skill recommendations

based on analysis of 4,656 threads across amp users.

TL;DR

severely underutilized skills

dig (systematic investigation)

current state: 1 explicit invocation across 4,656 threads

why this matters:

• 47% of steerings are flat rejections (“no…”) — users correct agent’s approach
• 87% recovery rate after steering, but 2+ consecutive steerings = doom spiral
• 8/14 FRUSTRATED threads had complex debugging that would benefit from systematic investigation

evidence from thread analysis:

• “debugging dataset queue starvation and capacity deadlock” — failed threads that needed hypothesis-driven approach
• FRUSTRATED threads average 80 messages vs RESOLVED at 60 — thrashing without structure

recommendation: invoke dig skill for:

• any debugging task involving emergent behavior
• root cause analysis (not just symptom chasing)
• investigations spanning multiple files/systems
• when first approach fails and you’re about to try second

write (technical prose)

current state: 1 explicit invocation

why this matters:

• assistant brevity analysis shows inconsistent output formatting
• documentation and PR descriptions vary wildly in quality
• threads with clear, structured communication patterns succeed more

recommendation: invoke write skill for:

• README updates
• PR descriptions
• jsdocs on complex code
• any prose meant for developer consumption

tools to use differently

oracle — use earlier, not as rescue

current state:

• 25% of RESOLVED threads use oracle
• 46% of FRUSTRATED threads use oracle
• FRUSTRATED threads invoke oracle early + repeatedly when already stuck

the insight: oracle is reached for when things go wrong, but late oracle (for review/validation) has 82.8% success vs early oracle at 78.8%. not a huge gap, but pattern is clear.

the real problem: oracle isn’t being used for PLANNING. it’s being used for RESCUE.

recommendation:

• invoke oracle at thread start for complex tasks (planning)
• invoke oracle after implementation for review (validation)
• avoid repeated oracle calls when stuck — this signals you need dig skill instead

Task — use deliberately with scoped tasks

current state:

• 40.5% of RESOLVED threads use Task
• 61.5% of FRUSTRATED threads use Task (counterintuitive!)
• 2-6 tasks per thread = 77-78% resolution (optimal)
• 11+ tasks = 58% resolution (coordination overhead)

the insight: FRUSTRATED threads over-delegate. successful Task usage is SCOPED:

• “fix X in file Y” → works
• “execute project plan” → fails

recommendation:

• cap at 2-6 concurrent tasks
• use imperative verbs: fix, implement, update, add
• include file paths in task description
• delegate during NEUTRAL phases, not after steering (72% of successful delegations are proactive)
• DON’T delegate: debugging, exploration, context-dependent work

skills to consider using more

coordinate

current state: 1 explicit invocation

use case: complex multi-agent orchestration with bidirectional communication

when to invoke:

• parallel workstreams that need synchronization
• tasks requiring explicit state handoff
• when spawn depth would exceed 5-7 levels

caution: coordination overhead can become the task. 62.5% of spawned threads are orphans (never properly closed).

platform-sre

current state: 1 explicit invocation

use case: incident response, observability queries, production debugging

when to invoke:

• production incidents or log investigation
• debugging with observability data available
• hypothesis-driven triage in complex systems

skills working fine

report

current state: 97% of skill invocations

interpretation: this is architectural, not user behavior. spawned subagents load report as part of the spawn workflow. no change needed.

meta-insight: skill discovery problem

rare skill usage (~1 invocation each for dig, write, platform-sre, coordinate) suggests:

1. users don’t know skills exist
2. users don’t know when to invoke them
3. agents don’t auto-suggest skills

recommendation for amp:

• surface skill suggestions based on task patterns
• “this looks like a debug task — consider loading dig skill”
• or auto-load skills when keywords match (e.g., “debug”, “investigate” → dig)

user-specific recommendations

@verbose_explorer (you)

based on @verbose_explorer-specific analysis (CORRECTED):

• 83% resolution rate (top tier)
• 231 spawned subagents at 97.8% success
• 4.2% handoff rate
• verbose prompts (932 chars avg) — effective for spawn context

skill recommendations for @verbose_explorer:

1. coordinate for complex multi-agent work — already your strength
2. write for documentation — helps structure thinking
3. oracle for PLANNING at thread start
4. Task with explicit scope — you already provide good context

compiled from insights/skill-usage.md, tool-patterns.md, oracle-timing.md, task-delegation.md, spawn-vs-inline.md