pm

Three things pm enforces

Implemented as Claude Code hooks — hard-blocks on edits and prompt injection on every interaction.

What the agent sees

When the agent tries to edit a file but hasn't pulled the doctrines your settings require, the pre-edit hook returns this:

At the strongest settings, this is a hard-block — the pre-edit hook returns an error and the agent cannot write to any file until it runs the listed commands.

What a session with pm looks like

1. 1

   Session starts

   pm injects the doctrine router into context, briefs Claude on what's in progress, and resets the doctrine pull tracker. Claude sees the active task, recent work, and every recorded decision waiting to be searched.

2. 2

   You give a prompt

   The prompt-context hook fires. Claude sees the active work, your workflow settings (planning=all, questions=thorough, followup=thorough), relevant decisions matched against your prompt, and — if required doctrines weren't pulled yet — a BLOCKING notice with the exact commands to run.

3. 3

   Doctrines pulled

   Each pm doctrine <name> call is recorded silently. Once all required doctrines are read, the pre-edit hook unblocks.

4. 4

   Discovery

   Claude reads the planning doctrine, runs pm recap, runs pm why "<keyword>" against past decisions, presents an approach, asks clarifying questions per your settings, and waits for your confirmation.

5. 5

   Execution

   Tasks are created. Claude starts the first one and writes code. Each edit is allowed because there's active work. Each scope creep past 4 files is hard-blocked. Each Read and Grep is counted — past 3, the prompt-context surfaces a nudge to delegate exploration to a subagent.

6. 6

   Decisions recorded

   Every choice future-you would benefit from knowing gets logged with pm decide — what, why, and how to apply it. The next session will see them.

7. 7

   Sweep

   pm sweep closes outstanding items at the end of the conversation. Issues, stale tasks, orphaned drafts — gone. Every session leaves the project clean.

Decisions are always on

A decision is a five-line record: what you chose, why, and what future sessions should do about it. There is no toggle — decisions are recorded and surfaced on every prompt regardless of settings.

The prompt-context hook also surfaces matching decisions on every prompt — Claude rarely runs pm why on its own, so pm pushes them in.

Ten focused docs, pulled on demand

Small markdown files. The router is always loaded; the rest are pulled with pm doctrine <name>. Each pull is tracked, and the strongest settings hard-block edits until the right ones are read.

Five hooks in .claude/settings.json

pm writes these into .claude/settings.json on first run and re-verifies them on every invocation. They merge with your existing hooks. If any are missing, pm restores them.

Three settings, per project

Run pm settings to cycle. The values appear in every prompt-context injection so the agent can't drift away from your configured depth mid-feature.

Everything stays on your machine

All state lives in .pm/ as JSON files. No database, no server, no network calls, no telemetry.

Add .pm/ to your .gitignore and you're done.

Two commands

Every subsequent pm invocation re-verifies the hooks and silently restores them. Read the docs →