Unified Engineering 0.3.1
One skills-only engineering plugin combining Ponytail's simplicity, Superpowers' development methods, Addy Osmani's specialist engineering expertise and selected Matt Pocock methods, with refinements informed by the Karpathy-inspired community guidelines. This is an independent adaptation, not an official release or endorsement by those projects.
Use it
Select the plugin/skill with ChatGPT's @ picker where available, or ask in ordinary language:
- “Use Unified Engineering to implement this change.”
- “Use Lite mode and fix this bug.”
- “Use Ultra mode, but follow this approved design.”
- “Review this diff only; do not edit files.”
- “Use the security skill to check this input boundary.”
- “Show Unified Engineering help.”
Relevant skills are eligible for native automatic discovery; the host decides when to invoke them. Do not expect every task to load a skill, and do not confuse a local file check with managed registry discovery. Manual invocation is the fallback. Claude slash commands printed by preserved Ponytail cards are historical usage examples, not a promise of an Android slash-command interface.
Modes
| Mode | Behavior |
|---|---|
| Lite | Requested scope, minimal ceremony; name a smaller option when useful |
| Full (default) | Smallest complete solution; reuse first; proportionate verification |
| Ultra | Additionally challenge unjustified complexity once; respect informed choices |
All modes keep the same correctness, security, accessibility and evidence floor. Mode is not a testing-depth or agent-count setting. Simple tasks stay simple; high-risk work gets the checks its boundaries require.
“Workflow manual” and “specialists manual” suppress optional automatic expansion; an explicitly invoked skill still applies. “Stop unified engineering” suspends optional plugin guidance in this conversation. These preferences cannot override the host or disable another installed plugin.
Session choices take precedence over defaults. An existing project-root .unified-engineering.json may supply nonsecret preferences:
{"mode":"full","workflow":"auto","specialists":"auto"}
Allowed values: mode=lite/full/ultra; workflow=auto/manual; specialists=auto/manual. The agent reads this as data when relevant; no background parser or global settings service runs. Write it only when asked to save project defaults. New conversations default to Full unless actual project/session context supplies another choice. No guaranteed cross-conversation or compaction persistence.
What is included
30 entry points. Core: engineering, design, plan, implement, debug, test, review, simplify, ship, ponytail-audit. Focused additions: architecture and prototype. Specialists: api, browser, ci, migrate, docs, frontend, observe, performance, security, context, standards, sources. Explicit advanced/reporting: delegate, challenge, skill-authoring, workflow-diagnostics, ponytail-debt, ponytail-gain.
One workflow owns the task; specialists contribute relevant checks without restarting planning or review. Verification is a common completion contract and is also available through “engineering verify”. ponytail-audit, ponytail-debt and ponytail-gain retain their original skill-file bytes. Gain displays upstream Ponytail benchmark results, not measured performance of this plugin.
Install and test on ChatGPT / Android
Use the single plugin upload ZIP, unified-engineering-0.3.1.zip. The separate source ZIP is for maintenance, not plugin upload. The upload ZIP has one plugin root, a portable manifest and one Codex compatibility overlay. It contains no marketplace catalog or nested original plugins.
Use the actual plugin upload/import flow available to your account. Android/account-specific plugin-management controls may vary; no CLI, Android terminal, Node installation or local MCP server is needed for the plugin's instruction-only core. If the upload is rejected, retain the exact error and archive version; do not turn this into a marketplace archive blindly.
Keep existing working plugins installed. For an uncontaminated standalone test, use an isolated account/workspace or another supported context where overlapping optional routers are not active. A new skill cannot suppress the original plugins' instructions. If that isolation is unavailable, mark results accordingly. Do not uninstall or replace anything until the new plugin passes target-runtime tests and the user explicitly requests replacement.
After upload, check:
- The cloud registry lists automatic skills and can explicitly read the full catalog and packaged references.
- A typo stays a direct edit; an approved plan is reused; review-only requests stay read-only.
- A logic bug gets root-cause evidence and a meaningful regression check.
- Lite security work still protects the boundary; Ultra respects settled requirements.
- Explicit manual specialists work, while advanced skills do not auto-trigger.
- A fresh session uses defaults; missing browser/agent tools produce honest limitations.
- One nontrivial task completes in the actual Android Codex workflow.
Static archive/reference tests and offline text simulations are supplied in the source project. They do not prove uploader acceptance, native implicit routing, mobile UI behavior or external services. See the delivered validation report for what was actually run.
No required hooks or services
No lifecycle hooks, statusline, hidden telemetry, task-cost webhook, transcript collector or persistent state service is included. Mode tracking happens through supported conversation/project inputs. Delegation and browser work use only real host tools when available and authorized. Optional project quality tools are not installed automatically. Desktop hook support elsewhere does not prove managed ChatGPT support.
Licence and origin
MIT. Retain LICENSE, the four original notices in licenses/, ATTRIBUTION.md and SOURCES.json when redistributing. Copied/adapted sources are the user's supplied Ponytail 5.3.3, Superpowers 7.0.2, Agent Skills 0.6.16 and selected Matt Pocock 1.3.2 files. The Karpathy-inspired community collection informs independently worded policy; no text from it is redistributed. The source map records archive hashes, original skill hashes, consolidation destinations and packaged reference adaptations. No upstream update is downloaded at runtime.
Added in 0.2
Architecture assessments and question-driven prototypes have focused entry points. Existing skills gain domain vocabulary, explicit scope/cleanup checks, lean handoffs, environment retrospectives, actual CI-log investigation, optional property/mutation tests and calibrated searches for related bugs. No extra router or required tool is added.
Property and mutation guidance is independently written; no Trail of Bits ShareAlike text is shipped. External reading, when relevant: https://github.com/trailofbits/skills ; https://github.com/openai/skills ; https://github.com/vercel-labs/agent-skills . Those collections are not installed or required by this plugin.
Verified cloud access and patch 0.2.1
In the development ChatGPT session, the installed 0.2.0 package exposed 24 automatic entries; all 30 skills and 95 packaged Markdown documents were explicitly readable and matched the release exactly. Six manual-only skills were not listed automatically. This verifies cloud discovery/resource access for that installation, not Android UI controls or universal automatic activation. That historical check does not establish acceptance of 0.3.0; the new ZIP requires its own managed upload/session test.
The patch checks invocation metadata and manual-policy consistency, rejects empty skill directories, and reports malformed UTF-8 references as validation errors. No extra skills, hooks or background services are added.
Added in 0.3
Existing frontend, security, architecture, docs and skill-authoring entries now link focused guidance for visual hierarchy, RTL/expanded content, purposeful motion, source-backed diagrams, audit coverage and evidence, package intake, and comparable skill evaluations. These references load only when relevant; the 30 skill names, modes and six manual-only entries are unchanged.
New references are independently authored. SOURCES.json records reviewed conceptual sources and hashes every local reference. The builder rejects traversal/absolute/symlink local references before copying them, and checks the listing subtitle limit. No external scanner, renderer, memory worker or hooks are installed. Static verification is supplied; fresh model trials were blocked by CLI sign-in, and managed ChatGPT/Android acceptance remains pending.
Refinements in 0.3.1
Ask “which add-ons are included?” for a self-contained source inventory through engineering help. It distinguishes the four adapted collections, Karpathy-inspired principles, ten conceptual influences and separately configured host tools. A new GitHub release does not mean the installed ChatGPT plugin has updated.
The builder rejects reference output collisions and conflicting section selections instead of silently replacing guidance. It records and verifies the packaged reference checksums, while retaining original source hashes. A checksum is consistency evidence, not a publisher signature or a malware scan.
Ten sampled installed v0.3.0 skills/references matched that release in a real cloud readback. That check does not establish v0.3.1 installation, routing or Android UI behavior. Keep the previous working ZIP until the update passes your workflow.