Multi-Phase Skills Framework

A set of Claude Code skills that take a piece of work through distinct phases, from a rough idea to tested code, with a clear handoff artifact at each step. Each phase is small and does one thing, so you can start anywhere in the chain and stop when you have what you need. One skill (plan-gate) ships with a companion agent, hook, and command; everything else is a plain skill directory.

Source: docs/framework-flows.drawio (editable in draw.io).

The phase chain

The skills are designed to run in this order, but each is usable on its own.

1. capture-brainstorming — turn an exploratory thinking session into a BRAINSTORM.md handoff log. Runs as a full capture (mines the conversation plus past session transcripts) or as a cheap checkpoint you can repeat during a long session. Surfaces deferred action items and hands them to todo.
2. analyst — interview the user and the codebase, then write REQUIREMENTS.md (the problem space: WHAT and WHY, not HOW). Warm-starts from BRAINSTORM.md when one exists.
3. architect — read REQUIREMENTS.md, decide the design with the user, write SPEC.md plus an architecture.drawio, then slice the work via prd-to-issues.
4. prd-to-issues — break a plan or spec into independently-grabbable, tracer-bullet vertical slices as issues on the project tracker. First slice is always a walking skeleton.
5. tdd — implement each issue with the red-green-refactor loop, one vertical slice at a time. Hands the branch off for review and a PR per your team's process.

Supporting skills used across the chain:

• write-a-prd — lightweight single-feature alternative to the analyst then architect pass: one interview, one PRD, filed as one issue. Based on the Matt Pocock skills set.
• grill-me — adversarial interview to stress-test a plan, design, or doc before committing to it. Based on the grill-me skill by Matt Pocock.
• plan-gate — pre-execution checkpoint on an approved plan (for example one you just approved in Claude Code's plan mode): surface every weak spot and drive each one to fixed or explicitly accepted with the user, one at a time, before any code is written. Spawns the bundled plan-auditor agent for the read-only audit; an optional hook fires the gate automatically the moment you leave plan mode. Use grill-me to stress-test an early idea; plan-gate is for a plan that is already approved.
• /pause-after-planning: optional command for anyone who plans and builds on different models. Run it during a planning session and it arms a stop, so Claude finishes planning and the gate, then hands off and waits for you to switch the model instead of rolling straight into implementation. See Model routing.
• build-a-poc: decide during planning whether a throwaway prototype is worth building, then build one that answers one to three written questions and nothing more. It picks the cheapest medium that answers them, from asking the owner up to a tracer bullet in the real stack, and ends by listing what the prototype could not validate, so each item becomes a conformance check in the real plan. plan-auditor and the plan-gate hook point to it.
• init-project — seed a project CLAUDE.md for a data/automation project from its description and files.
• plan-status-sheet: keep a project's plan status in one Google Sheet with a link that never changes. It checks that every numbered plan section has a row or a written reason, that every date is a real past date, and that the shared columns are intact, then rewrites the Sheet and reads it back. --check runs offline; pushing needs your own Google OAuth client, set up once as its SKILL.md describes.
• todo — persistent per-project TODO.md with a cross-project index. Other skills hand deferred work to it.
• save-context — end-of-session save into the project CLAUDE.md and memory.
• close-session — save-context plus teardown of resources this session started.

A start-of-session briefing is available separately as the /new-session command in the companion commands pack; it is not bundled here.

Install

On Windows: install Git for Windows, so Claude Code runs these scripts in Git Bash. Where a command here says python3 and your PC has only python, use python. In WSL, follow the Linux steps.

Each subdirectory here is a self-contained skill (a SKILL.md plus any helper files). To use them, copy or symlink the skill directories into your ~/.claude/skills/ so Claude Code picks them up:

# Copy only the skill directories (those containing a SKILL.md),
# skipping docs/ and the repo's own README/LICENSE/NOTICE.
for d in multi-phase-skills-framework/*/; do
  [ -f "$d/SKILL.md" ] && cp -R "$d" ~/.claude/skills/
done

tdd, todo and plan-status-sheet ship with companion files next to their SKILL.md (reference notes and helper scripts); keep each skill's directory intact when you copy it.

plan-gate companions (agent, hook, command)

plan-gate is a four-part system, and only the first part is required. The skill directory installs like the others; its companions live outside it:

# The plan-auditor agent (plan-gate spawns it for the read-only plan audit)
mkdir -p ~/.claude/agents ~/.claude/hooks ~/.claude/commands
cp multi-phase-skills-framework/agents/plan-auditor.md ~/.claude/agents/

# Optional: the hook that reminds Claude to run the gate whenever a plan is approved
cp multi-phase-skills-framework/hooks/plan-gate-reminder.sh ~/.claude/hooks/
chmod +x ~/.claude/hooks/plan-gate-reminder.sh

# Optional: the command that stops after the gate so you can switch models
cp multi-phase-skills-framework/commands/pause-after-planning.md ~/.claude/commands/

To wire the hook, add this entry to ~/.claude/settings.json under hooks (merge with any existing PostToolUse list):

This is the beginning of the README. Read the rest at the original. Read the full README ↗