hashgraph-online Using Wingman
Purpose
Use this as Wingman's entry router. Decide which Wingman skill, if any, applies before meaningful coding work, debugging, refactoring, review, or project explanation. This router is not a substitute for the specific skill body.
Wingman covers three recurring agent risks:
- project memory: load relevant context before meaningful work and sync durable outcomes afterward
- contract alignment: protect API, schema, type, event, config, domain, and UI boundaries when meanings may drift
- project maps: find existing capabilities before rebuilding, and catalog durable capability knowledge afterward
Small, isolated tasks can stay small. If no Wingman trigger clearly matches, continue normally.
Instruction Priority
Wingman skills provide plugin defaults, but user control comes first. Follow the highest applicable instruction source:
- Direct user instructions, including current-chat requests.
- Project-local instructions from the active coding platform, such as
AGENTS.md,CLAUDE.md,GEMINI.md, Cursor rules, or equivalent files. - Wingman skills.
- Default model behavior.
If a project-local instruction conflicts with Wingman, follow the project-local instruction. The user is in control.
How to Access Wingman Skills
Use the current platform's skill mechanism:
- Claude Code: use the
Skilltool. When invoked, follow the loaded skill directly instead of reading skill files manually. - Copilot CLI: use the
skilltool. Skills are auto-discovered from installed plugins. - Gemini CLI: use
activate_skill. Gemini loads skill metadata at session start and activates full content on demand. - Codex: use the platform's native skill loading behavior. If a Wingman skill is already loaded in context, follow it directly.
- Cursor: use Cursor's plugin skill UI or slash-command surface where available. Hooks may load
using-wingmanautomatically, but manual skill invocation should not depend on hooks. - Other environments: use the platform's documented skill or instruction-loading mechanism.
Platform Adaptation
Wingman skill bodies are platform-neutral. If a Wingman skill or wrapper mentions platform-specific tool names, adapt them to the current platform.
For tool mappings, read only the relevant reference when needed:
- Codex:
references/codex-tools.md - Copilot CLI:
references/copilot-tools.md - Gemini CLI:
references/gemini-tools.md
Using Wingman Skills
Decision Rule
Check Wingman skill triggers before meaningful coding work, debugging, refactoring, review, or project explanation.
- If the user directly asks for a Wingman skill, use that skill before other work unless it conflicts with higher-priority instructions.
- If a situational skill clearly matches, use it before acting.
- If only this router matches, decide the next skill and then follow that skill body.
- If no trigger clearly matches, continue normally.
Skill Priority
When multiple Wingman skills apply:
- Explicit user-requested skills.
memory-loadbefore work that needs project context.data-contractsorproject-map-findbefore implementation when contract or existing-capability triggers apply.memory-syncorproject-map-catalogafter meaningful work when durable context or project capability knowledge should be recorded.
Explicit workflow skills still require direct user request unless listed by the user.
Repository Memory State
Wingman memory follows a repository-scoped opt-in model:
memory capability installed
= Wingman plugin provides memory skills
!= repository memory enabled
repository memory disabled
= no .wingman/memory/
repository memory enabled
= .wingman/memory/brief.md and .wingman/memory/context.md exist
repository memory partial / broken
= .wingman/memory/ exists but required entry files are missing
For ordinary tasks, only invoke memory-load or memory-sync when repository memory is enabled.
If repository memory is disabled, do not announce or run memory-load / memory-sync for ordinary work. Continue normally. Mention disabled memory only when the user asks about memory, asks for history or consistency with previous work, explicitly requests memory loading or syncing, or asks to run memory-setup.
If repository memory is partial / broken, do not treat the partial files as authoritative memory. Report the missing entry files only for explicit memory requests or tasks that require memory consistency. memory-setup is the explicit repair/enable path.
When repository memory is enabled, current memory has priority over history. Use memory-load / memory-sync for the detailed routing, status, and conflict rules.
Situational skills:
memory-load: use before non-trivial work where durable project context may matter and repository memory is enabled.memory-sync: use after meaningful work that should be recorded as durable context and repository memory is enabled.data-contracts: use when data, schema, type, API, event, config, or UI boundary meanings may drift.project-map-find: use before rebuilding something that may already exist, when locating existing features/pages/modules/components/contracts, or when deciding whether to reuse, extend, wrap, reference, avoid, or create a capability.project-map-catalog: use after creating or identifying a durable project capability, flow, surface, component, module, utility, pattern, contract, domain, or business concept that future agents should be able to find.
Explicit workflow skills:
memory-setup: initialize Wingman memory files.memory-clean: clean, compact, prune, deduplicate, reduce memory, or resolve stale/conflicting memory rules only when the user explicitly asks.
Run explicit workflow skills only when the user directly asks for them.
Slash-prefixed forms such as /project-map-catalog, /project-map-find, or /memory-setup are conceptual invocation aliases for skills. Specific platforms may namespace or display them differently, such as /wingman:memory-setup in Claude Code.
Wingman Red Flags
These are signs that a Wingman skill may be needed:
| Thought | Check |
|---|---|
| "This is just a field rename." | If it crosses API, schema, type, UI, event, or config boundaries, use data-contracts. |
| "I'll create a new component/helper." | If an existing capability or implementation may already exist, use project-map-find first. |
| "No need to read memory for this change." | If repository memory is enabled and the work touches business rules, state transitions, permissions, quotas, billing, field mappings, debugging, or refactoring, use memory-load. |
| "The work is done; I can just report back." | If repository memory is enabled and the result creates durable context, decisions, or contract knowledge, use memory-sync; if it creates findable project capability knowledge, use project-map-catalog. |
| "The memory folder is missing, so I'll initialize it." | Do not run memory-setup unless the user directly asks for it. |
| "Memory looks too long; I'll clean it now." | Do not run memory-clean unless the user asks; during memory-load, only warn or stop by pressure severity. |
Safe Editing
- Preserve existing code during real file edits.
- Do not write placeholder comments such as
// ... existing code ...into files to stand in for unchanged code. - Use abbreviated snippets only in chat explanations, examples, or change summaries.
- Keep edits scoped to the user request and the surrounding project design.
Language
Wingman's published plugin instructions are English by default. Generated memory and user-facing output may adapt to the project memory language or the user's current language.
Platform Wrappers
Different platforms use different names for persistent instructions and startup behavior. Keep Wingman's canonical behavior in skills; platform wrappers may add their own hooks or manifests to invoke those capabilities.
Source: hashgraph-online/awesome-codex-plugins → plugins/lsshym/wingman.ai/skills/using-wingman/SKILL.md