| name | skill-self-audit |
| description | End-of-cycle audit that detects repeated processes and creates skills for them. Invoked by the orchestrator before wrap-up. If a process was repeated more than once during the cycle, the orchestrator invokes boomerang-agent-builder to formalize it as a skill before signing off. |
Skill Self-Audit
Description
The orchestrator runs this audit at the end of every cycle. The audit scans the cycle for processes that were repeated more than once — and if any are found, the orchestrator creates a skill for them. This is how the Boomerang v3 skill library grows organically: as patterns repeat, they get formalized.
The Rule
If we did a process more than once this cycle, it should be a skill.
This is not optional. The orchestrator's wrap-up protocol includes this step explicitly. The user said:
"We need to at the end of every boomerang look at if we did a process if it would make sense to make it a skill. We need this also built in."
So this skill is built in.
When to Use This Skill
- The orchestrator is about to wrap up a cycle (every cycle)
- The user says "/boomerang-handoff" or the cycle is naturally ending
- A session is being terminated
- The user explicitly invokes "skill audit" or "self-audit"
Audit Process
1. Review the cycle's actions
Walk through the orchestrator's actions in the current cycle:
- Memory queries
- Architect dispatches
- Worker dispatches (Task tool calls)
- File reads / searches
- Skill invocations
- Context Package builds
- Status changes
Look for processes that recurred — meaning the same multi-step pattern was executed more than once, even if the targets were different.
Examples of "recurring process":
- Built 3 different Context Packages for 3 different cards (→ the "build Context Package" process should be a skill)
- Dispatched 4 coder tasks each with the same permission-check step (→ the "broker permission gate" should be a skill)
- Read AGENTS.md, TASKS.md, HANDOFF.md at the start of two different cycles (→ already a skill: the Session Start Protocol)
- Had to look up the same docs/roadmap path 5 times during dispatch (→ the orchestrator should pin the roadmap as a session variable)
- Translated the same kind of card evidence into the same kanban-board-manager call 6 times (→ the wrap-up evidence structure is already formalized; no new skill needed)
Examples of "not recurring" (do NOT make a skill):
- Read 1 file once to get context
- Asked the architect to design 1 thing
- Performed 1 dispatch
2. For each recurring process, decide
Ask three questions:
- Is the process already a skill? If yes, no action. Example: "build Context Package" is a documented pattern in the orchestrator skill, not a new skill.
- Is the process too narrow to be a skill? A skill should be reusable across projects and sessions. If the process is project-specific or one-off, document it in HANDOFF.md instead of creating a skill.
- Would formalizing it reduce token cost or increase reliability? If yes, create the skill. If it's just a "nice to have," defer.
If all three answer "no," don't create a skill.
3. If a new skill is warranted
Invoke the boomerang-agent-builder skill with:
- Process name — what to call the new skill
- Process description — what it does, when to use it
- Process steps — the exact sequence of operations
- Inputs — what the skill takes
- Outputs — what the skill returns
- Example invocation — a worked example
boomerang-agent-builder handles the rest: frontmatter, naming convention, syncing across the 3 skill locations, validating with npm run fix-perms and yaml_check.
4. If a skill should be UPDATED (not created)
If an existing skill is being applied in a way that wasn't anticipated, the orchestrator should:
- Note the deviation in HANDOFF.md
- Add a TODO to the kanban board for "improve to handle "
- Continue with the current cycle; do not block on the skill update
5. Record the audit result
Whether or not a skill was created, the audit result is recorded:
- Skills created this cycle: T-XXX → new skill name
- Skills marked for improvement: T-YYY → skill name → reason
- No new skills needed: a one-liner explaining why the cycle's processes were all unique or already covered
This is saved to memoryManager with project metadata so the next session can see what was learned.
Output Format
The audit returns a single Markdown block:
## Skill Self-Audit Result
### Skills Created
- **<skill-name>** — <one-line description>
- Triggered by: <which actions in the cycle>
- Replaces: <ad-hoc process description>
### Skills Marked for Improvement
- **<skill-name>** — <gap description>
- Tracked as: T-XXX on the kanban board
### No Action Needed
- <one-liner per non-issue, e.g. "The Context Package template is already in boomerang-orchestrator SKILL.md">
### Cycle Summary
- Total actions reviewed: N
- Recurring processes found: N
- Skills created: N
- Skills marked for improvement: N
Anti-Patterns
- Do not create a skill for a process that was performed only once. "Once" is data, not a pattern.
- Do not create a skill that overlaps with an existing skill. Update the existing one instead.
- Do not create skills for project-specific processes. Those belong in HANDOFF.md or the project's own docs.
- Do not skip the audit. The user asked for it to be "built in," and it is.
Model
Use Gemini for the audit reasoning. Skill creation itself (via boomerang-agent-builder) uses the agent-builder's recommended model (deepseek-v4-pro:cloud or higher).
