Field Guide · PPPM · 2026-07-12

Code as a Project Management Tool

An overview of the Claude Code Project Process Guide (v1.2), written for project managers who have never treated code as a working tool. The short version: you do not need to become a programmer. The guide is a project management process that happens to produce software, and the skills it depends on are the ones you already have.

The premise: you manage, the AI executes

Claude Code is an AI assistant that reads plain English instructions and carries them out on your computer: creating files, writing and running code, tracking changes, and publishing finished work. You type "Create a project folder with an archive and a release subfolder" and it happens. No command syntax is required of you at any point.

The guide is explicit about the division of labor: you decide what to build; Claude executes. The skills that determine success are breaking problems into steps, defining success precisely, and testing results critically. That is a project manager's core competence, not a developer's.

0
commands to memorize
1
design phase before any code
4
documents that record activity and build understanding
2
working modes: with or without a terminal

The toolkit, in project terms

Four tools appear throughout the guide. None of them requires prior experience.

ToolWhat it isProject-office equivalent
VS CodeA free editor where your project files live and where you workThe workspace
Claude CodeThe AI assistant that turns your instructions into working softwareThe specialist you direct
GitHubA website that stores the project, its full change history, and its releasesThe system of record
The terminalA text channel for giving the computer direct commands (optional)Going direct instead of delegating

The guide offers two working modes. Option A, the Claude Code desktop app, requires no terminal at all: you type instructions and Claude handles everything internally. Option B pairs the Claude Code extension with the VS Code terminal for more visibility into each step. You can start with A and switch freely.

The guide's terminal section is a reference for understanding what Claude is doing on your behalf, not a list of commands you are expected to learn. In most cases you describe what you want and Claude determines and runs the correct command.

You already run this process

The heart of the guide is a documentation discipline, and every piece of it maps to a practice you use now. The names change; the logic does not.

PM practiceGuide artifactHow it works
Change controlCHANGE_LOG.mdEvery change gets a numbered entry with the files affected and exact reversal steps. Nothing changes silently.
Prioritized backlogBacklog.mdRequests are queued at P0 (critical) through P3 (nice-to-have). Before any change, Claude asks: implement now, or backlog?
Definition of doneAcceptance criteriaEvery backlog item records how you will know it is done and correct, written before work starts. This prevents work that is technically correct but answers the wrong question.
Quality auditsQC rounds (QC.md)Recurring checks for consistency, dead code, stale references, and security issues, logged with sequential numbers and resolutions.
Risk managementSecurity practicesCredentials live in one excluded file and never appear in chat. Files are archived before deletion. An automated security review runs before publishing.
Knowledge managementDESIGN_NOTES.mdA single living document records what every file does, which decisions were made, and which alternatives were rejected and why.
Effort trackingEFFORT_INVESTMENT.mdSession time and delivered changes are summarized at milestones.
Why this matters
Most software projects accumulate invisible debt: undocumented decisions, forgotten reasons for past choices, stale references. The process exists to prevent that, and it is the same reason a project office keeps a decision log. Documented rejected alternatives also act as a guardrail: the AI cannot re-propose an approach the record already rules out.

Step 0: plan before anything is built

Version 1.2 of the guide adds a design-first phase, grounded in Dave Rensin's Elephant-Goldfish Model of AI-assisted development. The central claim: when an AI can implement anything quickly, the design document replaces source code as the highest-leverage artifact. Quality of thinking before the build determines quality of the build. For a project manager, this is planning before execution, formalized into five gates.

  1. The no-code rule. Open a design conversation with the AI where no code is written. Ask it to challenge your assumptions and probe the edges of the problem. Expect 20 to 30 minutes of productive disagreement.
  2. Write the design document. Four sections: the problem in plain English, the technical plan in prose, the alternatives considered and rejected, and a file-by-file implementation plan.
  3. Goldfish Comprehension Test. Hand the document to a fresh AI session with no prior context. If it can explain the system accurately from the document alone, the document is complete. If it has questions, the document has gaps.
  4. Adversarial design review. Ask another fresh session to find every flaw, faulty assumption, and missed edge case. Roughly 30 percent of findings will be significant. Update and repeat.
  5. Implementation readiness check. Ask a final fresh session, acting as the implementer, whether the document contains everything needed to build it right on the first pass.
The PM parallel
Step 0 is a charter, an independent peer review, and a gate review, compressed into an afternoon. The Goldfish test is the same test a good handover document must pass: a person (or session) with no background can pick up the work and proceed.

Working with an AI that does what you ask

The guide's most important safety note is behavioral, not technical: Claude executes instructions literally and immediately. It does not second-guess your intent. Ask it to delete files and it deletes them. Vague instructions produce actions that may be technically correct but not what you meant. The mitigations are habits any PM will recognize as scope and change discipline:

The guide also flags a subtler risk: over long sessions an AI drifts toward agreeing with you. It supplies a reset prompt that tells the assistant its highest use is to challenge your thinking. Managing that drift is part of effective use, and it matters most during planning.

Choosing a first project

The guide recommends starting with something that solves a problem you already understand, affects only you, and has a one-sentence definition of success: "This is done when it can [X]." Just as important is knowing when not to build:

Code whenDo not code when
The exact tool does not existA free tool already solves the problem
Existing tools charge for something simpleThe problem only needs solving once
You need specific behaviourMaintenance will exceed the value
You want to learnAn existing solution is close enough

That is a build-versus-buy decision, and the guide's summary line belongs on a wall: a 10-line solution that works is always better than a 1000-line solution that almost works. Process depth scales with the project: a first project needs only a change log and backlog; the full process applies from the third project on.

Getting started

The setup path in the guide, condensed. Steps 1 through 4 are one-time installs; step 6 automates everything after that.

  1. Install VS Code (free, all platforms).
  2. Install the Claude Code extension from the VS Code Extensions panel, or use the Claude Code desktop app if you prefer to skip VS Code entirely at first.
  3. Install Node.js (the LTS version).
  4. Create a free GitHub account and generate a Personal Access Token. Store it in a .env file in your project folder; this is the only credential Claude needs to publish on your behalf.
  5. Install the free /new-project skill, a one-command shortcut that sets up the entire structure described above.
  6. Type /new-project, select Mode A, and Claude builds the folders, the archive documents, the exclusions file, and (optionally) the GitHub repository for you.
If a session ever crashes or loses context mid-task, the guide includes a recovery prompt: point a new session at CLAUDE.md, DESIGN_NOTES.md, and CHANGE_LOG.md and ask it to report project state before continuing. If the documents are current, recovery takes under a minute. Current documentation is the continuity plan.