Agent Goal
Write one /goal line that a coding agent can keep pursuing until the transcript proves the work is done.
This skill owns the slash-goal artifact. Do not use it for ordinary prompt
drafting, handoffs, continuation prompts, delegation briefs, or prompt
engineering.
Shape
/goal <single objective> in <lane> until <observable evidence proves completion>. First read <starting context>. Work in checkpoints and surface <validation evidence> after each checkpoint.
Answer these in order, then compress them into one line:
What should change?
Where should the agent start?
What evidence proves it changed?
When should the agent stop?
Include only execution-critical detail:
- Objective: one concrete outcome, not a backlog of unrelated wishes.
- Lane: files, packages, apps, issues, specs, or the intended work area.
- Starting context: the first files, plans, logs, screenshots, or acceptance criteria to read.
- Evidence: command output, tests, build result, screenshot check, eval score, file count, or reviewed artifact.
- Stop condition: the exact transcript-visible state that means the goal is achieved.
Rules
- Start the answer with
/goal when the user asks for the goal text.
- Make completion judgeable by someone who can only read the transcript.
- Prefer exact checks over vague proof: "
bun test packages/auth exits 0" beats "tests pass."
- Point long requirements at a file instead of pasting them into the goal.
- Ask for checkpoints when the work spans multiple turns.
- Bound repeated failure: after three failed attempts on the same check, report the root cause and next decision.
- Do not create a
/goal for vague wishes, open-ended research, unrelated chores, or work with no observable evidence.
Platform Note
Different agents may evaluate goals differently, but the portable rule is the
same: the goal should not rely on hidden state. Tell the agent to run checks and
surface evidence in the transcript.
Check
Before handing back a goal, verify:
- It has one main objective.
- It names where to start.
- It names the evidence that proves completion.
- It tells the agent to surface that evidence.
- It tells the agent when to stop.
1---2name: agent-goal3description: Draft a `/goal` line that gives long-running Codex or Claude Code work an objective, validation evidence, and a stop condition. Use when the user explicitly asks for `/goal`, an agent goal, or a completion condition.4---56# Agent Goal78Write one `/goal` line that a coding agent can keep pursuing until the transcript proves the work is done.910This skill owns the slash-goal artifact. Do not use it for ordinary prompt11drafting, handoffs, continuation prompts, delegation briefs, or prompt12engineering.1314## Shape1516```txt17/goal <single objective> in <lane> until <observable evidence proves completion>. First read <starting context>. Work in checkpoints and surface <validation evidence> after each checkpoint.18```1920Answer these in order, then compress them into one line:2122```txt23What should change?24Where should the agent start?25What evidence proves it changed?26When should the agent stop?27```2829Include only execution-critical detail:3031- Objective: one concrete outcome, not a backlog of unrelated wishes.32- Lane: files, packages, apps, issues, specs, or the intended work area.33- Starting context: the first files, plans, logs, screenshots, or acceptance criteria to read.34- Evidence: command output, tests, build result, screenshot check, eval score, file count, or reviewed artifact.35- Stop condition: the exact transcript-visible state that means the goal is achieved.3637## Rules38391. Start the answer with `/goal` when the user asks for the goal text.402. Make completion judgeable by someone who can only read the transcript.413. Prefer exact checks over vague proof: "`bun test packages/auth` exits 0" beats "tests pass."424. Point long requirements at a file instead of pasting them into the goal.435. Ask for checkpoints when the work spans multiple turns.446. Bound repeated failure: after three failed attempts on the same check, report the root cause and next decision.457. Do not create a `/goal` for vague wishes, open-ended research, unrelated chores, or work with no observable evidence.4647## Platform Note4849Different agents may evaluate goals differently, but the portable rule is the50same: the goal should not rely on hidden state. Tell the agent to run checks and51surface evidence in the transcript.5253## Check5455Before handing back a goal, verify:5657- It has one main objective.58- It names where to start.59- It names the evidence that proves completion.60- It tells the agent to surface that evidence.61- It tells the agent when to stop.
epicenterhq