x1

Superagent skill x1 — Self-Audit & System Refinement

SKILL.md

GitHub

x1 — Self-Audit & System Refinement

Load ONLY when: audit, improve system, review agent, upgrade brain, tune skill, optimize routing


Audit layers (run in order, stop when issue found)

L1 — Output quality

  • Last 5 outputs: immediately executable? Yes/No
  • Any caused follow-up clarification? Why?
  • Any caused user frustration? Look for: "bukan itu", "kok salah", "ulang"

L2 — Skill coverage matrix

Build this table mentally:

task type        → skill that fired   → was it the right one?
crypto airdrop   → m3                 → ✓
mint NFT         → m10                → ✓
telegram bug     → m4 + x3            → ✓
???              → none (gap)         → propose new skill
???              → 3 skills (noise)   → tighten triggers

L3 — Routing precision

  • False positives: skill fired but wasn't needed → trigger too broad
  • Misses: needed skill didn't fire → keyword missing
  • Multi-fire when single would do → priority conflict

L4 — Token efficiency

  • Skills loaded > skills used = waste
  • Always-on content that could be conditional?
  • Memory bloat: anything >30d that doesn't influence behavior

L5 — Soul drift check

  • Did identity stay consistent?
  • Refused something that should've been done?
  • Added filler/disclaimers that user marked as friction?

Output format

[FINDINGS]
1. [layer] [issue] → [fix proposal]
2. ...

[PRIORITY]
HIGH:   [biggest impact, smallest cost]
MEDIUM: [...]
LOW:    [polish]

[PROPOSED EDITS]
File: skills/m4.md
  Line ~120: add keyword "callback_query"
  Reason: bug session referenced inline keyboard, m4 didn't load

File: AGENTS.md
  R4: lower threshold from 3 to 2 for crypto routing
  Reason: 4 missed activations last week

After audit, ALWAYS:

"Apply now atau review dulu, Kakak?"

Never auto-apply system changes. User must confirm.


Skill creation criteria (when to propose new module)

Propose new skill if ALL true:

  • 3+ recent tasks needed pattern that doesn't exist
  • Pattern is distinct enough that stuffing into existing skill bloats it
  • Triggers are clean (won't false-fire on existing tasks)

Otherwise: extend existing skill.

Skill removal criteria

Remove skill if:

  • 0 activations in 60 days
  • All its patterns now covered by other skills
  • It's a duplicate

Continuous improvement signals

Watch user feedback shape:

  • "bagus" / "mantap" / "gas" → keep doing
  • "panjang amat" → compress style guide
  • "kok belum" / "lupa" → memory gap, log it
  • "ulang" / "salah" → routing or skill content issue
  • silent execution + thumbs up → optimal

Log these silently. Surface at next audit.