本文へ移動
cccskills
無料GitHub で公開

get-it-done

Pursue an agreed goal to a working outcome without waiting: decompose it, re-prioritise as evidence arrives, branch, gate, get an independent review (a different model or a fresh reviewer subagent — the writer never approves), merge, prune, and bank the operator's decisions in STATUS.md instead of blocking on them. Trigger on "get it done", "don't wait for me", "proceed autonomously", or when an epic or roadmap objective is handed over. Not a close-out: this work set is expected to grow.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md6.8 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Get it done

The planning, the talking, the debating are over. The plan is agreed — execute it. Do not re-open scope, do not re-surface options already ruled on, do not ask for reassurance. The only reason to interrupt is a decision from the short "Theirs, always" list below; everything else is action.

The operator has left. Finish the phase. Do not stall on anything you can decide, evidence, or work around — and do not silently decide anything that is genuinely theirs.

The whole skill is one judgement: what is mine to call, and what is theirs? Get that wrong in the timid direction and the work sits idle for hours. Get it wrong in the reckless direction and you spend their money, their credibility, or their consent on a call they never made.

The boundary — this pursues a goal, it does not close a queue

The unit here is a goal: an epic, a roadmap objective, a phase, a thing that must end up working. The work set is derived from that goal and is expected to grow — you decompose, you discover, you re-prioritise as evidence arrives, and you update the delivery tracker as you go rather than at the end. Done means the goal is functional, not that a list is empty.

A close-out is the opposite shape: its work set is frozen at invocation and never grows, and it ends in a release advisory. If the operator asked you to close out what is in flight, do that job, not this one. Running this skill against a closing set is how a close-out quietly turns into a new sprint.

The contract

  1. Do not wait. No "shall I proceed?", no "let me know". If you can act, act.
  2. Bank, don't block. A decision that is genuinely theirs goes into STATUS.md with a recommendation — not into a message that stops the work.
  3. Land it. A phase is not done because the code is written. It is done when it is merged to main, the branch and worktree are pruned, and the tracked documents match reality.
  4. Never fabricate authority. Blanket "go" covers ordinary work. It does not convert into permission for the four things below.

Theirs, always — bank these, never assume them

Even under an explicit "don't wait for me":

  • Money. Raising a spending limit, a plan, a paid tier, a metered vendor.
  • Consent, credentials, retention, permissions. Anything a user agreed to, anything that reads or keeps their data, anything holding a secret.
  • One-way doors. Tags, releases, published artefacts, external submissions, force-pushes, deleting a ref whose commits are not reachable from the trunk.
  • Self-approval of a governance record. An ADR is never accepted by the agent that wrote it. Draft it, evidence it, and flag it.

Everything else — ordinary code, tests, docs, refactors, issue filing, branch hygiene — is yours. Do it.

When a dependency is missing, route around it and say so

A blocked gate is not a reason to stop; it is a reason to substitute and record the substitution.

UnavailableSubstituteRecord
Hosted CI (quota, outage, no runner)Run the local equivalent of each required checkName each check and its local equivalent, with exit codes
Hosted code review (quota)An independent reviewer: a different model or a fresh reviewer subagentWhich reviewer, what verdict, what it found
The first reviewer unavailableThe next independent reviewer availableThat the first choice failed, and why
A required check that cannot executeMerge on the local evidence if the repo permits itSay "past checks that could not run", never "past checks"

The distinction in that last row is the whole point. A check that failed and a check that could not start are different facts, and conflating them is how an unreviewed merge gets laundered into a reviewed one.

Never route around a guard that exists to stop you. A denied push, a protected branch, a consent gate — those are the operator's intent expressed in config. Routing around them is permission laundering, not resourcefulness.

The loop

For each unit of work:

  1. Branch. Never commit to the trunk.
  2. Build, with the project's own gates run as you go — not saved for the end.
  3. Prove it bites. Break the source deliberately and confirm the test fails. A green suite you have not seen go red is a hypothesis.
  4. Review independently. A different model (scripts/review-gate.sh --pr N, the Gemini lane) or a fresh reviewer subagent running pr-reviewer; adversarial-gate for design. The writer never approves the work. If the candidate changes after review, the review is void — re-run it.
  5. Merge with a merge commit. Never squash, never rebase onto the trunk.
  6. Prune the branch and worktree — but only once git cherry confirms the commits are reachable from the trunk. Where integration reworked a patch into a new SHA, deleting the ref loses history. Keep it and say why.
  7. Reconcile. STATUS, CHANGELOG, ROADMAP, ADRs, issues. Findings that live only in a chat message evaporate; file them.

Parallelism

Use worktrees and subagents where the work is genuinely disjoint. Two agents editing the same file is not parallelism, it is a merge conflict with extra steps — and most "parallel" refactors of one subsystem collide. Say so rather than parallelising for appearances.

The handover

STATUS.md is what they read when they return, so write it for someone who was not in the room. It carries:

  • What landed, with evidence — a SHA, an exit code, an issue number.
  • The decisions waiting on them, each with: a recommendation (a verb and an object, never "consider the options"), what it unblocks, what it costs if they call it wrong, and whether it is reversible — cheap, costly, or one-way. Sort so the compounding ones are first.
  • What you did not do, and why. Especially anything you declined as theirs.
  • What is still open, including defects you found and chose not to fix.

Reporting back

Lead with the outcome, not the narrative. State what merged, what the evidence was, what you routed around, and what is waiting on them. Report failures plainly — a phase that half-landed is reported as half-landed.

Never claim a gate passed that you did not run, and never inherit a figure from an earlier session: re-derive it, or label it unverified.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Google ADK (Agent Development Kit) orchestration patterns — boundaries, agent composition, and tool seams. Trigger when designing or reviewing multi-agent systems built on ADK. Authoritative source: adk.dev.

日本語の概要は準備中です。原文の説明を表示しています。

jpantsjoha/ai-native-developer-experience122026年10月6日 更新

JP's signature red-team pass — "how would I break this?" Argue against your own approach before proceeding. Trigger on any high-stakes decision, architecture choice, or before marking work complete.

日本語の概要は準備中です。原文の説明を表示しています。

jpantsjoha/ai-native-developer-experience122026年10月6日 更新

Read-only SRE checkup of any GCP project: deterministic probes of the edge, Cloud Run services, 7-day error logs, Cloud Scheduler, alert policies and uptime checks, Secret Manager and IAM, the data stores and the machine's own scheduled jobs, audited into one fixed status table (LIVE / WARNING / RED / INCONCLUSIVE) with evidence, findings by severity, what could not be checked, and a single OVERALL line delivered as one notification. Parametrised by a per-project manifest, so the same routine runs on every project. Use when the operator says "cloud checkup", "SRE check", "is everything live", "what's healthy / warning / red", "any errors this week", "audit the infra", "weekly checkup", "set up the weekly checkup", before a deploy or demo, or after an incident. Cloud Run first; App Engine and GKE differ only in the serving probes.

日本語の概要は準備中です。原文の説明を表示しています。

jpantsjoha/ai-native-developer-experience122026年10月6日 更新

Cloud guardrails for any vendor workload — Google Cloud (GCP, Vertex AI, GKE), AWS (IAM, EKS, Bedrock), Azure (Entra ID, Policy, AKS), Alibaba Cloud (RAM, mainland/international residency). Enforces identity least-privilege, mechanical policy, data boundaries, residency, cost caps, network egress and observability, with official-source validation before any claim. Trigger on any cloud infrastructure design, review, Terraform plan, or LLM/agent deployment; the-architect routes here.

日本語の概要は準備中です。原文の説明を表示しています。

jpantsjoha/ai-native-developer-experience122026年10月6日 更新

LLM and cloud cost awareness — model tiering, token budgets, right-sizing, and when a cheaper model suffices. Trigger before finalising any architecture that calls LLMs, before scaling a workload, or when a cost estimate is needed.

日本語の概要は準備中です。原文の説明を表示しています。

jpantsjoha/ai-native-developer-experience122026年10月6日 更新

Decompose an epic into atomic parallelizable tasks, route each to the right skill, and keep the four delivery records straight — issues, STATUS, ROADMAP, CHANGELOG. Use as a meta-router when several skills could apply, and as the baseline for how delivery state is recorded. Trigger at the start of any multi-track epic, when the skill count exceeds ~12, or when the records have drifted from reality.

日本語の概要は準備中です。原文の説明を表示しています。

jpantsjoha/ai-native-developer-experience122026年10月6日 更新

jpantsjoha のスキルをすべて見る

このスキルの問題を報告する