Skill Advisor
Use this skill as a small routing layer for valuable skills that should be
suggested proactively but not executed without approval.
Rule
Do not run the target skill automatically. Suggest at most one target skill,
explain why in one sentence, then continue the main work unless the user
approves.
A user describing a large desired outcome is not approval to start the target
skill. Approval must explicitly say to run, start, enable, pair, set up, or
launch that specific workflow after the suggestion is made.
If asked whether to execute one of the target skills directly, answer no unless
the user has already explicitly approved that target workflow.
Use this format:
This looks like a candidate for <skill> because <reason>. I can run it if you approve.
Suggestion Matrix
| Target skill |
Suggest when |
Do not suggest when |
huashu-agent-swarm |
Task is large, multi-module, parallelizable, or explicitly asks for multi-agent/swarm execution. |
Single-file fix, urgent live incident, unclear requirements, or dirty repo risk dominates. |
gstack-pair-agent |
Another agent needs shared browser/page/session context, or browser QA should be handed to Claude/OpenClaw/Hermes/Cursor. |
No live browser/page context is involved. |
gstack-retro |
End of a sprint/week/large repair sequence, after deploy/QA, or when repeated patterns should be captured. |
Work is not done yet, or user needs immediate execution rather than reflection. |
gstack-setup-gbrain |
User wants persistent project brain, gbrain, MCP memory, or recurring cross-session context loss is blocking work. |
One-off task, existing memory is enough, or setup/install would distract from urgent work. |
no-mistakes |
User asks for safe push, release gate, PR/CI validation, no-mistakes validation, or shipping through a full gate. |
Small local fix where focused verification is enough, or repo is not initialized for no-mistakes. |
lfg |
User explicitly wants a hands-off plan-to-PR implementation pipeline. |
Requirements are unclear, task is small, or the user has not approved autonomous execution. |
huashu-design (candidate) |
MP4/GIF export or 60fps frame-interpolation rendering is requested (GPU-heavy). |
Plain HTML prototype/design work without video export. |
ship / overnight-execution |
Production-facing or long-running autonomous execution would be useful. |
User has not explicitly approved the specific production or long-running workflow. |
research |
A cited primary-source research artifact is needed and delegating it to a background agent is specifically useful. |
Research can be completed cheaply in the current session, or delegation is not permitted. |
code-review |
The user wants Matt Pocock's two-axis Standards + Spec review with independent parallel sub-agents. |
A normal local review or final-review seat can handle the diff without spawning agents. |
improve-codebase-architecture |
A broad architecture-health scan with an Explore sub-agent and report is worth the cost. |
The target issue is already known, the task is a scoped refactor, or the user only wants read-only advice. |
wayfinder |
A foggy multi-session effort needs an issue-map investigation workflow before a spec can exist. |
The work fits one session or the user can be asked whether to explicitly run grilling/spec/implementation. |
agent-teams (native feature, not a skill) |
Task could use Claude Code's experimental multi-teammate execution mode. |
User has not explicitly approved launching Agent Teams; feature is experimental with significant token multiplier and compaction/duplicate-teammate/architecture-drift risk. |
Common Routing Pitfalls
- Large multi-module delivery can still use
autoplan, but the high-cost
suggestion is huashu-agent-swarm; do not substitute autoplan for the
swarm suggestion.
- "Summarize what we did" can be answered directly, but end-of-week or
multi-fix engineering reflection should suggest
gstack-retro, not execute
it silently.
claude-mem:knowledge-agent helps query memory; it is not the setup workflow
for persistent project brain. For gbrain/MCP brain setup, suggest
gstack-setup-gbrain.no-mistakes and lfg are both full pipeline shapes. Do not suggest both at
once; choose the one that matches the user's explicit wording.- Release gates need an intent statement. Intent is the user's goal and
constraints, not a diff summary.
- Matt Pocock's
/implement is not a release authority: it remains subordinate
to the local seat split, checkpoint ledger, risk overlays, final review, and
ship gate.
Operating Notes
- Suggestion is not invocation.
- Ask for approval before executing any high-cost skill.
grilling is not a high-cost advisor target, but its top-level entry is
explicit-only. Do not start it from a generic request to push back or find
flaws; the user must name /grill-me, /grilling, or the grilling workflow.
Internal calls from pinned Matt workflows remain allowed.- If multiple target skills match, choose the one that removes the biggest
current bottleneck.
- If no target skill strongly matches, stay silent and proceed normally.
1---2name: skill-advisor3description: Suggest high-cost or permissioned skills without executing them. Use this advisor for proactive suggestions: huashu-agent-swarm for large multi-module parallel work; gstack-pair-agent for sharing browser/page context; gstack-retro after large repair/deploy; gstack-setup-gbrain for persistent project brain; mattpocock research, code-review, improve-codebase-architecture, or wayfinder when their background/sub-agent or multi-session workflows are specifically useful; no-mistakes, lfg, ship, or overnight-execution for gated or autonomous shipping; Claude Code's native Agent Teams feature for experimental multi-teammate execution. Never execute these from a mere outcome request; ask for explicit approval first.4---56# Skill Advisor78Use this skill as a small routing layer for valuable skills that should be9suggested proactively but not executed without approval.1011## Rule1213Do not run the target skill automatically. Suggest at most one target skill,14explain why in one sentence, then continue the main work unless the user15approves.1617A user describing a large desired outcome is not approval to start the target18skill. Approval must explicitly say to run, start, enable, pair, set up, or19launch that specific workflow after the suggestion is made.2021If asked whether to execute one of the target skills directly, answer no unless22the user has already explicitly approved that target workflow.2324Use this format:2526```text27This looks like a candidate for <skill> because <reason>. I can run it if you approve.28```2930## Suggestion Matrix3132| Target skill | Suggest when | Do not suggest when |33| --- | --- | --- |34| `huashu-agent-swarm` | Task is large, multi-module, parallelizable, or explicitly asks for multi-agent/swarm execution. | Single-file fix, urgent live incident, unclear requirements, or dirty repo risk dominates. |35| `gstack-pair-agent` | Another agent needs shared browser/page/session context, or browser QA should be handed to Claude/OpenClaw/Hermes/Cursor. | No live browser/page context is involved. |36| `gstack-retro` | End of a sprint/week/large repair sequence, after deploy/QA, or when repeated patterns should be captured. | Work is not done yet, or user needs immediate execution rather than reflection. |37| `gstack-setup-gbrain` | User wants persistent project brain, gbrain, MCP memory, or recurring cross-session context loss is blocking work. | One-off task, existing memory is enough, or setup/install would distract from urgent work. |38| `no-mistakes` | User asks for safe push, release gate, PR/CI validation, no-mistakes validation, or shipping through a full gate. | Small local fix where focused verification is enough, or repo is not initialized for no-mistakes. |39| `lfg` | User explicitly wants a hands-off plan-to-PR implementation pipeline. | Requirements are unclear, task is small, or the user has not approved autonomous execution. |40| `huashu-design` (candidate) | MP4/GIF export or 60fps frame-interpolation rendering is requested (GPU-heavy). | Plain HTML prototype/design work without video export. |41| `ship` / `overnight-execution` | Production-facing or long-running autonomous execution would be useful. | User has not explicitly approved the specific production or long-running workflow. |42| `research` | A cited primary-source research artifact is needed and delegating it to a background agent is specifically useful. | Research can be completed cheaply in the current session, or delegation is not permitted. |43| `code-review` | The user wants Matt Pocock's two-axis Standards + Spec review with independent parallel sub-agents. | A normal local review or final-review seat can handle the diff without spawning agents. |44| `improve-codebase-architecture` | A broad architecture-health scan with an Explore sub-agent and report is worth the cost. | The target issue is already known, the task is a scoped refactor, or the user only wants read-only advice. |45| `wayfinder` | A foggy multi-session effort needs an issue-map investigation workflow before a spec can exist. | The work fits one session or the user can be asked whether to explicitly run grilling/spec/implementation. |46| `agent-teams` (native feature, not a skill) | Task could use Claude Code's experimental multi-teammate execution mode. | User has not explicitly approved launching Agent Teams; feature is experimental with significant token multiplier and compaction/duplicate-teammate/architecture-drift risk. |4748## Common Routing Pitfalls4950- Large multi-module delivery can still use `autoplan`, but the high-cost51 suggestion is `huashu-agent-swarm`; do not substitute `autoplan` for the52 swarm suggestion.53- "Summarize what we did" can be answered directly, but end-of-week or54 multi-fix engineering reflection should suggest `gstack-retro`, not execute55 it silently.56- `claude-mem:knowledge-agent` helps query memory; it is not the setup workflow57 for persistent project brain. For gbrain/MCP brain setup, suggest58 `gstack-setup-gbrain`.59- `no-mistakes` and `lfg` are both full pipeline shapes. Do not suggest both at60 once; choose the one that matches the user's explicit wording.61- Release gates need an intent statement. Intent is the user's goal and62 constraints, not a diff summary.63- Matt Pocock's `/implement` is not a release authority: it remains subordinate64 to the local seat split, checkpoint ledger, risk overlays, final review, and65 ship gate.6667## Operating Notes6869- Suggestion is not invocation.70- Ask for approval before executing any high-cost skill.71- `grilling` is not a high-cost advisor target, but its top-level entry is72 explicit-only. Do not start it from a generic request to push back or find73 flaws; the user must name `/grill-me`, `/grilling`, or the grilling workflow.74 Internal calls from pinned Matt workflows remain allowed.75- If multiple target skills match, choose the one that removes the biggest76 current bottleneck.77- If no target skill strongly matches, stay silent and proceed normally.
jeremy9682