Goal Framing Agent Skill

Aegis Goal Framing

Use this skill to create a thin goal frame before execution. It is opt-in and

boundary-setting only.

Do not use it for tiny edits, one-command checks, or ordinary fast-path Q&A

unless the user explicitly asks for /aegis-goal or Aegis goal:.

Authority Boundary

Current owner:

• Method Pack task framing

Not owned here:

• authoritative GateDecision
• final evidence sufficiency
• final completion authority is not owned here
• host daemon / automatic stop enforcement

Input Forms

Treat these as equivalent:

• /aegis-goal <task description>
• Aegis goal: <task description>
• "Define the goal / stop condition before we start"

Slash commands are optional host shortcuts. The natural-language form is the

portable fallback.

Example:

Aegis goal: Fix the auth refresh bug without rewriting the auth system.

Output

Produce the smallest useful frame, then continue into the routed workflow in

the same turn.

TaskIntentDraft:
- Requested outcome:
- Goal:
- Success evidence:
- Stop condition:
- Non-goals:
- Constraints:
- Scope:
- Risk hints:
- Aegis Visibility:
- Route:
- Next:

Default behavior:

• Do not stop after TaskIntentDraft.
• Treat the frame as the start protocol for execution, not as the final answer.
• After the compact frame, immediately take the Next action for the selected

route when the user asked to do the work.

• Keep the visible frame natural and short when the goal is clear; do not emit a

large internal-looking card unless the user asked for a formal frame.

• Use Aegis Visibility to say why the goal frame constrains the route, stop

condition, or non-goals. Do not add trace ceremony unless the user explicitly

asks for auditability.

Frame-only behavior:

• Stop at the frame only when the user explicitly asks to only define the goal,

only define the stop condition, not execute, not implement, not write a plan,

or wait for confirmation before continuing.

• If required information is missing, state the missing input and stop with

blocked rather than pretending to continue.

Stop condition must distinguish:

State set: done, blocked, needs-verification, scope-exceeded.

• done: success evidence is satisfied
• blocked: required dependency, permission, or information is missing
• needs-verification: implementation exists but evidence is insufficient
• scope-exceeded: continuing would exceed the goal or non-goals

Routing

After framing:

• Low-risk single-owner work continues through the normal fast path or TDD
• Ambiguous product / architecture / contract work routes to brainstorming
• Approved requirements route to writing-plans
• Multi-step, compaction-prone, handoff, or subagent work routes to

long-task-continuation

• Bug diagnosis routes to systematic-debugging

Route Matrix

| Goal signal | Route |

| --- | --- |

| single-owner, low-risk, clear verification | fast path or test-driven-development |

| bug, failure, regression, unexpected behavior | systematic-debugging |

| ambiguous product, architecture, contract, cross-module behavior | brainstorming |

| approved spec, stable requirements, implementation slicing | writing-plans |

| multi-step, compaction-prone, handoff, subagent work | long-task-continuation |

| completion, release, handoff, "is this done?" | verification-before-completion |

Only create docs/aegis/ records when the routed workflow needs persistent

evidence. Goal framing alone does not create project files.

Subagent Context Packet

When delegating work, pass a compact packet instead of the full conversation:

SubagentContextPacket:
- Task:
- Goal:
- Stop condition:
- Relevant baseline refs:
- Relevant files:
- Known facts:
- Unknowns:
- Non-goals:
- Expected output:
- Verification expected:
- Must-read excerpts:
- Unsafe assumptions:

The packet reduces repeated file reading, but it does not replace evidence.

Subagents should still read the smallest raw file/log/test excerpt needed to

verify critical facts.

Do not paste full chat transcripts, full session history, or unbounded logs into

the packet. If a fact matters, include a file ref, line/window hint, or compact

must-read excerpt.

Drift Rule

If the goal changes mid-task, do not silently overwrite it. Record old goal,

new goal, changed scope, new risks, and route through DriftCheckDraft when a

long-task record exists.