Goal Os Optimize

Goal OS Optimize — Self-Directed Agentic OS Maintenance

Uses the /goal slash command to have Claude Code autonomously audit and improve its own workspace — skill library, rules, CLAUDE.md files, and project catalog — with a judge agent on a separate LLM validating each iteration.

How /goal Works

/goal accepts a goal description (up to 4,000 characters) and runs a loop:

• Primary agent: works toward the goal
• Judge agent: runs on a different LLM model; after each iteration checks whether the done condition has been met
• Loop ends when the judge confirms the condition is satisfied, or the max iteration budget runs out
• Typical duration: 3 minutes (simple cleanup) to 1 hour (complex multi-project work)

The judge operates independently from the primary agent, providing adversarial validation on the primary agent's own progress claims.

Five Modes

1. Clean — Reduce and de-duplicate

Cuts skill count, removes contradictions in rules files, archives rather than deletes.

/goal
Audit the skills folder at ~/.claude/skills/ and any rules files for this workspace.
Goal: reduce the skill count by consolidating or archiving redundant and stale skills.
Archive all removed skills to ./archived-skills/ with a brief archival note — do not delete permanently.
For rules files, identify any contradictions between rules and write them to contradictions.md.
Done when: skill count is reduced, all removed items are archived (not deleted), and contradictions.md lists any conflicts found.

Adjust scope to your skills path and the consolidation threshold you want. The archive-not-delete constraint matters — state it explicitly so the judge enforces it.

2. Sharpen — Optimize against a rubric

Scores each skill against explicit criteria you define; rewrites skills that fail.

First, write a rubric.md that lists your quality criteria. Example:

# Skill Quality Rubric
1. Trigger phrase appears within the first 150 characters of the description field
2. Behavior steps are numbered and concrete — no vague "process the input"
3. Output contract is explicit: the user knows exactly what they will see
4. No hardcoded absolute paths or user-specific values
5. Skill body is under 200 lines

Then run:

/goal
Read rubric.md. Audit every skill in ~/.claude/skills/.
For each skill: score it against each rubric criterion (pass/fail per criterion).
For any criterion where fewer than 4 of 5 skills pass: rewrite the failing skills to meet that criterion.
Write scoring-log.md with before/after pass counts per criterion and a list of each skill that was changed.
Done when: all criteria score at least 4/5 skills passing and scoring-log.md is complete.

Writing your own rubric prevents the agent from setting a low bar it can easily meet. Do not skip this step.

3. Revive — Resurrect dormant projects

Classifies half-built projects, creates revival checklists for viable ones, archives the rest.

/goal
Go through every subfolder in ~/projects/ (or the specified projects directory).
For each project: check for git commits, runnable entry points, tests, and a README.
Classify each as: (a) production-ready, (b) revivable with under 2 hours of work, (c) dormant with no useful prior work.
For class (b): write a revival-checklist.md inside the project folder listing the exact steps to make it runnable.
For class (c): move to ./archived-projects/ with a one-line archival note explaining why.
Done when: all projects are classified, each (b) project has a revival checklist, each (c) project is archived.

Adjust the projects path to match your setup.

4. Forge — Extract recurring patterns as skills

Mines past Claude Code session transcripts for prompt patterns that recur without an existing skill.

Claude Code stores session transcripts as JSONL files in ~/.claude/ (and per-project .claude/ folders).

/goal
Go through all session transcripts in ~/.claude/ (look for *.jsonl files).
For each transcript: scan the user turns for prompt patterns that recur across multiple sessions with no associated skill covering them.
Identify the top 3 most frequent patterns that have no existing skill.
For each pattern: draft a SKILL.md using the standard format (YAML frontmatter with name + description, a Trigger section listing invocation phrases, numbered Behavior steps, and an Output Contract).
Save each draft to ~/.claude/skills/<name>/SKILL.md.
Done when: 3 skill drafts are written.

Review drafts before installing — extracted skills may have over-broad triggers that collide with existing skills.

5. Maintain — Continuous background audit via /loop

Combines /loop with /goal for a recurring maintenance cycle that runs as long as the session is open.

/loop 30m
/goal
Audit skills in ~/.claude/skills/ for staleness.
Archive any skill not modified or invoked in the last 30 days to ./archived-skills/ with a timestamp note.
Check CLAUDE.md and any rules files for contradictions between rules; log findings to maintenance-log.md.
If CLAUDE.md exceeds 150 lines, propose consolidations in maintenance-log.md (do not edit the file directly).
Done when: archived list is current, maintenance-log.md has a new entry dated today.

This runs every 30 minutes while the session is open. The maintenance log accumulates over time — review it periodically to see what keeps going stale.

Note: /loop keeps running until you close the session. Use this during active work periods, not for overnight runs.

Template Variables

Common Pitfalls

• Don't let the agent write its own rubric in Sharpen mode. It will pick criteria it can already meet. Always write rubric.md yourself first.
• Archive, don't delete. State this explicitly in the done condition — the judge will enforce what you describe.
• Review Forge output before installing. Skills extracted from transcripts may have vague or colliding trigger phrases.
• Maintain mode keeps running. Set a reasonable interval (30-60 min); shorter intervals waste turns on no-op audit passes.

Source Attribution

Technique: /goal for self-directed agentic OS optimization Source: Mark Kashef YouTube URL: https://www.youtube.com/watch?v=5xrjO38WUYY Published: 2026-05-17