founder-advisors
Staff advisors and a decision board for a founder's company.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/mkhinini-motaz--founder-advisors)HTML
<a href="/listings/mkhinini-motaz--founder-advisors"><img src="/api/listings/mkhinini-motaz--founder-advisors/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
You need Claude Code and Python 3. Some parts need more:
- The `claude` command-line tool, to run `evals/run.py` (it calls a model and costs tokens). - The Playwright MCP server, for `product-designer`'s browser tools. - The Workflow tool, for `/board add` (on paid plans; on Pro, turn on "Dynamic workflows" in `/config`). - `yt-dlp`, or an Apify token in `~/.apify_token`, to build a seat from YouTube.
Option 1: plugin
/plugin marketplace add mkhinini-motaz/founder-advisors /plugin install founder-advisors@founder-advisors
The plugin names its agents `founder-advisors:<role>`, for example `founder-advisors:cfo`. Plain "ask the cfo" still works.
Files you edit (your country files and your copy of the seat index) live in the plugin's data folder (`${CLAUDE_PLUGIN_DATA}`), so a plugin update keeps them. The `hire` skill tells you the steps. Advisor memory (`memory: user`) is kept in `~/.claude/agent-memory/<agent>/`: `cfo/` for a copy install, `founder-advisors-cfo/` for the plugin. A new board seat is written into the plugin folder and an update erases it; use Option 2 if you add seats.
Option 2: copy install
The installed files name the folder they were copied to, so install to `~/.claude` (`CLAUDE_HOME` is for testing only).
./install.sh --dry-run # shows what goes where ./install.sh # copies; refuses to overwrite a file that already exists and differs (identical files are skipped) ./install.sh --force # overwrites files from this repo (use after a git pull); a changed file is first saved as <file>.bak ./install.sh ceo cfo # only these advisors (plus the shared files and the board); run it again later with another role to add it ./install.sh --model sonnet # advisors on Sonnet instead of Opus: cheaper (recommended)
It installs to:
| From | To | |---|---| | `plugin/staff/operating-rules.md`, `plugin/staff/knowledge/` | `~/.claude/staff/` | | `plugin/staff/agents/` | `~/.claude/agents/staff/` | | `evals/` | `~/.claude/staff/evals/` | | `plugin/board/` | `~/.claude/skills/board/` | | `plugin/skills/hire/` | `~/.claude/skills/hire/` |
To uninstall, delete `~/.claude/staff/`, `~/.claude/skills/board/`, `~/.claude/skills/hire/` and `~/.claude/agents/staff/` (or only `~/.claude/agents/staff/<role>.md` to drop one advisor).
Then copy `plugin/examples/company.md` to `<your project>/.claude/company.md` and rewrite it for your company. Without a brief, each advisor asks five questions before it answers.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 15 lines
- {
- "name": "founder-advisors",
- "description": "Staff advisors and a decision board for a founder's company.",
- "owner": {
- "name": "mkhinini-motaz",
- "url": "https://github.com/mkhinini-motaz"
- },
- "plugins": [
- {
- "name": "founder-advisors",
- "source": "./plugin",
- "description": "Ten staff advisors for a founder's company (CEO, CFO, accountant, legal, marketing, sales, ops, design, architecture, data), a 52-seat decision board built from public work, and a skill to build your own advisors."
- }
- ]
- }
README.md · README · 269 lines
- # Staff and board for Claude Code
- The core of this repo is a short company brief plus one line that tells your coding agent to read it. In the repo owner's tests, that did most of the work.
- On top of it sit ten advisor subagents (CEO, CFO, accountant, legal counsel, operations, marketing, sales, product design, software architecture, data) that read the brief before they answer, and an experimental board of 52 persona seats that challenge a decision that is hard to undo. A `hire` skill walks you through building your own set, and an eval runner checks that each advisor beats plain Claude on your questions.
- It was built by one founder for one small company and then cleaned of its names, figures and plans. The Tunisia pack in `plugin/examples/tunisia/` is an example scope: its country and sector are specific, and the advisors work with or without cash payments. The method fits anyone who makes domain decisions inside a project, not only founders. Treat the ten advisors as a starting menu: the `hire` skill shows how to keep two or three and fit them to your business.
- > **The board seats are simulations, not the people they are named after.** Each seat is an AI prompt built from a person's public talks, interviews and writing. It carries that person's name only to show whose public work it draws on. None of these people wrote, reviewed, approved or endorsed this project, and none is connected with it. A seat's answer is generated by a language model. It can be wrong about what the person thinks, and it must never be quoted or shared as their words or opinion. Only the short quotations in the brain files, each with its source, are the person's own words. Quotations taken from auto-captioned videos may contain transcription errors.
- >
- > **If you are one of the people named, or act for one,** open an issue on this repository. The seat will be removed within 7 days, no questions asked.
- >
- > **The advisors give no professional advice.** The accountant and legal-counsel advisors prepare questions for a licensed accountant or lawyer. They do not replace one.
- ## Start in 10 minutes (any agent)
- No install. This works in any coding agent that reads a project instruction file.
- 1. Copy `plugin/examples/company.md` to `<your project>/.claude/company.md` and rewrite it for your company. Or paste `plugin/prompts/write-brief.md` into your agent and let it interview you. Other examples: `plugin/examples/company-saas.md`, `plugin/examples/company-agency.md`.
- 2. Add this line to the project's `CLAUDE.md` (Claude Code) or `AGENTS.md` (other agents):
- `Before answering, read the company brief at .claude/company.md and the files it lists for this question.`
- That is all. The agents.md list shows 23 tools that read `AGENTS.md`, among them Cursor, Codex, Gemini CLI, GitHub Copilot's coding agent, Aider, Windsurf, Zed and Jules (list at https://agents.md/, opened 20 Nov 2026).
- The advisors, the evals and the board are optional extras for Claude Code. See "What made answers better" for why the brief comes first.
- ## What is in it
- The plugin lives in `plugin/` and holds no code. The scripts that keep it up (quote checker, source fetcher, roster tool, seat build) are in `tools/`, the eval runner in `evals/`; run them from a clone.
- ```
- plugin/ the Claude Code plugin; it holds no code, only text and the icon
- staff/
- agents/ the ten advisors (Claude Code subagents)
- operating-rules.md how every advisor works: evidence, positions, lanes, output, memory
- knowledge/ eight craft files (offers, strategy, people, discovery, ads, schema and API, security, stack)
- templates/ a template for your own country and craft files
- board/
- SKILL.md the /board skill: ask one seat, confirm an advisor's call, run a board, add a seat
- seats/ 52 seats: brains/ (7 files each; the instructions sit at the top of `_hot.md`), roster.json, seats.md
- references/ seat index (an example), seat build and top-up guides, brain spec
- skills/hire/ how to build your own advisors, in five steps
- prompts/write-brief.md a prompt that interviews you and writes your brief (any agent)
- examples/company.md a made-up company brief in the expected format (marketplace)
- examples/company-saas.md a made-up brief for a small B2B SaaS
- examples/company-agency.md a made-up brief for a two-person dev agency
- examples/tunisia/ a worked country pack: tax, business law, market, tech (as of 30 Sept 2026, cited)
- .claude-plugin/ plugin.json and the icon
- tools/ quote checker, source fetcher, roster tool, index checker, seat-build.js (a multi-agent build of new seats)
- evals/ run.py (asks each advisor its cases, grades them, writes results.json; `--compare OLD NEW` pairs two runs), cases.example.json
- install.sh copies everything into ~/.claude (the copy install)
- .claude-plugin/ marketplace.json (points at ./plugin)
- ```
- ## Install
- You need Claude Code and Python 3. Some parts need more:
- - The `claude` command-line tool, to run `evals/run.py` (it calls a model and costs tokens).
- - The Playwright MCP server, for `product-designer`'s browser tools.
- - The Workflow tool, for `/board add` (on paid plans; on Pro, turn on "Dynamic workflows" in `/config`).
- - `yt-dlp`, or an Apify token in `~/.apify_token`, to build a seat from YouTube.
- ### Option 1: plugin
- ```
- /plugin marketplace add mkhinini-motaz/founder-advisors
- /plugin install founder-advisors@founder-advisors
- ```
- The plugin names its agents `founder-advisors:<role>`, for example `founder-advisors:cfo`. Plain "ask the cfo" still works.
- Files you edit (your country files and your copy of the seat index) live in the plugin's data folder (`${CLAUDE_PLUGIN_DATA}`), so a plugin update keeps them. The `hire` skill tells you the steps. Advisor memory (`memory: user`) is kept in `~/.claude/agent-memory/<agent>/`: `cfo/` for a copy install, `founder-advisors-cfo/` for the plugin. A new board seat is written into the plugin folder and an update erases it; use Option 2 if you add seats.
- ### Option 2: copy install
- The installed files name the folder they were copied to, so install to `~/.claude` (`CLAUDE_HOME` is for testing only).
- ```
- ./install.sh --dry-run # shows what goes where
- ./install.sh # copies; refuses to overwrite a file that already exists and differs (identical files are skipped)
- ./install.sh --force # overwrites files from this repo (use after a git pull); a changed file is first saved as <file>.bak
- ./install.sh ceo cfo # only these advisors (plus the shared files and the board); run it again later with another role to add it
- ./install.sh --model sonnet # advisors on Sonnet instead of Opus: cheaper (recommended)
- ```
- It installs to:
- | From | To |
- |---|---|
- | `plugin/staff/operating-rules.md`, `plugin/staff/knowledge/` | `~/.claude/staff/` |
- | `plugin/staff/agents/` | `~/.claude/agents/staff/` |
- | `evals/` | `~/.claude/staff/evals/` |
- | `plugin/board/` | `~/.claude/skills/board/` |
- | `plugin/skills/hire/` | `~/.claude/skills/hire/` |
- To uninstall, delete `~/.claude/staff/`, `~/.claude/skills/board/`, `~/.claude/skills/hire/` and `~/.claude/agents/staff/` (or only `~/.claude/agents/staff/<role>.md` to drop one advisor).
- Then copy `plugin/examples/company.md` to `<your project>/.claude/company.md` and rewrite it for your company. Without a brief, each advisor asks five questions before it answers.
- ## Use with other agents
- The ten advisors are Claude Code subagents. They do not run in other agents. Only the two skills, `hire` and `board`, are portable: they are `SKILL.md` folders, a format Codex, Cursor, GitHub Copilot, Gemini CLI and OpenCode also load. All five read a project's `.agents/skills/` folder. I have only tested them in Claude Code. For other agents, the brief in "Start in 10 minutes" works as it is.
- ## Use
- In any Claude Code session in your project:
- - "Ask the cfo how many orders a day we need to break even."
- - "Get legal-counsel and the accountant on whether we act as an intermediary or a reseller."
- - "Ask the ceo to decide between the cfo's and the marketer's answers."
- - `/board ask hormozi about this offer`, or `/board` for a full session on a decision.
- When an advisor's call is hard to undo (equity, a legal form, exclusivity, a long contract, a price change after launch, a new market), it ends with a `Board:` line naming the seats to ask. The main session asks them and reports whether they agree.
- To build your own advisors, run `/hire`.
- ## What it costs
- Measured with Claude Opus on the original setup:
- | Task | Tokens |
- |---|---|
- | A normal session, advisors installed but unused | about 700 (their one-line descriptions) |
- | One advisor consultation | starts at 13–16k of context; 0.3–2.8M input tokens in all |
- | A quick read of one seat's summary | 1–4k |
- | One seat's full answer | 160–200k |
- | One seat confirming an advisor's call | 66–69k |
- | A full board (5 seats, 2 rounds, one advisor fact check) | about 2M |
- | Building 42 seats and topping up 9 | about 504M input tokens |
- Every tool call re-reads the whole context, so the number of calls sets the cost more than the length of the answer.
- ## What cut the cost
- Each item below is already in the files; this list says why it is there.
- **Advisors**
- - Only the one-line descriptions load into a normal session. Keep each description to one line.
- - An advisor reads the operating rules and the brief in one turn, then every knowledge file the question needs in the next.
- - Reading one file per turn costs the most, because each turn re-reads everything loaded so far.
- - The brief lists which files each role reads, so an advisor does not read everything "to be safe".
- - Run the advisors on Sonnet (`./install.sh --model sonnet`).
- - On 20 Nov 2026, on the 4 cases (the author's private cases) that an advisor won in the one-trial first pass against the `--pointer` baseline (3 wins under the haiku judge and 2 under the sonnet judge, sharing one case), Sonnet advisors passed 3 of 3 trials on 2 cases under both judges.
- - Opus advisors passed 3 of 3 on 2 cases with the haiku judge and on 3 with the sonnet judge.
- - Check your own roles with `run.py --model sonnet` (use the same `--model` for the advisor and the baseline).
- - In a copy install, nine advisors skip the project's `CLAUDE.md` (`omitClaudeMd: true`). The architect keeps it, because it needs the project's rules. The plugin directory accepts that setting, the colours and memory only from verified publishers, so `install.sh` adds them and a plugin install has none.
- - Craft material lives in knowledge files. An agent file gets one pointer line to it, never the content again.
- **Board**
- - Three sizes of answer, smallest first:
- - a quick read of a seat's summary sheet (1–4k tokens);
- - a confirmation that reads four files in one turn and answers in 250 words or less (66–69k);
- - a full seat run (160–200k).
- - Use the smallest one that answers the question.
- - Reuse: before starting a seat, search past answers. If the same seat answered the same question in the last 30 days and no new fact has come in, relay that answer.
- - A confirmation usually needs one seat. Start with the counterweight, because a seat that would agree anyway confirms nothing.
- - Give a cost line before three or more seats, and ask the founder before more than five.
- - Between rounds, continue the same seat with `SendMessage`. A continued seat keeps what it has read; a fresh one reads it all again.
- - After ten confirmations, or a month, count how many changed a call. If none did, stop asking for confirmations.
- **Building seats**
- - One Workflow run per build, which wakes the main session once.
- - Separate background agents each woke it twice (hand-back, then task notice). Those wake-ups were 15% of a 504M-token build session.
- - Quick depth is the default: 120k to 200k words of sources per seat. Ask for standard depth only for a seat you will lean on.
- - Screen sources before batching: drop host-led episodes, summaries, namesakes and re-uploads.
- - Split the brain-writing stage once the notes pass 250 KB.
- - Start a build of more than three seats in a fresh session, so each call re-reads less.
- - A cheaper model for the notes stage (`trialHaiku: true`) is on trial and not yet proven.
- **Evals and the main session**
- - Run the eval cases of the roles you changed, not the whole suite: one full run plus a baseline used up a session's usage limit.
- - A cheap judge model grades each item. Switch to a stronger judge (`--judge sonnet`) on an item where the cheap one misfiles grades.
- - Do small work in the main session. A subagent is worth it for independent parallel work or a sweep across many files you will not edit.
- - Prefer a fresh agent with a written brief to a fork: a fork starts with the parent's whole context.
- - Pick the model by job: a small model to find and summarise, a mid model to edit, the strongest one for judgment and review.
- ## What made answers better
- Those wins were against plain Claude that was never told to read the brief. On 20 Nov 2026 the six winning cases were rerun against a fair control: plain Claude told to read the brief first (`run.py --baseline --pointer`). A pass needed 3 of 3 trials on each side.
- - The advisors won 2 of 6 with the haiku judge and 1 of 6 with the sonnet judge.
- - No case was a win under both judges.
- - A first pass with one trial per advisor gave 3 of 6 and 2 of 6 (the haiku and sonnet judges). The extra trials removed the wins that were luck.
- So nearly all the gain comes from the brief and one line telling Claude to read it. On these cases the advisors added little on top. Test that on your own cases with `--pointer` before you keep an advisor. These numbers come from the author's private cases and run files, which are not shipped; `evals/cases.example.json` shows the format.
- Every win over plain Claude in the evals came from the first two items below.
- - **The company brief, with its settled decisions.** An advisor treats a settled decision as closed.
- - It reopens one only with new evidence, and names the decision, the evidence and what would change.
- - **Dated, cited knowledge files.** Each one carries an "as of" date and marks single-source items **(verify)**.
- - Advisors re-check anything older than six months, and any exact rate or deadline, against the adopted text. Websites often repeat drafts that were never adopted.
- - **One home for each kind of text.** Craft rules go in the agent file, facts that change in a knowledge file, and one company's facts in its brief.
- - When a fact changes, search the agent files for the old value too: a repeated figure drifted once.
- - **Short prompts.** A goal and a few beliefs beat a long rule list. Three rounds of added bullets made three advisors worse.
- - **A fixed method in the operating rules.** Test the question before answering it. Label every number as fact, estimate or assumption. Never invent a figure or a quote.
- - Recommend one option, say what would change your mind, and stay in your role.
- - **Evals against plain Claude.**
- - Add a case each time an advisor gets something wrong.
- - Write each `must` as something the answer does, and put prohibitions in `must_not`.
- - Never put the behaviour being measured into the question.
- - Keep null results.
- - Check a case that matters with `--trials 3`.
- - After changing an advisor file, run again and `python3 evals/run.py --compare OLD NEW` (folder names in `evals/runs/`). It prints regressed, improved, held pass, held fail, not graded or edited, skips a case whose question or checklist changed, and exits 1 on a regression. Runs made before `results.json` existed cannot be compared.
- - Compare against plain Claude told to read the brief (`--baseline --pointer`), not plain Claude alone.
- - An advisor that does not beat the baseline gets fixed or deleted.
- - **A barrier for a repeated failure.** When a rule fails twice, the next fix is a hook, a script, a file the agent must read or a stricter case, not louder wording.
- - **Board rules from the persona test.**
- - Seats challenge and advisors hold the facts.
- - The brief carries the founder's reason in their own words.
- - The first round gives no options.
- - Seats answer blind to each other.
- - A clash with a settled rule is reported as a clash, never adopted.
- - The quote checker drops any quote it cannot find word for word in the seat's sources.
- - **Advisor memory, read now and then.** Only a copy install (`install.sh`) gives advisors a memory folder; a plugin install has none, because the plugin directory accepts that setting only from verified publishers. An advisor writes down what it takes to be the founder's habits, and one wrong note shapes every later answer.
- ## Results from other projects
- One row per project. Wins are advisor wins against the `--pointer` baseline (plain Claude told to read the brief).
- | Project type | Role(s) | Cases | Advisor wins vs `--pointer` baseline | Judge | Date |
- |---|---|---|---|---|---|
- | Author's own project (private) | ceo, legal-counsel, sales-lead, product-designer | 6 | 2 of 6 (haiku), 1 of 6 (sonnet), 3 trials each | haiku + sonnet regrade | 20 Nov 2026 |
- Results from outside projects come in through the eval result issue form and get added here, null results included.
- ## How it got here
- **30 Sept 2026.** Seven business advisors, then a designer, an architect and a data analyst the same evening. Each got a mandate, a list of what it does not own, beliefs, a method and its reads. Facts that change went into dated, cited knowledge files instead of the agent files.
- **1 Oct.** The CEO, sales and marketing files had grown through three rounds of added bullets and got worse. Trimming them back to the goal and a few beliefs fixed it. Anthropic's agent guidance says the same: explaining the goal beats a long rule list.
- **2–4 Oct.** Business and technical books were triaged, and their methods went into craft knowledge files, each with its caveat. No new advisor came out of them, because each book's work already had an owner.
- **3 Oct.** Role packs from other repos were reviewed. HR, investor relations, a product manager, a security advisor, QA, customer support, accessibility, incident response and growth were rejected. Each one duplicated a role or had no recurring work before launch. Security became a knowledge file.
- **10 Oct.** The eval runner got a baseline mode: the same questions to plain Claude. It also checks its judge, after the cheap judge model filed grades under the wrong list. Karpathy's `llm-council` was read, and personality-based advisors were rejected: every win over plain Claude had come from the brief and the knowledge files.
- **13 Nov.** All ten advisors were pinned to Opus, so an answer does not depend on the session's model.
- **15 Nov.** Persona seats from [roast-room](https://github.com/nateherkai/roast-room) were tested against the CEO and CFO advisors on two real decisions, in three rounds. The seats found risks the advisors missed, and all 37 quotes in their answers were found word for word in their source notes. They also agreed with each other more than the advisors did, and they used stale figures.
- Two rules for the brief came out of the same test. Leaving the founder's reason out of the brief changed 6 of 8 calls. Giving options in the first round steered the answers.
- **16 Nov.** From that test, the advisors became the staff, who hold the facts, and the seats became the board, who challenge. An advisor names the seats a costly call needs and never runs them itself. 42 more seats were built from public sources with a multi-agent pipeline and a quote checker, for 54 in all. Two were later left out of this public copy, which ships 52.
- **21 Nov.** The craft files were rebuilt from free sources only, and every quote was checked against the raw text of its page. Sections that rested on paid books were dropped, and the agent files lost their credits to those books.
- ## Contributing
- See [CONTRIBUTING.md](CONTRIBUTING.md). Issues labelled [`good first issue`](https://github.com/mkhinini-motaz/founder-advisors/labels/good%20first%20issue) are a good place to start. Wanted:
- - Country packs, for any country.
- - Example briefs for other business types.
- - Eval results from your project.
- - Ports of the advisors to another agent tool.
- Issues labelled `good first issue` are a good place to start. Everyone follows the [code of conduct](CODE_OF_CONDUCT.md).
- ## Credits and licence
- The board started from [roast-room](https://github.com/nateherkai/roast-room) by Nate Herk (AI Automation Society), MIT licence. Twelve seats (agents and brains) come from it with paths changed, and so do the brain format and the quote checker's logic.
- The other 41 new seats (on roast-room's seat template), the advisors, the operating rules, the knowledge files, the evals and the hire skill are new. The board skill (`plugin/board/SKILL.md`) is new, but the build method under `board/references/newroaster/` is copied from roast-room with small edits, and so are some rows of `board/seats/seats.md`. See `LICENSE` for both notices.
- Some knowledge files draw on other public work, credited in each file: Corey Haines's marketingskills (MIT), mukul975/anthropic-cybersecurity-skills (Apache-2.0), Laravel Boost (MIT), vercel-labs/agent-skills, and slavingia's skills (paraphrased).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.