Plan-Driven Development skills for coding agents.
PDD is a small skill pack that reduces agent rework by putting decision gates before expensive actions: product planning, code mutation, and release/push work.
Coding agents are fast enough to create new failure modes:
- They start building before the product direction is stable.
- They implement a broader change than the user asked for.
- They mix finished work, experiments, generated files, and release metadata.
- They push or tag before blockers are visible.
PDD is not a bigger prompt. It is a set of small, installable gates that make those mistakes harder.
| Gate | Install | Best for | Output you get |
|---|---|---|---|
pdd-zero |
--skill pdd-zero |
New projects, rough app ideas, MVPs, PRDs, vibe coding, 从0到1, 新项目 |
Project Launch Packet plus a Decision Ledger |
pdd-scope |
--skill pdd-scope |
Implementation, fixes, refactors, scripts, migrations, automation | Scope Lock with reuse-before-build and verification |
pdd-gate |
--skill pdd-gate |
Release planning, version bumps, commits, tags, pushes, 发版, 上线 |
Release Map separating ready, local, excluded, and blocked work |
User: Build this feature.
Agent: I made a broad plan, edited six files, added a helper, and updated release notes.
User: The release notes were not part of this. Also that helper duplicates existing code.
Scope Lock
Goal: Add the requested feature through the existing public interface.
Required behavior: ...
Non-goals: No release notes, no broad refactor, no dependency change.
Existing reuse options: Reuse the current validator and test pattern.
Minimal route: One handler change plus one behavior test.
Verification: Run the focused test and the related suite.
Stop conditions: Ask before changing the API shape or adding a dependency.
User: Can we tag and push this?
Agent: Tests pass, so yes.
Release Map
Public/tagged: v0.2.0
Pushed: rename and packaging fixes
Committed-only: none
Uncommitted: local training samples
Excluded: assistant-sidecar, private data captures
Blockers: notarization not ready
Decision: Yellow. Safe for an internal artifact, not a public release tag.
Install all three skills globally for Codex:
npx skills add llfzzz/pdd --skill '*' -g -a codex -yInstall one skill:
npx skills add llfzzz/pdd --skill pdd-zero -g -a codex -y
npx skills add llfzzz/pdd --skill pdd-scope -g -a codex -y
npx skills add llfzzz/pdd --skill pdd-gate -g -a codex -yList the skills before installing:
npx skills add llfzzz/pdd --listUse the convenience script from a clone:
scripts/install.sh all
scripts/install.sh zero
scripts/install.sh scope
scripts/install.sh gateGitHub CLI users with agent skill support can also install a specific skill:
gh skill install llfzzz/pdd pdd-scopeSee docs/install.md for more install options.
npx skills add llfzzz/pdd --list finds the three independent skills:
Available Skills
pdd-gate
Govern releases, version bumps, commits, tags, pushes, and publishing actions...
pdd-scope
Lock coding scope before a coding agent edits files...
pdd-zero
Turn a rough product, app, website, SaaS, game, automation, or coding project idea...
Use pdd-zero to turn this rough Mac app idea into a launch packet before coding.
Use pdd-scope before implementing this feature. I want the smallest safe change.
Use pdd-gate to check whether this branch is ready to tag and push.
Chinese trigger phrases are included in the skill descriptions where useful, so prompts such as 从0到1, 新项目, 发版, 上线, and 推送版本 can route naturally.
| If you already use... | PDD adds... |
|---|---|
| A brainstorming or PRD skill | A Decision Ledger that records accepted assumptions, blocked decisions, reversible choices, and hard-to-change choices |
| A generic implementation planner | A Scope Lock that includes non-goals, existing reuse options, verification, and stop conditions before mutation |
| A release checklist | A Release Map that separates public, pushed, committed-only, uncommitted, excluded, and blocked work before Git actions |
| A personal prompt template | Installable, separately selectable SKILL.md packages with eval prompts |
skills/
pdd-zero/
SKILL.md
evals/evals.json
references/project-launch-packet.md
pdd-scope/
SKILL.md
evals/evals.json
pdd-gate/
SKILL.md
evals/evals.json
Each skill is independently installable from skills/<skill-name>/.
- Plain Markdown skills: inspect every instruction before installing.
- No API keys, credentials, telemetry, or hidden executable payloads.
- The only script is scripts/install.sh, a small wrapper around
npx skills add. - Evals are included so users can see the behavior each skill is meant to produce.
- Vercel skills CLI documents installing all skills from a repo or selecting specific skills with
--skill. - Claude agent skills docs describe the
SKILL.mddirectory model, frontmatter, supporting files, and testing workflow. - Anthropic skills repository shows the public skill set convention and basic skill package shape.
MIT. See LICENSE.