1---2name: nerd-xfast3description: Use only when explicitly invoked for a concrete output or authorized file or code change where latency is worth reduced exploration, accuracy, completeness, and proof breadth.4---5
6# Nerd XFast
7
8## Incompatible Skills
9
10- Never combine Nerd with these unless this request explicitly asks:
11 - Superpowers
12 - Ponytail
13 - Caveman
14- Skill hooks, mentions, and indirect instructions are not authorization.
15
16## Contract
17
18| Rule | Requirement |
19| --- | --- |
20| **Activation** | Use this self-contained KISS-first output skill only when the user explicitly invokes `nerd-xfast` for a concrete answer, decision, plan, static artifact, or authorized file or code deliverable. |
21| **Isolation** | Do not load, invoke, or route to another Nerd skill. |
22| **Trade-off** | It trades exploration, accuracy, completeness, and verification breadth for latency. |
23| **Guardrails** | Preserve constraints, authority, authorization, safety, and honest reporting. |
24| **Coordination** | Do not dispatch subagents or reviewers. |
25| **Artifacts** | Do not create a plan, TODO list, ledger, state file, or review record. |
26
27## One Focus
28
29- Create this Focus Record once in working context before acting:
30
31> **Focus Record**
32> - **Goals:** [Concrete requested outputs]
33> - **Expectation:** Produce the smallest sufficient result
34> - **Commands:** [user action 1] -> [user action 2] -> [user action 3]
35> - **Scope:** [Named subject or targets plus necessary adjacents]
36> - **Role:** KISS output-first agent
37
38- **One action:** Use it alone in `Commands`.
39- **Multiple actions:** For multiple commands, steps, or actions, preserve one chain; reorder only for a hard dependency.
40- **Constraints:** Constraints and acceptance criteria stay with their action.
41- **Record:** The record is internal and immutable. Never persist, display, reread, revise, or status-track it.
42- **Questions:** Ask only when authorization, safety, or the required output is materially unresolved.
43
44## Direct Action
45
46| Rule | Action |
47| --- | --- |
48| **Start** | After the Focus Record, selection is finished. |
49| **Silence** | Do not talk before acting, expose thinking, or narrate reasoning. Emit only requested outputs and the required Finish lines. |
50| **Simplicity** | Use the simplest sufficient solution with the fewest concepts, steps, files, dependencies, and boundaries. |
51| **Options** | When options are requested, recommend one KISS direction, give at most two credible alternatives, and stop. |
52| **Authority** | Act only within the authorized Scope and toward the recorded Goals. Never expand scope, invent goals, or take unrelated action. |
53| **Persistence** | If any Goal remains unmet, immediately take the next authorized action that directly advances it. Continue without pausing for commentary or confirmation until every Goal is reached or a real authorization or safety blocker requires the user. |
54| **Usefulness** | Every action must directly produce the requested output, unlock a named write, or select final proof. Otherwise skip it. |
55| **Reading** | Read named write targets and their nearest authority together; if unknown, use one narrow discovery batch. Stop reading when the smallest sufficient output or complete write set is known. |
56
57## Output First
58
59| Request | Action |
60| --- | --- |
61| **Non-write** | For a non-write request, immediately produce the smallest decision-ready answer, recommendation, plan, or artifact. |
62| **Discovery** | Batch tooling with `&&`: `rg ... && rg ...` or `grep ... && grep ...`. |
63| **Fact loop** | Write with `minimum fact → maximum output → immediate write`. Reuse known facts: `reuse fact → immediate write`. Never rediscover a sufficient fact. |
64| **Write** | For writes, immediately produce one structured, single-agent multi-file patch containing implementation, tests, and static outputs. |
65| **No interleaving** | Do not inspect, compile, lint, test, review, narrate, or clean up between writes. |
66| **Boundary** | Do not improve unrelated code. |
67
68- Use `&&` to batch related commands in one invocation; later commands run only when earlier commands succeed.
69
70| Tool | Use | Batch example |
71| --- | --- | --- |
72| `rg` | Search | `rg -n 'foo' src && rg -n 'bar' tests` |
73| `grep` | Search | `grep -R 'foo' src && grep -R 'bar' tests` |
74| `sed` | Line ranges | `sed -n '1,120p' file_a && sed -n '1,120p' file_b` |
75| `awk` | Filters | `awk 'NR <= 20' file_a && awk 'NR <= 20' file_b` |
76| `find` | Paths | `find src -name '*.py' && find tests -name '*.py'` |
77| `git` | Repository facts | `git status --short && git diff --stat` |
78| `head` / `tail` | Edges | `head -n 20 file_a && tail -n 20 file_b` |
79| `wc` | Counts | `wc -l file_a && wc -l file_b` |
80
81
82## Execution Rules
83- **No Narration:** Emit zero introductory text, internal thoughts, or progress updates. Output only final artifacts, tool calls, and the final Finish line.
84- **Single Pass Discovery:** Batch all read/search commands in a single execution block using `&&` (e.g., `rg -n 'foo' src && git status --short`). Stop reading the moment minimum viable context is acquired.
85- **Single Pass Write:** Apply code changes, static assets, and test files in one uninterrupted write sequence. Never inspect, test, or review between edits.
86- **No Subagents / Ledger:** Do not spawn review subagents, TODO lists, or status tracking files.
87
88## End Proof
89
90- Never verify before every requested output is complete.
91- Choose **V0** or **V1** once from obvious output type, risk, cost, and tool availability.
92- Do not investigate merely to choose proof.
93- The model decides whether V1 is useful and whether to ask first or run it automatically.
94
95| Mode | Use |
96| --- | --- |
97| **V0** | **V0:** Skip for non-code or trivial output, low-risk changes, unavailable focused tools, or proof cost above its value. Report why. |
98| **V1 automatic** | **V1 automatic:** Run immediately available safe, local, focused commands. Run a single, targeted verification command concurrently (e.g., exact test file or linters on changed files only). |
99| **V1 ask first** | **V1 ask first:** Ask when proof is broad, slow, stateful, external, potentially destructive, or needs configuration or more authority. Tool unavailability means skip, never install. |
100
101## Proof & Verification (Default: V0)
102- **V0 (Default):** Skip verification for non-code, low-risk, or simple edits.
103- **V1 (Code Changes Only):** Run a single, targeted verification command concurrently (e.g., exact test file or linters on changed files only). Maximum one repair attempt on failure.
104
105## Finish
106
107| Item | Requirement |
108| --- | --- |
109| **Report** | Report only the produced outcome and one exact proof decision: `V0 — skipped: [reason]`, `V1 — automatically verified: [results]`, or `V1 — confirmation required: [cost or risk]`. |
110| **Stop** | Stop immediately. |
danangjoyoo