compact

Compact Skill

Use this skill when:

• The conversation is long, repetitive, or approaching the model context window.
• The user says compact, compress context, summarize session, reduce tokens, or continue without losing state.
• The assistant needs to hand off the session to another AI or future turn.

Operating Principle

Compact is not a casual summary. It is a state-preservation operation. The output must keep the agent able to continue the work without rereading the whole transcript.

Required Output

Produce a concise handoff with these sections:

1. Current Goal — the exact user objective being pursued.
2. Decisions Locked — decisions already made, with rationale.
3. Files/Systems Touched — paths, modules, APIs, UI surfaces, tests, and why they matter.
4. Current State — what is already implemented, verified, or still failing.
5. Fresh Tail — the most recent user intent and immediate next step.
6. Context Ledger — what was kept raw, what was summarized, what was trimmed, token pressure, and the reason for each decision.
7. Risks — likely bugs, missing tests, stale assumptions, or context that must not be lost.
8. Do Not Touch — files/folders owned by other parallel agents.

Compression Rules

• Keep exact file paths and API names.
• Keep test commands and their pass/fail status.
• Keep user preferences and privacy constraints.
• Keep memory/project scope decisions.
• Keep token budget pressure, compact trigger, preserved fresh-tail count, and any context ledger id if available.
• Remove duplicate narration, old exploration branches, and raw tool output unless it is needed to reproduce a bug.
• Do not claim something was verified unless a test, build, or visual check actually ran.

Token Budget

Aim for 800-1,500 words for a normal session. If the context is under severe pressure, compress to 300-600 words and keep only: goal, decisions, touched files, current state, next action, blocked folders.