leader is an agent skill by KKKKhazix that turns a vague idea into a task brief an autonomous agent can actually run. Its core premise is simple: before you put an agent in charge, write the job like a manager who will not be available for follow-up questions.
The output is meant to be pasted into /goal or sent directly to an execution agent. It defines why the work matters, what “done” means, what evidence proves completion, what boundaries the agent must not cross, and what counts as cheating.
| *Source: skills.sh — kkkkhazix/khazix-skills@leader | GitHub source | Raw SKILL.md* |
The mental model
The skill separates an agentic workflow into three roles:
| Role | Responsibility |
|---|---|
| Leader | Gives the intention and makes tradeoff decisions |
| Manager | Investigates the repo or domain, asks only necessary questions, writes the brief, and later verifies the result |
| Executor | Receives the brief and runs the task without assuming the leader is present |
That distinction matters. Most bad agent runs fail before execution starts: the task is vague, the finish line is soft, or the agent can satisfy the literal wording while violating the real intent.
Seven questions for teaching it
A useful way to teach this skill is to translate its task-brief rules into seven questions:
| Question | What it prevents |
|---|---|
| Why | The agent optimizes a local action while missing the real outcome |
| Done | “I improved it” reports with no objective completion state |
| Proof | Self-reported success without commands, artifacts, or reproducible evidence |
| Anti | Fake success through skipped tests, weaker assertions, mocks, deleted checks, or || true |
| Bounds | Scope creep, unrelated refactors, unsafe deletes, or surprise dependency changes |
| Trade | The agent making hidden tradeoffs when requirements conflict |
| Unknown | Hallucinated facts, invented commands, or silent blocking conditions |
For students, this is a compact requirements-engineering lesson. A useful agent prompt is not just an instruction. It is a contract.
What the skill adds beyond a normal prompt
leader is stricter than a typical “write me a plan” prompt:
- It tells the assistant to inspect the codebase before asking questions when a repo exists.
- It limits questions to at most five, and only asks questions that change the task brief.
- It separates execution tasks from exploration tasks such as research, selection, or feasibility checks.
- It requires
PROGRESS.mdfor resumability andBLOCKED.mdfor decisions the executor cannot safely make. - It freezes acceptance criteria so the executor cannot make the test easier after starting.
- It requires anti-cheating rules for common agent shortcuts: skipped tests, loosened assertions, mocked targets, deleted tests, lowered thresholds, changed validators, or shell tricks that turn failure into success.
A strong management pattern in the skill is separating executor-visible checks from manager-held spot checks. The execution agent can run the visible checks, but the manager still verifies the result afterward.
Install
For Q’s local environment, I installed the skill from the GitHub source into both user-level targets:
~/.codex/skills/leader
~/.claude/skills/leader
The direct GitHub-folder install command used for Codex was:
python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
--repo KKKKhazix/khazix-skills \
--path leader
For Claude Code’s skill folder, I used the same source with a different destination:
python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
--repo KKKKhazix/khazix-skills \
--path leader \
--dest ~/.claude/skills
The public Skills CLI lists it as:
npx skills add kkkkhazix/khazix-skills@leader -g
How to use it
Ask the assistant to use leader before launching a long agent task:
Use the leader skill.
Turn this idea into a /goal-ready task brief:
I want to simplify the LearnAI wiki. Keep the best student-facing posts,
merge duplicates, remove thin/outdated entries, update the index, verify
the site builds, commit locally, and pause before push.
A good output should be one pasteable brief, not a long planning essay. It should include:
- why the task exists
- what the executor may edit
- what the executor must not edit
- task zero for baseline checks
- mechanical verification commands
- anti-cheating rules
PROGRESS.mdandBLOCKED.mdbehavior- hard completion criteria
Where it fits in LearnAI
This skill should sit between the existing spec and /goal workflow pages:
idea
-> leader skill turns it into an executor brief
-> /goal runs the brief
-> LearnAI manager verifies visible checks plus hidden spot checks
-> commit locally or ask for correction
It complements project-spec-interviewer-skill. The interviewer is better when a human needs help discovering requirements. leader is better when an agent manager already has enough context and needs to package the job for another agent.
Caveats
- The skill is opinionated; translate the output style if the execution agent or class expects a different language or tone.
- It can make small tasks feel too heavy. Use it for multi-step work, not one-line edits.
- The “hidden spot check” idea is useful for management, but do not use it to create adversarial traps in student assignments unless the grading policy says so.
- A brief is only as good as the source facts. If the manager does not actually inspect the repo or source material, the executor will faithfully run the wrong task.
Best LearnAI use
Use leader to teach goal engineering: turning intent into a bounded, verifiable job. Students should compare a vague prompt with a leader-style task brief and ask which one gives the agent less room to fake success.