本文へ移動
cccskills
無料GitHub で公開

deliver

Carry out a released charter to a proven result: plan only as much as the work needs, do the work through the project's own workers, prove every success check with a recorded real run, get one independent review, and close honestly. Invoke explicitly as `$deliver <charter>` (Codex) or `/deliver <charter>` (Claude, Cursor), or the way your agent invokes a skill, or when a dispatch names you the executor of a charter.

インストール方法を見る

含まれるファイル(5)

  • SKILL.md5.7 KB
  • references/proof.md1.8 KB
  • scripts/deliver-proof6.1 KB
  • scripts/deliver-proof.t12.9 KB
  • scripts/selftest/run-selftest.sh1.7 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。

deliver

Carry out the named charter. Its ## Success checks are the definition of done; nothing else is. Read the whole charter and the existing work first. The owner (whoever owns the work, often the user) can override this skill and any plan an agent wrote. Any kind of project uses it: for code and tests also read references/proof.md.

Prerequisites: sno deliver-proof needs bash 4+ and GNU coreutils (Linux) and prints one message and stops when they are missing; its selftest also needs python3. On macOS, install bash 4+ and GNU coreutils (for example with Homebrew) and put them first on PATH. If sno deliver-proof is not available, run scripts/deliver-proof from this skill's folder instead.

Keep the work useful

  • Before adding a prerequisite, test, record or approval step, name the requested result it proves and what changes if it fails. If nothing changes, leave it out.
  • Check an uncertain fact by reading or calling what exists. If checking needs something built first, building it is the first task, not a permission gate.
  • Records, telemetry, formatting, identifiers, checksums and review sign-offs never block the work. Log the failed action, its place and the actual error, and continue. Do not turn such a failure into an assertion, a silent return or a swallowed error. If the main operation fails, say so with its cause; never report success. Stop only what cannot proceed and keep independent work moving.
  • Time is finite and proof shares it: estimate a run from existing timings before starting, test only what changed and what directly depends on it, and by default run a full suite only when the owner explicitly asks. Do not stop a useful run just because an estimate has passed.
  • By default add no security check or new gate the owner did not ask for; propose it as a question that says what it would check and what it would block.

Steps

  1. Start. If the charter is missing, is not released, or has a success check nobody could observe, say so once and have charter revise it with the owner. Never invent a decision. If <charter-name>.state.md sits beside the charter, an earlier executor stopped partway: run sno handoff-checkpoint --verify on it (the handoff skill ships the command), read its Done and Next lists, treat Done as finished, and continue from Next instead of starting over.

  2. Plan, only as much as needed. Work with dependent steps gets a ## Plan in the charter: the steps in order and who does each. One or two steps need none. A plan of more than three steps gets one independent review, then work starts; there is no second plan document. A product decision the charter does not settle is recorded in the plan and asked once, while independent work continues.

  3. Do. Use the project's own workers and tools: for code, its coder, test and cleanup skills when installed; for documents, its writing and review tools. Build the smallest real integration early when later steps depend on it. Keep existing behavior unless the charter says otherwise. Reproduce a defect before fixing it. After every finished step run sno handoff-checkpoint <charter-name>.state.md (beside the charter) and update its Done and Next lists, so another agent can take over at any moment. If the command is not installed, skip the record and keep working.

  4. Prove. For each success check, record one of:

    • sno deliver-proof run <charter> <n> -- <command>: the real command, run in the current directory; its output and exit status are kept as the log and the table row.
    • sno deliver-proof see <charter> <n> <evidence-file> "<what was observed>": when inspection is the proof (a produced document, a message at its receiver, a screen). The evidence file is a nonempty file, named relative to the current directory.

    Prove the result where the requester would see it: read back the saved value, open the produced file, read the message at its receiver. Never substitute a mock's answer, a successful exit, a call count or the program's own narration. The expected value comes from the charter or independent data. Never edit the Proof table by hand, never mark a failed or unrun check passed, and after a repair rerun only the affected checks.

  5. Review once. One independent review of what changed (peer-review when installed, otherwise a second agent given the charter and the changed work). Fix ordinary-path defects in scope; report the rest without growing the job.

  6. Close. Run sno deliver-proof check <charter> and read its output. Only when it exits 0, write ## Report (what worked, which checks ran and how, anything unfinished) and set status: delivered. A failing record step is logged and repaired on its own; it does not undo working results. Tell whoever dispatched you the result and what remains.

Publish, send to other people, spend, delete, deploy and commit only within the owner's authorization. If ~/.config/sno/decision-rights.md exists (the pl skill creates it), it lists what needs the owner. Do not absorb the next charter's work. When you are blocked on a missing input, state the assumption you proceed under in the plan and keep working.

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

Answer how long agent work will take, how much longer active work needs, or what clock time applies. Estimate full AI-agent wall time, measure active progress first, and get owner-facing times from `sno report-time`. Never estimate human-developer time.

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。

sno-ai/sno-station-skills2882026年10月13日 更新

Prove that a reported run caused its intended real effect. Walk the executor path from entry to effect, capture one observable per hop, and report a result a no-op could not produce.

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。

sno-ai/sno-station-skills2882026年10月13日 更新

Give the owner one page when they come back from sleep, a meeting or a long run: what got done, what is stuck, what waits for their answer, and what it cost in quota. Use when the owner asks what happened while they were away, wants a status page, or when you finish a long stretch of work and must report it.

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。

sno-ai/sno-station-skills2882026年10月13日 更新

Report how many problems the two brains found: mutual review (the other model checks the work, from the peer-review records) and self-check (an agent finds its own error in its own conversation, judged by that agent's own model), side by side. Use when the owner asks whether the second agent or the reviews are worth it, or wants the week's review results.

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。

sno-ai/sno-station-skills2882026年10月13日 更新

charter

無料

Write, revise or review a charter: the one document that tells an executor what to deliver, what is out of bounds, what the owner already decided, and how each success check will be proven. Use when work needs a written brief before someone else does it, or when asked to write, change or check a charter.

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。

sno-ai/sno-station-skills2882026年10月13日 更新

cos

無料

MANUAL-ONLY: load solely when the human owner types `/cos`, uses `$cos`, or asks you to act as the Chief of Staff (COS). The COS is the one agent the owner talks to; it coordinates the owner and the per-project Project Leads (PLs). A task card, charter, or dispatch makes you an executor, not COS.

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。

sno-ai/sno-station-skills2882026年10月13日 更新

sno-ai のスキルをすべて見る

このスキルの問題を報告する