printed, `— nothing, you're clear` when nothing is) and any open question, a links list of everything the reader might open, and a summary of at most 3 lines. Use it whenever a task involved a pipeline or several sequential steps and you're about to write the wrap-up message — shipping a PR, reconciling branches, a deploy, a ba...

Render the state of every feature, task or mission in flight as a fixed six-line glance card — progress bar per front, what is running now, what is done, what is next, what needs the operator, what is open — followed by the links worth opening, built from mission_list/mission_get, the board and the PRs, never from memory. Use whenever the operator asks "status?", "where are we", "is it done", "what's left", "what do you need from me", "how's the mission", or opens a session after being away. ...