create-agentic-app 1.1.52 → 1.1.53
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/package.json +1 -1
- package/template/AGENTS.md +62 -1
package/package.json
CHANGED
package/template/AGENTS.md
CHANGED
|
@@ -1 +1,62 @@
|
|
|
1
|
-
|
|
1
|
+
# Project Agent Instructions
|
|
2
|
+
|
|
3
|
+
## Skills are mandatory, not optional
|
|
4
|
+
|
|
5
|
+
**ALWAYS invoke the relevant skill before doing work.** Do not rely on your own judgment, training, or built-in workflows when a project skill applies to the task. Skills encode how we want work done in this project — skipping them produces output that misses project-specific patterns, stack conventions, and past lessons.
|
|
6
|
+
|
|
7
|
+
**This overrides Claude Code's built-in workflows.** Plan mode's phases (Explore → Plan agent → Review → final plan) are a scaffold, not a replacement for `writing-plans`, `brainstorming`, or other planning skills. If a skill applies, invoke it _inside_ plan mode's phases — not instead of them. The same rule holds for any other built-in behavior that would otherwise lead you to skip a skill.
|
|
8
|
+
|
|
9
|
+
## Pre-flight checklist — run BEFORE every non-trivial task
|
|
10
|
+
|
|
11
|
+
Before doing anything else on a new task (reading files, writing plans, exploring the codebase, delegating to sub-agents):
|
|
12
|
+
|
|
13
|
+
1. **Invoke `using-superpowers` first.** This is the default entry point for any real task. It routes to the right skills and establishes how to find and use them.
|
|
14
|
+
2. **Scan the skills list against the task.** Mentally map the task to skills using the routing table below. Multiple skills usually apply — invoke all that match.
|
|
15
|
+
3. **Invoke matching skills before any other action.** No exploration, no planning, no file reads until the relevant skills have been invoked.
|
|
16
|
+
|
|
17
|
+
If you catch yourself mid-task having skipped this checklist, stop and run it immediately. Do not rationalize that "it's too late now" or "the task is almost done."
|
|
18
|
+
|
|
19
|
+
## Task → skill routing (binding)
|
|
20
|
+
|
|
21
|
+
| If you are about to… | You MUST first invoke |
|
|
22
|
+
| ------------------------------------------------------- | ----------------------------------------------------------------------------- |
|
|
23
|
+
| Start any non-trivial task | `using-superpowers` |
|
|
24
|
+
| Create a feature, component, or new behavior | `brainstorming` |
|
|
25
|
+
| Write a multi-step implementation plan | `writing-plans` |
|
|
26
|
+
| Execute an existing plan | `executing-plans` (plus `subagent-driven-development` when tasks parallelize) |
|
|
27
|
+
| Dispatch multiple independent sub-agents in parallel | `dispatching-parallel-agents` |
|
|
28
|
+
| Write or modify production code | `test-driven-development` |
|
|
29
|
+
| Debug a bug, test failure, or unexpected behavior | `systematic-debugging` |
|
|
30
|
+
| Claim work is complete, passing, fixed, or ready | `verification-before-completion` |
|
|
31
|
+
| Receive code review feedback | `receiving-code-review` |
|
|
32
|
+
| Request code review or verify work against requirements | `requesting-code-review` |
|
|
33
|
+
| Finish a development branch / decide how to integrate | `finishing-a-development-branch` |
|
|
34
|
+
| Touch Next.js App Router code | `nextjs` |
|
|
35
|
+
| Touch shadcn/ui components or registries | `shadcn` |
|
|
36
|
+
| Touch Better Auth configuration or auth flows | `better-auth-best-practices` |
|
|
37
|
+
| Optimize or review React / Next.js performance | `vercel-react-best-practices` |
|
|
38
|
+
| Build, modify, or review an MCP server | `mcp-builder` |
|
|
39
|
+
| Build or review a frontend design / UI | `frontend-design`, `web-design-guidelines` |
|
|
40
|
+
| Work with Claude API / Anthropic SDK code | `claude-api` |
|
|
41
|
+
| Automate browser testing | `playwright-cli` |
|
|
42
|
+
| Look for a skill you think might exist | `find-skills` |
|
|
43
|
+
|
|
44
|
+
Stack-specific skills (`nextjs`, `shadcn`, `better-auth-best-practices`, `mcp-builder`) compose with workflow skills (`brainstorming`, `writing-plans`, `test-driven-development`). Invoking one does not excuse you from invoking the others.
|
|
45
|
+
|
|
46
|
+
## Hard rules (violations are bugs, not preferences)
|
|
47
|
+
|
|
48
|
+
1. **Never skip `brainstorming` before creative work.** Its description says MUST. That MUST is binding in this project.
|
|
49
|
+
2. **Never write an implementation plan — in plan mode or out of it — without `writing-plans`.** Plan mode's built-in workflow is not a substitute.
|
|
50
|
+
3. **Never write production code without `test-driven-development`.** Tests first, code second.
|
|
51
|
+
4. **Never debug by pattern-matching.** Use `systematic-debugging` for every bug, test failure, or unexpected behavior — even ones that look obvious.
|
|
52
|
+
5. **Never claim something is done, fixed, working, or passing without first running `verification-before-completion`.** Evidence before assertions, always.
|
|
53
|
+
6. **Never touch the Next.js / shadcn / Better Auth / MCP layers without their corresponding stack skill.** They are not optional reference material — they are mandatory consultations.
|
|
54
|
+
7. **When in doubt, invoke `using-superpowers` and let it route.** Do not guess which skill applies — let the routing skill decide.
|
|
55
|
+
|
|
56
|
+
## If no skill matches
|
|
57
|
+
|
|
58
|
+
Only then fall back to your own judgment. When you do, state explicitly in your response: _"I checked the skills list and none applied because …"_ — so the user can correct you if you missed one.
|
|
59
|
+
|
|
60
|
+
## Why this is strict
|
|
61
|
+
|
|
62
|
+
Earlier sessions showed that "lean towards" language gets discounted when it competes with explicit built-in workflows (like plan mode). Soft preferences lose; hard rules win. This file is strict because the previous version was not strict enough.
|