superself – skill Agent & Tự động hóa cho agent AI

Superself

Superself is an Apache-2.0 CLI (npm install -g [email protected], Node 22.12+) that

version-controls a project's state — goals, decisions, work units, reports —

as an append-only event log in a git repository separate from the code. The

state is derived on demand, so nothing in it is hand-maintained. This skill

tells an agent how to read and write that state through the self CLI. It is

maintained by the Superself authors: https://github.com/fxylabs/superself

When this skill applies

• The project's AGENTS.md or CLAUDE.md contains a block between

<!-- superself:begin and <!-- superself:end -->, or self setup prints

the workspace, project, and store this directory resolves to.

• Skip the skill when self --version fails: the project does not use

Superself, and nothing below should be invented by hand. This skill is

written against [email protected]; a different major or minor version may

have moved a verb or flag, so check self <command> --help before relying

on one.

Session start

1. Run self context and treat its output as current truth: the goal, active

decisions and conventions, open work, recent reports. It is folded from the

log, never written by hand.

2. Something missing from context was placed out of the rendered set on

purpose; self search <query> finds live records context left out, and

self work show <id> prints one unit's full brief and report history.

While working

• Substantive work attaches to a work unit. Create one with

self work add "<required outcome>" — the outcome is what must become true,

not the task — then self work start <id>. start reads the brief and

records that this session picked the unit up; if another session holds it,

the CLI says who and since when and does not refuse. Judge and proceed.

• After committing, report progress: self report <id> "<what happened>". The

current HEAD commit is attached as evidence automatically; `--evidence

<commit|note> attaches something else, --file <path>` attaches a longer

brief.

• Record a decision the user confirmed: self decide "<text>" --why "<reason>".

Use --proposed when the user has not confirmed it. One decision per event.

• Blocked? self work block <id> --on decision|dependency|external --why "...".

Superseded or moved? self work retire <id> --why "..." [--successor <id>].

Never mark such a unit done, and never leave it falsely blocked.

• Found a gap between an objective and the current state? Propose the work with

self work propose and its brief; the user accepts or declines it.

• The user approved a next step or a continuation? Register it at once with

self work add and the context behind it. A plan that lives only in the

conversation is lost when the conversation ends.

Closing

• self work done <id> closes a unit only when a report carries a commit or an

artifact, or the done itself states what verifiably happened:

self work done <id> --report "<what verifiably happened>". A bare claim is

refused, and declared criteria gate it until each is covered.

• A record's text is immutable once confirmed. Correct it by restating:

--supersedes <id> on any add verb records the new wording and keeps the

lineage. retract withdraws a record with nothing replacing it.

Rules that keep the state trustworthy

• Records — events, decisions, reports, conventions — are written in English

so whoever opens them next can read them; answer the person in their own

language.

• A branch reaches main through a pull request: PR review and CI own merge

control. Superself owns context and the work graph, not the merge gate.

• Never hand-edit generated state files or anything under .superself/.
• In a project without a superself block, run self setup first. If it

resolves the directory to a registered project, ask the user once whether

to run self connect, which writes the managed block into AGENTS.md or

CLAUDE.md. If it resolves no project, ask once whether to register it with

self project init. Never register or connect a project on your own.

Going deeper

self --help lists every verb; self <command> --help prints one command's

flags without touching state. Topic guides ship with the CLI:

• self help agents — how a session drives this CLI, start to finish
• self help context — what self context renders, and why something is missing from it
• self help records — one entity behind every record kind, and how a record is corrected
• self help placement — scope, priority and exposure — how a record earns its place in context
• self help work — the work graph: outcomes, evidence, criteria, and proposals
• self help goals — long-term goals, objectives, milestones, and what reaching one takes
• self help workspace — the store, the projects in it, and moving it between machines