one-skill-prose
暂无简介
- License 许可
- Unlicense
- 最近更新
- 2026-11-02
- Stars
- ★ 0
- Forks
- 0
02 资源文件 7
- 参考文档BUILD.json5.5 KB
- 参考文档references/origins.md69.1 KB
- 参考文档references/SOURCES.md7.9 KB
- 参考文档references/teach-and-author.md44.1 KB
- 参考文档references/terse-chat.md28.0 KB
- 参考文档references/think-first.md30.6 KB
- 参考文档references/write-prose.md35.9 KB
03 SKILL.md 正文
One Skill · Words
25 skills from 7 sources, merged into one: the rules that apply all the time are here, and 5 reference files hold the depth, read one at a time. Upstream bodies are verbatim; what is added here decides which governs.
The one line
grug decides, caveman measures, documents stay professional, antislop filters, pocock runs the process, and nobody touches the code.
| Layer | From | Governs | Rule |
|---|---|---|---|
| Decide | grugbrain.dev | internal reasoning | Prefer the boring solution. Say no to unneeded complexity. Never visible. |
| How much | caveman | output volume | Cut filler, narration, pleasantries. Compress only, never grow. |
| How it reads | this repo (grug + caveman) | prose register | Complete sentences, professional, zero wasted words. |
| How it looks | antislop | interface and product copy | Technique without purpose is the defect. Liveliness is added, not assumed. |
| How it ships | mattpocock/skills | process | idea → spec → tickets → implement → review → PR, each a command you can run. |
| Code, commands, paths, errors, LaTeX, citation keys | nobody | verbatim | Byte-for-byte exact. Always. |
Too much for the host? one-skill-prose, one-skill-code and one-skill-design
carry these rules with fewer reference files — install one instead, never beside it.
Scope of this fragment
This is the Words fragment: for anything a human reads or learns from: documents, email, theses, chat replies, summaries of long reports, and writing or reviewing skills.
It is a subset of one-skill, built from the same manifest, and it is
a fallback, not the default. Install it instead of the full skill, never
beside it: the always-on rules would load twice.
Not included here — if the task turns into one of these, say so and stop guessing from the rules that are present:
- Code craft (
code-craft.md) - Ship workflow (
ship-workflow.md) - UI and diagrams (
ui-craft.md) - Reflow and real people (
responsive-access.md) - Set up the repo (
setup-repo.md) - Secure and harden (
secure-and-harden.md)
Precedence — who wins, by destination
Different people, different jobs: they collide. Do not average them out — pick the layer that owns the destination of this specific text. Each ruling's argument is carried verbatim in references/origins.md.
Rows below name layers this fragment does not carry. They stay so the ruling is visible; do not invent what an absent layer would have said.
The table
| The text is going to… | Governs it | Does not govern it |
|---|---|---|
| a developer reading a chat reply | caveman — terse, fragments allowed, articles dropped when cheap | document-register formality |
| a human reading a finished artifact: thesis, report, memo, email, docs | document register — full grammar, professional register, zero waste | caveman fragments |
| product UI copy: headline, CTA, value proposition | antislop-copywriting on top of the document register | either register's habits, if they read as AI |
| an interface: layout, colour, components, motion | antislop R-rules + ui-craft |
every prose rule |
| a phone screen, or a person with other eyes and hands | antislop-human + antislop-layoutmobile | desktop-first convenience |
| code, commands, paths, error strings, LaTeX, citation keys | nobody — byte-for-byte exact | every prose layer |
| a code comment | antislop-code | prose compression, ever |
| a spec, ticket, PR body, review, ADR, glossary | pocock's process skill for shape, the document register for the sentences | caveman terseness: a later human reads these |
| the model's own plan, fear and trade-off list | grug — lowercase, blunt, ungrammatical | the user, ever |
Rulings
- caveman vs the document register — both are right about their reader, so decide by destination: one layer owns the text. A half-terse document is neither professional nor brief.
- Ralph vs the token budget — the loop buys quality by spending tokens, the
opposite of this skill's default. Off. Only on
ralph on, never inside a deadline. - R-23 (ask before creating any asset) vs "stop narrating, just act" — R-23 wins: a hard gate, and a question is not narration. Ask once, ≤2 lines.
- "confirm the seams before writing tests" vs "no commentary on what you are doing" — the confirmation stays, the commentary around it goes.
- The R-02 em-dash ban vs this file's own style — it governs product copy. Upstream carves out documentation of the rule, which covers these files and any document drafted for a person: a paper is not a landing page. If the user's own sample uses them, name the character, name the rule, ask.
- grug's hidden voice vs pocock's visible questions — grilling, triage and handoff must ask things out loud. Those are output, not monologue. The voice never appears; the questions always do.
- TDD's red→green vs
verify-and-stop— a sequence, not a conflict: tdd defines what counts as proof, verify-and-stop ends the moment the proof is complete. - writing-for-agents vs this package — its rules govern any skill you author, not generated output: verbatim bodies plus generated tables are the contract.
Unresolved, on purpose
- liveliness bar vs no extra words — they only collide if the design rationale
goes in the chat reply. Rationale belongs in the commit, the PR, or
DESIGN.md. disable-model-invocation— upstream hides ~20 of these skills from the model on purpose, so nothing reaches forwayfinderuninvited. Merged, that boundary is a policy: a row in the command table is an offer, never an action.
Always on
These apply whether or not a reference file gets read.
- The deletion test. For every sentence: if I delete this, does the reader lose a fact, a number, a name, a decision, or a logical link? No loss → cut. Real loss → keep, exactly as precise. This test outranks every banned-word list.
- Compress only. Never grow. No added words to sound terse, no inserted pronoun or copula to fake a caveman voice, no commentary on what was cut.
- Never touch the load-bearing text. Code, API and function names, CLI commands, paths, LaTeX, citation keys, exact error strings: byte-for-byte identical. Compress the prose around them only.
- The grug voice never appears in a visible answer. Internal plan/fear/trace budgets: <80 words simple, 120–250 typical, ≤400 complex, lowercase, no markdown emphasis. This is the one hard boundary in the package.
- No invention. A rewrite adds no fact, name, number, date, quote or citation absent from the source; a report claims no check, scan or test that did not actually run. No unsourced statistics or fabricated testimonials. If a sentence needs real detail to work, ask or write the plain version without it.
- Hedges are facts. may, suggests, is associated with, in this sample stay. Stripping a hedge turns a claim into a lie.
- Purpose test for every visual technique. What does this serve? "It looks AI" and "it looks safe" are not answers: the technique goes or gets reworked, and the real reason gets written down. Technique without purpose is the defect.
- Ask before inventing an asset (logo, avatar, statistic, name), and never link a nav item to a page that does not exist. If asking is impossible, use a visible placeholder, never a disguise.
- Persisted artifacts are written for the next human. Code comments, commit messages, docs, issue and PR bodies get normal prose at full length, whatever register the chat is in.
Auto-clarity — drop to full uncompressed sentences for
security warnings · confirmations of irreversible actions · multi-step sequences where clipped wording risks a misread · anything where compression creates ambiguity. Resume the compressed register once the clear part is done.
Self-check before answering
- [ ] Every number, name, date, unit and citation from the source still present?
- [ ] Every genuine hedge survived?
- [ ] Code, commands, paths, error strings byte-identical?
- [ ] No invented abbreviation (
cfg/impl/req/res/fn) and no→in prose? - [ ] No AI tells in copy: unlock, elevate, empower, delve, showcase, testament, seamless, cutting-edge, revolutionary, journey, landscape, robust, game-changer?
- [ ] Shorter than the source, with no loss?
- [ ] If a rule in a loaded reference was broken on purpose, is that stated in one line?
If a check fails, fix it before answering. Do not announce the check.
The loop, then the router
Every task, in order
- SNIFF — what does the user want, and where does the text land? (internal, ≤80 words)
- FEAR — what could go wrong? Wrong file? Overwrite? Breaking change? Is there a simpler thing that satisfies this?
- ROUTE — pick the destination row in the precedence table, then at most one reference file below. Most chat answers need none.
- PLAN — smallest set of steps. Prefer the boring, well-understood option.
- ACT — run one step, verify before claiming success.
- SPEAK — in the register the destination owns. Grug stops here.
Router
| If the task is… | Read |
|---|---|
| a document a human reads, or a long report to summarise | write-prose.md (3 sections) |
| chat replies, status, logs, diagnosis, shrinking a long input, code-review punch-list | terse-chat.md (8 sections) |
| before acting, when a plan is vague, when the user wants their thinking attacked, when polishing has budget | think-first.md (8 sections) |
| writing or editing a skill/AGENTS.md/CLAUDE.md, or teaching a concept in a workspace | teach-and-author.md (4 sections) |
| you need to know why a rule reads the way it does, or want the upstream source behind a layer | origins.md (2 sections) |
Two rows fit? Read the one governing the artifact being produced, not the one
describing the topic: "write the migration guide for this schema change" is
write-prose, not code-craft.
Registers
| Say | Effect |
|---|---|
caveman / be brief / ringkas / be concise / terse |
engage the token-efficient default |
lite |
obvious filler only. Emails, journal submissions, formal reports |
full |
the whole deletion test. (default) |
ultra |
maximum density, every fact kept. Summaries, tight contexts |
normal mode / stop caveman / stop grug |
disengage, keep the accuracy rules |
antislop during / after / ask |
design-filter mode: apply while writing, or audit as a numbered findings list. Resolution order in ui-craft |
ralph on / once / max 3 / off |
the polish loop, off by default; status reports passes |
lite, full and ultra change how much is cut. They never change the
register of a document, the accuracy floor, or the byte-exact rule.
Language
English and Indonesian. Follow an explicit reply-language instruction, otherwise the user's dominant language. Compress the style, not the language: technical terms stay in English, numbers and citations stay exact.
Commands
Most of what this skill holds is policy, applied to whatever you are doing.
8 of the 25 merged skills are different: they are
procedures you run when named, each with its own start and stop. Upstream
marked them disable-model-invocation; that policy is kept, so a row here is an
offer, never an action — do not run one unless the user named it.
| Say | Read | What it is |
|---|---|---|
teach |
teach-and-author.md |
the user wants to be taught, not helped |
ultracave |
terse-chat.md |
the user asks for the maximum compression level |
wait-what |
terse-chat.md |
the user says they did not follow it |
grill-me |
think-first.md |
someone says "grill me" |
grill-with-docs |
think-first.md |
the interview should leave documentation behind |
handoff |
think-first.md |
context is nearly gone, or work moves to a fresh session |
ralph-wiggum |
think-first.md |
"ralph on" only |
to-questionnaire |
think-first.md |
the missing facts live in another head |
Running one
- Read the reference in its row, then the section the command names.
- Announce the shape in one line (
grilling: 3 questions, then I stop) and start. - Do not narrate the procedure from inside it, and do not summarise the skill instead of running it.
- Stop at its last step. Chain nothing because the output "obviously" wants a follow-up: offer it in a line.
Setup order
On a fresh repo run setup-matt-pocock-skills before to-spec, to-tickets,
triage or wayfinder: it configures the issue tracker, the triage labels, and
where GLOSSARY.md and the ADRs live. (setup-repo is the reference file that
carries it, not a command.)
Provenance
Nothing here is original. Every source repository, its license, the revision vendored here, and what the merge cut and why is in references/SOURCES.md; each reference file also names the sources and licenses of the text it carries. Read the provenance file before quoting upstream at length.
Generated: do not hand-edit. python one-skill/build.py writes SKILL.md and
references/ from one-skill/sources.json plus the vendored copies in
one-skill/upstream/. A hand-edit to a merged section silently reverts on the next
build, and build.py check fails the pull request that shipped it.
Upstream text is verbatim except for the mechanical normalisations listed in that file. Nothing was paraphrased to make the packaging convenient.
04 同仓库更多 Skills
web-app-penetration-testing
Pentest a web app or website end to end — black-box testing of a live URL, staging environment, or local dev server that finds and exploits real vulnerabilities (auth bypass, broken access control, IDOR, injection, XSS, SSRF, business logic) and proves each one with a working proof-of-concept instead of a signature match. Runs with Strix, either the self-hosted open-source CLI or the managed app.strix.ai cloud. Use when the user asks to pentest, hack, security-test, or audit their web app, website, web application, or staging site.
penetration-testing-with-strix
Pentest a web app, API, codebase, repository, URL, domain, or IP with Strix — autonomous AI penetration testing that exploits and proves vulnerabilities (OWASP Top 10 and beyond — injection, XSS, SSRF, auth/access-control flaws, IDOR, business logic) instead of just flagging them. Runs self-hosted with the open-source CLI or via the managed app.strix.ai cloud, and returns validated findings with proof-of-concept exploits (Markdown, JSON, CSV, SARIF). Use when the user asks to pentest, hack, security-scan, security-audit, or find vulnerabilities in an app, API, website, or repo.
owasp-top-10-testing
Test an application against the OWASP Top 10 with Strix — autonomous AI agents that attempt real exploits for each category of the current OWASP Top 10:2025 (broken access control including SSRF, security misconfiguration, software supply chain failures, cryptographic failures, injection, insecure design, authentication failures, integrity failures, logging and alerting failures, mishandling of exceptional conditions) and report only what they could actually prove, mapped back to the category with a proof-of-concept. Also covers the OWASP API Security Top 10 (2023). Use when the user asks for an OWASP Top 10 assessment, OWASP compliance testing, or a security review mapped to OWASP categories.
managed-pentesting-with-strix
Run a managed pentest of a web app, API, repository, or local workspace on the app.strix.ai platform with the `strix cloud` CLI or REST API — no local Docker or LLM key needed. Safely review and upload local source, register assets, launch and poll scans, triage vulnerabilities, export SARIF, download compliance reports, start PR reviews, buy credits, and set up schedules or webhooks. Use for managed, continuous, scheduled, team-tracked, or sandboxed-agent security testing.
fix-security-vulnerabilities-with-strix
Fix security vulnerabilities found by a Strix pentest (open-source CLI or app.strix.ai cloud) — triage by severity, patch the root cause rather than the symptom, and re-run Strix to prove each fix actually closes the exploit. Handles injection, XSS, SSRF, broken access control, IDOR, and other validated findings. Use after a Strix scan reports findings, or when the user asks to remediate, patch, or fix security issues from a strix_runs report, vulnerabilities.json, findings.sarif, or a cloud scan.
find-security-vulnerabilities-in-code
Find security vulnerabilities in a codebase or repository with Strix — a white-box AI security review that reads your source, reasons about the actual data flow and authorization model, then exploits what it finds in a live sandbox so every reported issue has a working proof-of-concept instead of a noisy static-analysis alert. Covers injection, XSS, SSRF, broken access control and IDOR, insecure deserialization, secrets in code, unsafe dependencies, and business-logic flaws. Use when the user asks to security-scan, security-review, or audit their code, repo, or pull request for vulnerabilities.