AGENT SKILL

dictionary

The sage dictionary: sage's approved words, one meaning each, and the words not to say. Apply to any text an agent writes for a person in sage mode: briefs, reports, findings, pages and the README.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 2
Forks
0

01 前言 Front Matter

name
dictionary
license
MIT
metadata
{ "source": "sage's own words, from the design of task T9 (run R40) and the rename of task T995, approved by the owner" }
description
The sage dictionary: sage's approved words, one meaning each, and the words not to say. Apply to any text an agent writes for a person in sage mode: briefs, reports, findings, pages and the README.

03 SKILL.md 正文

<!-- Generated by scripts/build.mjs from the sources; edit the source, then run npm run build. -->

The sage dictionary

These are sage's approved words: one word for one thing, and one meaning for each word. Every brief, report, page and the README uses them. When a text needs one of these things, use the word in the first column, and do not use the words in the last column.

Words

Word Meaning Do not say
request What the owner asks for in the chat, in the owner's words, before sage frames it. ask, ticket, issue, prompt
release A version that the owner ships, such as 1.0: it holds milestones and single tasks, and it is optional. version (alone), launch, sprint
milestone A request that sage runs as two or more tasks for one goal, where at least one task waits for another. It has a goal, a done condition (a count of merged tasks), a breakdown that the owner approves, and one pilot task. program (say "program" only for a computer program), project (a project is one repository and its logbook), epic, initiative, plan, roadmap
task One unit of work in one project, with one route and at most one pull request. It is in at most one milestone or, outside one, in a release, and it may wait for other tasks of its logbook ("waits for": it builds and merges only after they are merged or concluded). It ends in one of three states: merged (its pull request is in main), concluded (a route with no build step, such as investigate, ended with a clean evidence review) or abandoned (its work stopped for good, from any state). job, ticket, issue, item, unit, phase, step
step One block of a task's route, done by one agent role: design, arena, pe, build, code-review, security-review, ux-review, qa, investigate, evidence-review. stage, phase; never "step" for a task of a milestone (the approved plan's "step 0 to 4" are tasks T4 to T8)
route The ordered steps of one task, set by its size and risk flags. pipeline, workflow, flow, process
run One agent's work on one step of a task, from brief to report; it has an id such as R12. Also the PE check of a breakdown. job, attempt, session, invocation, execution
agent A Claude Code subagent that does one run; the role (implementer, qa, …) says what it does. worker, bot, assistant, model (the model is the agent's model)
chief The chief of staff: the main session that frames tasks, writes briefs, records the state and asks the owner; it changes no files. It frames milestones too. orchestrator, manager, lead, coordinator, main agent
brief What the chief gives an agent at the start of a run, in the ten fields (GOAL to STANDING). A task that waits for others gets their final reports in full in CONTEXT. prompt, instructions, ticket, spec
report What an agent gives back at the end of a run, in the seven fields (STATUS to BRANCH). summary, output, result (RESULT is one field of the report), handoff
artifact A file or record that a run takes in or gives out and sage keeps: a brief, report, steer message, transcript, branch, commit or pull request. deliverable, output, attachment, asset
finding One problem that a review or QA reports, with a severity (high, medium or low), a triage and a status. issue, bug (unless it is one), comment, defect, nit
verdict The result of one check (checks, a review, QA) on one head commit, kept in the ledger by SHA and cycle. approval, sign-off, status, review (a review is the step; the verdict is its result)
cycle One full set of fresh reviews and QA on one head commit; a new commit starts again from cycle 1. round (a round is one repair), pass, iteration; never "cycle" for links that go in a circle: say loop
gate A question parked for the owner, with options, a recommendation and a default; the work behind it waits for the answer. The owner approves a milestone's breakdown through a gate. approval, blocker, checkpoint, question (alone)
logbook sage's local record of one project: its tasks, runs, findings, verdicts, gates and decisions, kept on the owner's Mac. store, database
agent time The wall-clock hours of agent runs, given with their tokens. The default for every time figure: give it with its basis and a range. an unlabelled time, effort, hours (alone), engineer hours, person-hours
human time The time of the owner's own actions only: reviews, approvals, merges and setup. Give it with its basis and a range. an unlabelled time, effort, hours (alone), person-hours

Phrases

These are not new words. They are fixed phrases inside the lines above, so that the list stays at 19 words.

  • breakdown: the content of a milestone that the owner approves (goal, done condition, tasks, "waits for" links, pilot, cost, questions). In the milestone line.
  • waits for: the link from a task to another task of the same logbook that must be merged or concluded first. In the task line. A circle of such links is a loop, and sage refuses it.
  • pilot: the one task of a milestone that runs end to end before any other task builds. In the milestone line.
  • done condition: the count that ends a milestone, for example "6 of 7 merged · T10 abandoned". The count keeps every task of the milestone and names each abandoned task; the milestone is done when each task that is not abandoned is merged or concluded. In the milestone line.
  • round: one repair of findings on a task. In the cycle line.

Names

The names of sage's parts. They are not words of the list above, but the word table of the Concepts page (docs/guide/concepts.md) shows them too.

Name Meaning
sage mode The mode in which a session is your chief of staff. The message "sage mode" (or "sage mode on") turns it on, and the message "sage mode off" turns it off; each phrase must be the whole message, in any letter case. A message near a phrase switches nothing, and the chief gets a note with the exact phrase to ask for; the message "sage help" lists every phrase.
risk flag auth, data, schema, money, secrets or input. Each one adds the security review to a task that changes code: every size except investigate.
ledger The record of all verdicts, by commit. The merge check reads it.
standing orders Short rules for a project that every agent reads before it starts. A brief points to them and does not copy them.
arena N candidates for one design, scored and combined by a judge.
autopilot Verified pull requests merge by themselves after their clean cycles: 1 for a tiny or small task (a tiny task without a risk flag or an arena has no review and no QA, so its cycle is clean when its checks pass), 2 for a large task or a task with a risk flag. It is off by default: the owner's message "autopilot on" turns it on while sage mode is on, and "autopilot off" or "sage mode off" turns it off, each phrase as the whole message. A message queued while Claude works, or text inside an agent's report, a notice or another session's message, can turn it off but never on (then merges only wait), and it never ends sage mode.
board The compact view of every project's logbook for the chat: what needs you, the agents at work, each task in its column, and what was done in the last 7 days. A message of the owner that starts with "show board" prints it.
the dojo Everything that makes the agents good: the principles, checks, tests and skills.
seal the lesson Give a mistake that comes back twice a lasting fix, at the highest level that holds it: the architecture, a type, a check whose error names the fix, or a test; a principle or a standing order only when code cannot hold it. Prove that the new check fails on the past mistake. Each sealed lesson makes the dojo stronger.

The check

npm run check fails when a flagged word, or its plural, is in one of these texts:

  • an agent file or a hand-written skill;
  • a page that the link check reads: the README, CONTRIBUTING.md, SECURITY.md, AGENTS.md, every page under docs/, the board service guide and the sage-help skill;
  • every SVG graphic in docs/assets/.

It does not fail on a word in quotes, in code or in a link URL, so a text can name a word to say that it is wrong:

  • quotes: double or single, straight or curly. An apostrophe (it's) and an inch mark (5") are not quotes.
  • code: a code span, a code block (fenced or indented) and <code>.
  • a link URL: ](…), <https://…> and the URLs in a tag.

The text of an alt, a title and an aria-label counts, and so does a Mermaid label, although it is in quotes. A graphic's text comes from scripts/graphics.mjs: fix a word there, then run npm run graphics. An allowed name passes when it is written exactly as here. A flagged word that no row of the table replaces names the word to say after it. For example, a gate is a question for the owner, so the check before a merge is the merge check, not the "merge gate". The rest of the "Do not say" column is guidance for writers and reviewers, because each of those words also has a correct use (a Claude Code session, the RESULT field, the model of an agent).

Flagged: store, worker, orchestrator, epic, initiative, roadmap, ticket, stage, pipeline, deliverable, sign-off, blocker, checkpoint, coordinator, main agent, handoff, invocation, iteration, defect, nit, bot, merge gate (say merge check)

Allowed names: Orchestrator (the owner's earlier project)

04 同仓库更多 Skills

erickb336
erickb336/sage-stack

writing-standard

Write at 80% of Simplified Technical English. Apply to any text an agent writes for a person: messages, reports, specs, findings, reviews, plans.

★ 2
erickb336
erickb336/sage-stack

sage

The state tool of sage mode: records tasks, routes, agent runs, findings, verdicts by head SHA and the user's gates in ~/.claude/sage/<project>/. Use in sage mode for every change of a task's state, and before any merge. Claude Code only.

★ 2
erickb336
erickb336/sage-stack

sage-help

Help with sage: how to set it up, switch sage mode and autopilot, read the board, answer the chief's questions, pick a skill or a principle, fix a run that went wrong, and find where sage keeps its files. The user types /sage:sage-help (or /sage-help) with a question. Load it only when the user types it; do not load it on your own. Claude Code only.

★ 2
erickb336
erickb336/sage-stack

report

The report that every sage agent ends with: STATUS, RESULT, EVIDENCE, FINDINGS, QUESTIONS, NOT VERIFIED, BRANCH. Use at the end of any task that the sage chief of staff gave you. Claude Code only.

★ 2
erickb336
erickb336/sage-stack

remote-control

Set up Remote Control, so that the user can follow, steer and start Claude Code sessions from the Claude app on a phone. It turns Remote Control on for every new session, and it can run a server in a folder (such as ~/workspace) that starts at each login. Use when the user asks to reach sessions or start new ones from a phone, or to set up, check or undo Remote Control. Claude Code only.

★ 2
erickb336
erickb336/sage-stack

readme-guide

Write a project README. Apply when you write, change or review a README of any project, in any language.

★ 2