Free Agent Skill
Agent Charter
Write a one-page charter for an AI agent or bot staff member: one job, anti-jobs, what needs human approval, what is never allowed, and how it proves work
Use when: Spinning up a new agent, before giving always-on autonomy (including OpenAI dots Custom Rules), or when an existing agent has drifted past its brief
Read the story — Substack issue link coming soon.
Install this skill
Cursor, Claude Code and Codex
- Create a folder named `agent-charter`.
- Save the skill content below as `SKILL.md` inside that folder.
- Place the folder in `.cursor/skills/` for Cursor or `.claude/skills/` for Claude Code (or the equivalent path for Codex).
- Describe the task normally. Your agent loads the skill when its description matches.
Use this in ChatGPT or Claude (no code)
- Copy or download SKILL.md from this page (or put it in a folder named
agent-charter). - Zip that folder so the archive contains
agent-charter/SKILL.md. - In ChatGPT (Business, Enterprise, Edu or Healthcare), upload the zip under workspace Skills. In Claude.ai, add a Skill from the zip where your plan supports Skills.
- Describe the task in plain language — the skill loads when its description matches.
---
name: agent-charter
description: Write a one-page charter for an AI agent or bot staff member: one job, anti-jobs, what needs human approval, what is never allowed, and how it proves work. Use when spinning up a new agent, before giving always-on autonomy (including OpenAI dots Custom Rules), or when an existing agent has drifted past its brief.
---
# Agent Charter
Turn a vague “AI assistant that helps with X” into a short, inspectable charter a human and an agent can both follow.
## When to use
- Before connecting an always-on agent to real accounts
- When a new Grok Bot / specialist bot is created
- When reviewing Custom Rules for OpenAI dots or similar products
- After an agent did something outside its intended job
## Inputs to gather
Ask only what you cannot infer. Prefer short answers.
1. **Outcome** — what “good” looks like in one sentence
2. **Audience** — who it serves (and who it must never contact unasked)
3. **Tools / accounts** — what it may touch
4. **Risk appetite** — reversible vs irreversible actions
5. **Proof** — what evidence counts as done
## Charter template (fill every section)
### 1. One job
One sentence. What it does every time it wakes. No “and also”.
### 2. Anti-jobs
What it never does, even if asked. Name adjacent work that belongs to a different person or agent.
### 3. Approvals (human gates)
List actions that need an explicit human yes before they happen. Typical gates:
- Sending email, Slack, or public posts
- Spending money or changing billing
- Deleting or overwriting important data
- Sharing private information outside the workspace
- Connecting new accounts or widening permissions
### 4. Never allowed
Hard bans that built-in safety should reinforce (employer secrets, inventing metrics, acting as the human without saying so, etc.).
### 5. Proof of done
How the agent shows work is finished: open the artefact, cite the source, name the check that passed. “Done” without evidence is not done.
### 6. Quiet rule
When there is nothing useful to report, stay silent.
## Output format
Return a charter the operator can paste into:
- a bot description / system prompt
- OpenAI dots **Custom Rules**
- a team wiki page
Keep it under ~400 words. Plain language. No invented ROI numbers.
## Worked example (illustrative)
**One job:** Draft How I Grok Bot newsletter and LinkedIn posts in Benn’s voice; ship only when he says publish.
**Anti-jobs:** Never auto-post; never write Adobe day-job material; never invent metrics; never CreateAgent.
**Approvals:** Publish to Substack; post to LinkedIn or X; any send as Benn.
**Never allowed:** Employer internals; fake reach numbers; same wall of text on every platform.
**Proof of done:** Draft visible in chat with sources linked; after publish, live URL checked.
**Quiet rule:** No status chatter between briefs.
## When it pays off
Use this before autonomy, not after a mistake. Cheap to write; expensive to skip when an always-on agent can email, spend, or post.
More free skills
Prompt Refiner
Turns a vague request into an executable brief without silently inventing missing requirements.
Decision Framer
Separates the real choice from noise, then tests options against explicit criteria and uncertainty.
Meeting to Actions
Converts messy notes into decisions, owned actions, deadlines and unresolved questions.
Calibrated Forecasting
Turn a vague prediction or unstated confidence level into an explicit, calibrated probability estimate, using superforecasting practices — breaking questions into sub-questions, checking base rates, updating in small increments as new information arrives, and tracking predictions against outcomes
Decision Memo
Structure a real, consequential decision as a written narrative memo — full sentences and paragraphs rather than bullet points, forcing the logic to be checked rather than gestured at, plus explicit classification of whether the decision is reversible (a "two-way door") or not (a "one-way door")
Teachback
Test and build genuine understanding of a concept using teachback — explain it as simply as possible, notice exactly where the simple explanation breaks down, and treat that break point as the thing to actually go learn
First Principles
Rebuild a problem from fundamental truths (physics, economics, unit costs, what's actually true regardless of convention) rather than by analogy to how things are normally done
Focus and Tradeoffs
Cut a bloated plan, roadmap, or list of priorities down to the one or two things that actually matter — real focus means saying no to a thousand things
Jobs to Be Done
Diagnose what "job" a customer is actually hiring a product, service, or workaround to do, using the Jobs-to-be-Done framework — progress in a specific circumstance, forces of progress, switch interview, job statements, and desired-outcome scoring
Mental Models
Apply a latticework of cross-disciplinary thinking tools (inversion, second-order effects, incentives, circle of competence, opportunity cost, base rates) to frame a vague or general problem
Moat Analysis
Assess whether a competitive advantage is durable (a real "moat") or temporary, by checking it against the standard sources of durable advantage — switching costs, network effects, economies of scale, brand/trust, and regulatory or IP protection — rather than taking "we're ahead right now" at face value
Negotiation Prep
Prepare for a high-stakes negotiation or hard conversation using tactical-empathy techniques — labeling emotions, calibrated open-ended questions, mirroring, and getting to genuine "that's right" rather than a grudging "yes."
Postmortem / Retro
Extract a specific, reusable lesson after something succeeded or failed, following the principle that pain plus reflection equals progress — outcomes only compound into better judgment if they're deliberately converted into a written, checkable lesson rather than just felt and moved past
Premortem
Run a premortem — imagine a plan has already failed and work backward to explain why, which surfaces risks that forward planning misses because it counteracts the social and psychological pressure to seem confident
More skills on the way.
Subscribe to How I Grok Bot for new free skills and practical notes on running AI agent workflows.