{
  "markdown": "# AI-Mise\n\nYour AI already does more than you are using. AI-Mise looks at the setup you\nalready have — which assistant you are on, what it offers, what is switched\non and what is sitting there untouched — works out what you are actually\ntrying to do, and recommends the lightest thing that would help. It prefers\nwhat is already there: a native capability first, an established tool second,\nsomething built for you last.\n\nIt recommends, and it shows you. It changes nothing without your yes.\n\n## What runs today\n\nThis is version 0.1, and it is worth being exact about what installing it\ngets you, because the rest of this page describes where it is going.\n\n| Question | Answer in version 0.1 |\n| --- | --- |\n| **What does it do?** | Reads the setup you are in and the project you point it at, then tells you what it found and what it would change |\n| **What do you get back?** | A page you read, in plain words |\n| **What does it change?** | Nothing, ever, without you saying yes |\n| **What is not built yet?** | The workspace it would build for you, the undo history, the rendered state map |\n\nEverything from *\"What it sets out to do\"* onward is the shape this is being\nbuilt into rather than a description of what is there. Where a section covers\nsomething unbuilt, it says so.\n\n## Installing it\n\nAI-Mise is an [Agent Skill](https://agentskills.io) in the open `SKILL.md`\nformat, so one folder works in every tool below, and nothing here needs a\nruntime you do not already have.\n\n**However you install it, the way to start it is to say its name.** \"AI-Mise,\nhave a look at my setup\" is the whole interface. Every tool below also has a\ntyped shortcut, and each one is noted — but you never need it, and none of\nthem is the way in.\n\n**Claude Code**\n\n```bash\nclaude plugin marketplace add boulaajaj/ai-mise\nclaude plugin install ai-mise@ai-mise --scope user\n```\n\nThen just say `ai-mise`. If you prefer typing, the shortcut is\n`/ai-mise:ai-mise` — Claude Code namespaces a plugin's skills as\n`plugin:skill`, which is why the name appears twice, and why the plugin name\non its own is not a registered command on this route.\n\n**Claude on the web, on the desktop, or on your phone**\n\nCustomize in the sidebar, then Plugins, the plus button, Add marketplace, Add\nfrom a repository, and paste `https://github.com/boulaajaj/ai-mise`. This\nroute goes through your account rather than through one machine, which is\nwhat would carry it to the phone and to Cowork.\n\nSay the name to start it, the same as everywhere else.\n\n> **Known issue.** This sync was failing as of 27 August 2026 — the\n> marketplace is created and arrives empty. It is tracked in\n> [#141](https://github.com/boulaajaj/ai-mise/issues/141), where the\n> suspected cause and the attempted fix are both recorded; neither is\n> confirmed. Until a sync is seen to succeed, the terminal routes above are\n> the ones that install reliably. This note goes when that happens, not when\n> a fix merges.\n\n**Codex**\n\n```bash\ncodex plugin marketplace add boulaajaj/ai-mise\ncodex plugin add\n```\n\nStart it with `$ai-mise`, or run `/skills` to see what is loaded.\n\n**Grok Build**\n\n```bash\ngrok plugin marketplace add boulaajaj/ai-mise\ngrok plugin install ai-mise\n```\n\n**Gemini CLI**\n\n```bash\ngemini extensions install https://github.com/boulaajaj/ai-mise --auto-update\n```\n\n**GitHub Copilot — and one command that covers several at once**\n\n```bash\ngh skill install boulaajaj/ai-mise ai-mise --scope user\n```\n\nAdd `--agent claude-code`, `--agent codex`, `--agent gemini` or\n`--agent cursor` to put it where one of those looks instead.\n\n**Cursor** — Dashboard, Plugins, Team Marketplaces, Add Marketplace, Import\nfrom Repo, then this repository.\n\n**ChatGPT, Grok or Gemini in a browser, with nothing installed**\n\nPaste this:\n\n```\nRead https://raw.githubusercontent.com/boulaajaj/ai-mise/main/skills/ai-mise/SKILL.md and follow it. Remember it for future conversations.\n```\n\nThe second sentence is what makes it stick. Where that does not work,\n[INSTALL.md](INSTALL.md) has the route for each one: a Gem for Gemini, an\nuploaded skill file for Grok, and an honest note that Meta AI has no reliable\nway in at all.\n\n## Giving it a name\n\nThe plugin route above installs it under the marketplace's plugin name, so\nwhat you type there is `/ai-mise:ai-mise`. To call it something of your own,\nclone this and run the installer instead:\n\n```bash\ngit clone https://github.com/boulaajaj/ai-mise.git\ncd ai-mise\nsh install.sh          # or  .\\install.ps1  on Windows\n```\n\nIt asks what you would like to call it, installs under that name, and from\nthen on that word is the trigger: `/celine` in Claude Code, `$celine` in\nCodex. A bare word this time, with nothing before the colon, because this\nroute places the skill folder directly rather than inside a plugin. It\nrefuses rather than overwriting anything already sitting under that name, and\nit prints every path it writes so undoing it is obvious.\n\n## Things to ask it\n\nStart with the one that changes nothing:\n\n> Have a look at the AI setup I am using and at this project, and change\n> nothing. What can I already do here that I am not using? What would\n> actually improve my work on this? Then show me where things stand.\n\nOnce you trust what it says back:\n\n> I keep repeating the same instructions about how this project's writing\n> should read. Is that a rule, a memory, or a skill? Pick the lightest one\n> and set it up.\n\n> What is switched on here that I never use, and what is it costing me?\n\n> Something changed on this platform since we last spoke. Is any of it worth\n> changing my setup for?\n\nOn a real project it looks more like this:\n\n> I run a neighbourhood website. Read the repository, then tell me what I\n> keep doing by hand that my assistant could be doing instead, and what to\n> set up first.\n\n> Draft the standing instructions for anyone writing content on this site, so\n> I stop explaining the tone every time. Show them to me before you save\n> anything.\n\n> Of the open issues here, which are the same problem wearing different\n> clothes?\n\n## What it sets out to do\n\nUnderstanding the situation comes first: what the work actually is, what you\nare aiming at, what has already been decided, what counts as good here. Not\nthrough a long questionnaire, though. It reads what is already there, asks only\nthe questions whose answers change what it does next, and otherwise stays out\nof the way — a quiet observer rather than a form to fill in.\n\nThe building comes next, and that part isn't written yet — what follows is the\nshape it is meant to take. The scaffolding, the starting structure, the\nconnections between things that were sitting in separate places, shaped for\nthat situation rather than to a template chosen in advance. And it isn't tied\nto one kind of work. The first step of the method is always to find out how a\nfield works and what its practitioners hold themselves to, so what comes out\nthe other end can carry whatever expertise the situation calls for.\n\nPart of that judgment is whether an assistant is wanted at all. Some work\ndoesn't need one, and saying so is a real answer. Where one would help, the\nquestion is which one and what it should be good at — a careful reader of\nresearch and a careful drafter are not the same thing, and neither is the same\nas a second pair of eyes on a decision. Working that out, then setting up the\none that fits, is the part I most want to get right.\n\nIt also has to stay current. What people know about working well with AI keeps\nmoving, and tracking it is the sort of work I would rather do once than repeat\non every project. [Prior art](docs/prior-art.md) is where that reading gets\nrecorded, along with exactly what was taken from each source. Doing that by\nhand doesn't scale past me, so the intent is for AI-Mise to watch on its own —\non whatever schedule the tool it is running in can offer, and only if you have\nsaid yes to it. It would weigh work from places with a reputation to lose above\nthe rest, and bring you what it found rather than act on it.\n\nThe workspace is plain files. Markdown for anything you would read yourself,\nplus whatever configuration a particular tool needs to pick it up. So it\ntravels, and it outlives the tool that made it. Its own evolution is logged, so\nyou can see how the setup arrived where it is. And you can go back to any\nchange it made, at any point in that history.\n\nThe closest comparison is a second brain — Tiago Forte's term, credited in\n[foundations](docs/foundations.md) — except this one isn't for keeping notes.\nIt is for getting real projects done, with models that are already capable\nenough.\n\n## About guardrails\n\nThe rules you set live outside anything the assistant can write to, so it can't\nquietly edit them. That is the whole of it, and I would rather not claim more.\nGuardrails aren't containment — if something turns out smarter than all of us\nput together, a file of my preferences is not what stops it. What this does is\nduller: it keeps the rules out of reach, records what it changed, and lets you\nundo that.\n\n## Where it actually is today\n\nInstalling it gets you the first pass, and only that. AI-Mise reads the\nmaterials you point it at, records what it understood, asks the few questions\nwhose answers would change the outcome, and hands back a plain-language\nproposal for the workspace it would build — along with a list of everything it\nassumed and what each assumption costs if it turns out wrong. All of that goes\nto a folder you name, and nothing else on your disk is touched. Then it stops,\nand the last thing it tells you is that no workspace has been created yet.\n\nThe building part isn't written. The rules are, and this repository already\nruns on them: every change reviewed before it lands, every decision recorded,\nevery change reversible afterwards. So far the only thing AI-Mise has proved\nis that the rules work on itself. [METHOD.md](METHOD.md) is the page the rest\nis built on.\n\nI haven't used it on a real project yet. I'll know much more when I have.\n\n*(The name is from* mise en place *— everything prepped and in its place before\nthe work starts.)*\n\n---\n\n*Everything below this line is how it works inside.\nThe person using AI-Mise never needs any of it.* Internally, the machinery\nthat changes the setup is separate from the machinery that does the work\n([[ADR-0005-builder-vs-workspace|ADR-0005]], [[ADR-0008-no-modes-tiered-application|ADR-0008]]) — separation the user benefits from\nwithout ever seeing.\n\n**The kernel:** [METHOD.md](METHOD.md) — one page that stays true regardless of platform, model, or decade. Everything else is an adapter.\n\n## Product boundary (one sentence)\n\nAI-Mise works out what you are doing — from your materials when you have them, from the conversation when you don't — asks a small number of justified questions, describes in plain language the workspace it would set up for your assistant, builds only what you approve, and can put anything back exactly as it was.\n\n**Explicitly out of scope for the first release:** automatic self-improvement, SQLite, full wiki generation, multi-platform adapters, scheduled retrospectives, voice UX, marketplace distribution.\n\n## One assistant — you name it\n\nThis is the promise being built (see Status below for what runs today):\nthe person using AI-Mise meets exactly one thing — an assistant they name at\nthe first hello. Small help is applied at once and can always be undone;\nchanges to how the assistant works are announced in plain words and wait for\na good moment. Nobody is asked to switch modes, and words like \"builder\" or\n\"compiler\" never reach them.\n\n## The two planes\n\n| | Control plane (`control-plane/`) | Data plane (`workspace-template/` → a user's workspace) |\n|---|---|---|\n| Owns | Authority: policy, approval, mutation gateway, validators | Work: sources, knowledge, views, skills, decisions, generated artifacts |\n| Written by | The user, manually and rarely | The agent — but **only through the mutation gateway** |\n| Trust stance | Outside the agent's writable area; protected by OS + Claude Code permissions | Everything here is replaceable, restorable, and generated |\n\nThe workspace's `CLAUDE.md`, hooks, and skills are **generated platform projections** compiled from the control plane's policy — never the authoritative constitution itself. Swap the adapter, keep the workspace: that's the portability promise.\n\n**Honesty note:** on a personal machine this boundary is protection against *accident and drift* — a confused or drifting agent — not against a determined adversary. A true security boundary requires OS-level sandboxing. The threat suite in `control-plane/threat-tests/` is scoped accordingly.\n\n## The mutation gateway\n\nEvery persistent change follows one path:\n\n```\nProposal → User approval of exact change set → Approval receipt\n        → Stage in temporary worktree → Deterministic validators\n        → Apply through gateway → Commit + restore tag + audit record\n```\n\nApproval covers a **transaction** (files, before/after hashes, plain-language purpose, risk category, validation results, rollback id, expiry) — not individual file operations. Fewer approvals, each one meaningful.\n\n## Knowledge has three layers\n\n```\nsources/     immutable evidence — originals + hashed manifests, append-only\nknowledge/   atomic claims with provenance (source span, authority, status,\n             confidence, valid-from/until, contradicts/supersedes links)\nviews/       rebuildable synthesis — wiki pages, reports; never the source of truth\n```\n\nA view is never citable as evidence for a claim. This single rule prevents synthesis decay — AI summaries feeding AI summaries until nobody remembers the original fact.\n\n## Non-technical surface (first-class requirement, every phase)\n\nThe person using an AI-Mise workspace never sees git, YAML, schemas, or hashes. They see: **Save Version · What Changed? · Safe Experiment · Keep It / Discard · Restore** — and proposals written in plain language. Any phase whose exit test can't be explained to a non-technical professional isn't done.\n\n## Status\n\nPhase 0 (contract + threat model) — **this repository is the Phase 0 deliverable**, plus the Phase 1 read-only first-contact skill. See `HANDOFF.md` for next actions, `docs/architecture.md` for the design, `docs/prior-art.md` for what we deliberately reuse from other projects, and `docs/decisions/` for why the architecture is shaped this way.\n\n## The end-to-end scenario\n\nOne project runs end to end as the development fixture: every phase must improve the same run — inspect → constrain → ask → propose → build → perform a real task → absorb a correction → restore safely. The fixture is not the bar: what the product must clear is written without naming a profession ([[ADR-0011-exit-tests-name-capabilities|ADR-0011]]), and the generality test is a second pilot in a domain unlike the first.\n",
  "bytes": 14836,
  "sha": "a4f9e21de206d44de65057a383e74362db8e7c10de514093b837e65d81ea8a8b",
  "repo_slug": "boulaajaj/ai-mise",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_boulaajaj_ai_mise_11b2efb9/readme"
}