Intuitive Init

Intuitive Init

Use this skill to create, audit, slim, merge, or refresh project-local AGENTS.md and CLAUDE.md guidance for Claude Code and Codex. The output should make the target repo easier for future agents without replacing local truth with a generic process manual.

This compact entrypoint preserves the full original guidance in references/detailed-guidance.md. Read that file for edge cases, full merge rules, init discovery variants, symlink migration, or mode-specific detail. Read references/lsp-and-mcp.md before editing LSP, Serena, MCP, language-server, or agent-facing tool setup.

Source Priority

Respect this source hierarchy:

1. current system/developer/user instructions;
2. repo-local human truth: README.md, ARCHITECTURE.md, STATUS.md,

docs/human/**, equivalent files named by the repo, and executable repo evidence such as package metadata, scripts, CI config, and tests;

1. existing project-local agent guidance and docs/agents/** operational

runbooks;

1. generated /init or codex init suggestions;
2. intuitive workflow defaults.

Generated init output is input to merge, not authority to overwrite local guidance.

Default Workflow

1. Inspect the target repo's current root docs and existing agent guidance.
2. Classify the request:

audit, apply/create, refresh, slim/cleanup, symlink migration, or LSP/MCP setup.

1. Preserve project-specific commands, hazards, test gates, and current source

of truth.

1. Move long operational procedures out of root guidance. Prefer the standard

docs/agents/operating-runbook.md for mixed repo-specific agent procedures; use topic-specific docs/agents/.md only when the scope is clearly independent. Use skills, hooks, scripts, or human docs when those are the better owner.

1. Check the high-frequency startup context: AGENTS.md, CLAUDE.md, and the

root orientation docs they require such as README.md, ARCHITECTURE.md, and STATUS.md. The first-read path should be bounded, newest-first, and task-routed; it should not require agents to reread injected files or consume every long root doc before a simple command.

1. For Codex/Paseo harnesses, preserve or add a short rule that XML-like host

control envelopes such as `, , , , and ` are orchestrator metadata unless accompanied by natural-language user intent. They must not be treated as a human stop request by themselves.

1. Prefer deterministic hooks/tools for lint, format, setup, and verification

rules instead of expanding root prose. Always inspect Python environment and submodule signals while refreshing startup guidance: .venv, uv.lock, pyproject.toml, .python-version, .gitmodules, and populated vendor/ submodule paths. When the repo already has .venv or uv/ pyproject.toml signals, explicitly recommend or add a .githooks/post-checkout script that prepares the worktree .venv with the repo's existing convention, plus a short root hint to enable it with git config core.hooksPath .githooks. When the repo has submodules, explicitly recommend or add the same hook to run git submodule sync --recursive for URL config only, and add root guidance for inspecting existing submodules before initializing local copies. Do not auto-run git submodule update --init --recursive for all worktrees by default; for mostly read-only submodules, tell agents to inspect the main checkout's initialized submodule and initialize a local worktree copy only when modification or isolated verification needs it.

1. For large repos and monorepos, prefer nested AGENTS.md / CLAUDE.md files

only when local scope differences are real.

1. Verify the final files are concise, local, and non-contradictory.

Root Guidance Shape

Root AGENTS.md and CLAUDE.md should answer:

• Why this repo has special rules.
• What the agent must know before acting.
• How to run setup, tests, verification, demos, and safe workflows.
• Where to find durable human docs and agent runbooks.
• Where fixed plan contracts, active capsules, and GSD-owned execution state

live.

• What must not be done in this repo.
• How host control metadata affects stop/continue decisions when the repo uses

Paseo or another orchestrator.

When root guidance names planning surfaces, keep it stable: canonical plans are docs/plans/.md, compact resume state is docs/status/active/.md, and .planning/* is owned by GSD tools. Do not recommend .continue-here.md, manual .planning/HANDOFF.json, or one-off prompt folders as default repo surfaces.

Avoid copying broad official docs or all intuitive workflow rules into each repo. Distill only the local invariant and point to local tools/docs for detail.

Orientation Doc Hygiene

When root guidance tells agents to read human docs before acting, make those docs fit that role:

• STATUS.md should be current-state first: latest material, next action, and

blockers near the top. If it includes recent changes, list them newest-first. Old shipped detail should be removed or replaced by links to plans, ADRs, retrospectives, or docs/human/**.

• README.md should orient and route, not become a full manual.
• ARCHITECTURE.md may be longer, but its first screen should summarize the

code map and layer contract so agents can decide what to read next.

• First-read policies should usually require only the already-injected

AGENTS.md plus STATUS.md or the active task/plan. Read README.md, ARCHITECTURE.md, CLAUDE.md, and longer runbooks only when the task needs them.

If the user asks for startup-context or first-read optimization, $intuitive-init may edit these root orientation docs in the same scoped pass. For broad human doc cleanup, route to $intuitive-doc.

LSP And MCP Setup

When the task is LSP, language-server, Serena, or MCP setup:

1. Read references/lsp-and-mcp.md.
2. Check whether the target repo already has a working project-local setup.
3. Prefer Serena MCP as the agent-facing LSP path for coding-agent repos unless

the target repo already has a better concrete setup or the host cannot run it.

1. Keep credentials, machine-local paths, and private endpoints out of committed

guidance.

Modes

| Mode | Use when | Output | Redirect when | | --- | --- | --- | --- | | Audit | Existing guidance may be stale, bloated, or missing local hazards. | Guidance health report and recommended edits; no changes unless asked. | The issue is human docs, tests, or code layout. | | Apply/create | A repo lacks local AGENTS.md / CLAUDE.md or needs initial setup. | Project-local guidance from repo evidence and accepted defaults. | The user wants only current-state human docs. | | Refresh | Existing local guidance needs current repo truth or init suggestions merged. | Updated guidance with stale/generic content removed. | Generated init output should only be reviewed, not applied. | | Slim/cleanup | Root guidance is overgrown, generic, duplicated, or stale. | Shorter root guidance with long detail routed to durable homes. | The long detail belongs in human docs owned by $intuitive-doc. | | Startup-context cleanup | First-read policy or orientation docs force too much context before work. | Bounded first-read policy, reasonable README / ARCHITECTURE / STATUS entrypoints, and standard docs/agents/operating-runbook.md extraction when needed. | The request is broad human documentation cleanup. | | Symlink migration | Root guidance is linked to shared/external files. | Project-local files preserving target repo rules. | The repo intentionally owns external guidance as its contract. | | LSP/MCP setup | The request names LSP, language server, Serena, or MCP setup. | Agent-facing setup or a concrete blocked/parked reason. | The target already has a better concrete setup. |

For non-trivial runs, state Selected mode:, Why:, and Redirect: before auditing or editing. For tiny direct changes, one sentence can carry the same information. Add a final Mode note: only when manual invocation, ambiguity, or a better owner matters.

Stop Conditions

Stop when root agent guidance is local, concise, and aligned with current repo truth; any required first-read docs are bounded and latest-first; long detail is routed to appropriate local docs/skills/scripts; and any LSP/MCP setup is either configured, documented as blocked, or explicitly parked.

Report changed files, verification run, remaining risks, and whether generated init output was merged, rejected, or not used.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: MiaoDX
• Source: MiaoDX/intuitive-flow
• License: MIT

Install and usage instructions live in the source repository linked above.