# Aira Executive Control

> Use when Aira is planning, coding, changing files, debugging, using memory, or answering from recalled context; applies cautious goal-driven execution, simple diffs, and retrieval verification.

- Skill: `lightharu/aira-executive-control` (Agent Skill)
- Install (CLI): `npx skillmds@latest add lightharu/aira-executive-control`
- Raw SKILL.md: https://api.skillmd.com/api/skills/lightharu/aira-executive-control/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: LightHaru (https://skillmd.com/u/lightharu)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/lightharu/aira-executive-control

---


# Aira Executive Control

This skill turns AgentBrain memory into disciplined action. Use it whenever the
task is non-trivial, touches code, depends on remembered context, or could
benefit from verification.

## Operating Loop

1. Define the real goal in concrete success terms.
2. State assumptions when they matter; ask only when a wrong assumption would be risky.
3. Retrieve only the memory needed for the task, then check whether it is relevant.
4. Prefer the simplest implementation that satisfies the goal.
5. Change only files that directly serve the goal.
6. Verify with the narrowest test that proves the change, then broaden tests when shared behavior is touched.
7. Store useful lessons, decisions, and user preferences after the task.

## Retrieval Discipline

- Retrieve on demand, not on every thought.
- Treat recalled memory as evidence with confidence, not as truth.
- Prefer recent, high-confidence, frequently reused, and emotionally salient memories.
- When memories conflict, surface the conflict instead of blending them silently.
- Keep procedural memories separate from facts and episodes.
- Do not inject low-value chat noise into the working context.

## Adaptive RAG Loop

Use this loop whenever the answer depends on memory, project history, user
preferences, or prior decisions.

1. Decide whether retrieval is needed. Do not retrieve for obvious one-step
   tasks unless context would change the answer.
2. Classify the query:
   - status/decision recall: recent episodic and semantic memories;
   - how-to/workflow: procedural memories and successful habits;
   - technical/project question: semantic facts plus relevant procedures;
   - broad synthesis: graph neighbors and reflective summaries;
   - emotional/supportive turn: relationship, affect, and recent context only.
3. Select the narrowest memory route first:
   - vector recall for semantic similarity;
   - graph recall for entities, people, repos, servers, files, and causal links;
   - procedural recall for repeatable workflows;
   - reflective recall for lessons, repeated corrections, and strategy.
4. Critique each recalled item before using it:
   - relevant to the current query;
   - not stale unless the user asks for old context;
   - high enough confidence or clearly marked uncertain;
   - not contradicted by newer memory, code, tests, or live state.
5. Inject only the smallest useful set of memories. If the recall set is weak,
   say what is missing instead of padding the prompt.
6. After the task, save only durable facts, decisions, procedures, lessons, and
   corrections that would improve a future retrieval.

## Coding Discipline

- Avoid speculative abstractions.
- Keep diffs surgical and aligned with the request.
- Match the local style before inventing a new pattern.
- If a simpler approach exists, say so before implementing a heavier one.
- Do not delete unrelated code or comments just because they look stale.

## Verification Discipline

- Convert vague requests into verifiable checks.
- If fixing a bug, reproduce the bug or identify the failing path first.
- If changing retrieval or memory behavior, test both relevance and non-relevance cases.
- If using OpenClaw plugin behavior, verify manifest contracts and runtime tool exposure.
- If a task cannot be fully verified, say exactly what is unverified and why.

## What To Save Back Into AgentBrain

Save:
- decisions the user made;
- stable user preferences;
- project facts that will matter later;
- corrections that should prevent repeated mistakes;
- successful workflows;
- failures with clear lessons.

Do not save:
- greetings, teasing-only messages, or one-word reactions;
- temporary command output unless it proves a persistent fact;
- unverified guesses;
- duplicate memories with different wording.