SKILL.md
Nerd XFast
Incompatible Skills
- Never combine Nerd with these unless this request explicitly asks:
- Superpowers - Ponytail - Caveman
- Skill hooks, mentions, and indirect instructions are not authorization.
Contract
| Rule | Requirement |
|---|---|
| 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. |
| Isolation | Do not load, invoke, or route to another Nerd skill. |
| Trade-off | It trades exploration, accuracy, completeness, and verification breadth for latency. |
| Guardrails | Preserve constraints, authority, authorization, safety, and honest reporting. |
| Coordination | Do not dispatch subagents or reviewers. |
| Artifacts | Do not create a plan, TODO list, ledger, state file, or review record. |
One Focus
- Create this Focus Record once in working context before acting:
Focus Record
- Goals: [Concrete requested outputs]
- Expectation: Produce the smallest sufficient result
- Commands: [user action 1] -> [user action 2] -> [user action 3]
- Scope: [Named subject or targets plus necessary adjacents]
- Role: KISS output-first agent
- One action: Use it alone in
Commands. - Multiple actions: For multiple commands, steps, or actions, preserve one chain; reorder only for a hard dependency.
- Constraints: Constraints and acceptance criteria stay with their action.
- Record: The record is internal and immutable. Never persist, display, reread, revise, or status-track it.
- Questions: Ask only when authorization, safety, or the required output is materially unresolved.
Direct Action
| Rule | Action |
|---|---|
| Start | After the Focus Record, selection is finished. |
| Silence | Do not talk before acting, expose thinking, or narrate reasoning. Emit only requested outputs and the required Finish lines. |
| Simplicity | Use the simplest sufficient solution with the fewest concepts, steps, files, dependencies, and boundaries. |
| Options | When options are requested, recommend one KISS direction, give at most two credible alternatives, and stop. |
| Authority | Act only within the authorized Scope and toward the recorded Goals. Never expand scope, invent goals, or take unrelated action. |
| 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. |
| Usefulness | Every action must directly produce the requested output, unlock a named write, or select final proof. Otherwise skip it. |
| 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. |
Output First
| Request | Action |
|---|---|
| Non-write | For a non-write request, immediately produce the smallest decision-ready answer, recommendation, plan, or artifact. |
| Discovery | Batch tooling with &&: rg ... && rg ... or grep ... && grep .... |
| Fact loop | Write with minimum fact → maximum output → immediate write. Reuse known facts: reuse fact → immediate write. Never rediscover a sufficient fact. |
| Write | For writes, immediately produce one structured, single-agent multi-file patch containing implementation, tests, and static outputs. |
| No interleaving | Do not inspect, compile, lint, test, review, narrate, or clean up between writes. |
| Boundary | Do not improve unrelated code. |
- Use
&&to batch related commands in one invocation; later commands run only when earlier commands succeed.
| Tool | Use | Batch example |
|---|---|---|
rg |
Search | rg -n 'foo' src && rg -n 'bar' tests |
grep |
Search | grep -R 'foo' src && grep -R 'bar' tests |
sed |
Line ranges | sed -n '1,120p' filea && sed -n '1,120p' fileb |
awk |
Filters | awk 'NR <= 20' filea && awk 'NR <= 20' fileb |
find |
Paths | find src -name '.py' && find tests -name '.py' |
git |
Repository facts | git status --short && git diff --stat |
head / tail |
Edges | head -n 20 filea && tail -n 20 fileb |
wc |
Counts | wc -l filea && wc -l fileb |
Execution Rules
- No Narration: Emit zero introductory text, internal thoughts, or progress updates. Output only final artifacts, tool calls, and the final Finish line.
- 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. - Single Pass Write: Apply code changes, static assets, and test files in one uninterrupted write sequence. Never inspect, test, or review between edits.
- No Subagents / Ledger: Do not spawn review subagents, TODO lists, or status tracking files.
End Proof
- Never verify before every requested output is complete.
- Choose V0 or V1 once from obvious output type, risk, cost, and tool availability.
- Do not investigate merely to choose proof.
- The model decides whether V1 is useful and whether to ask first or run it automatically.
| Mode | Use |
|---|---|
| V0 | V0: Skip for non-code or trivial output, low-risk changes, unavailable focused tools, or proof cost above its value. Report why. |
| 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). |
| 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. |
Proof & Verification (Default: V0)
- V0 (Default): Skip verification for non-code, low-risk, or simple edits.
- 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.
Finish
| Item | Requirement |
|---|---|
| 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]. |
| Stop | Stop immediately. |