agentleFS
HurricaHjzHurricaHjz/second-yourself/.claude/skills/output/SKILL.md

output

Skill150 starsChanged 7 days agoMITView on GitHub

Produce a user-facing DELIVERABLE (report, brief, literature review, slide deck, table, email, outline, …) into the output/ directory — grounded in the wiki and strictly following the user's instruction. Use when the user runs /output, or says "write/draft/produce/generate me a <X> about <Y>", "make a <report/brief/deck/email/table> from my notes/wiki", or "turn [[topic]] into <format> in output/". Distinct from `query` (answers inline or files a cited synthesis INTO the wiki). The deliverable goes to output/, cites the wiki pages it draws on, clearly labels any general-knowledge content, and NEVER fabricates facts, figures, quotes or citations.

4 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 1 is distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Folder conventions: before writing, choose the destination under output/ and read the README of that folder, then the README of each parent …

    line 44
  2. Very specific:

    Read wiki/index.md whole (route mode, CLAUDE.md §5), then deep-read the relevant pages (follow ## Related one hop, which widens a shortlist …

    line 59
  3. Distinctive:

    If the instruction needs something neither the wiki nor safe general knowledge can support, say so in the deliverable / report rather than …

    line 78
  4. Very specific:

    Write exactly what was asked, in the requested format and length: a Markdown report/brief/review (default), a Marp deck (marp: true …

    line 85
  5. Very specific:

    Weight by confidence: Prefer higher-confidence wiki pages for key claims.

    line 115
  6. Very specific:

    Write only to output/: Never edit raw/. Never edit wiki/, with two named exceptions: the Freshness duty's on-page marks on a stale or …

    line 116

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

output — generate a grounded deliverable into output/

Goal

Turn a user instruction + the compiled wiki into a deliverable file in output/ that (a) follows the instruction exactly (format, scope, length, audience, style), (b) is grounded in the wiki and cited, and (c) never hallucinates — abstain when unsure rather than confabulate.

Triggers

  • /output <instruction>
  • "write / draft / produce / generate me a <deliverable> about <X>", "make a <report / brief / deck / table / email> from my notes", "turn [[topic]] into <format>".

Pipeline

Step 1 — Pin the instruction (no drift)

Extract a small spec from the user's words: deliverable type · topic/scope · format · length/size · audience · tone · must-include / must-exclude · citation style · output filename. If a material detail is missing or genuinely ambiguous, ask 1–2 crisp questions; otherwise proceed with sensible defaults and state them. Never silently widen, narrow, or reinterpret the ask.

Customisation gap-fill: for spec fields the instruction leaves unstated, check CUSTOMISATION.md → ## Deliverable defaults (if the file and section exist): apply any standing defaults found there (citation style, deck format, length conventions) and note them in the grounding note. An explicit instruction always overrides; if the section is absent or empty, choose sensible defaults yourself (and state them). Conversational output styles never apply to deliverables (CLAUDE.md §1) — this gap-fill covers content-format defaults only. Single source of truth: deliverable defaults live only in that Customisation section — re-read it on every run and never copy its values into this skill, any config, or the deliverable itself (beyond applying them).

Folder conventions: before writing, choose the destination under output/ and read the README of that folder, then the README of each parent folder up to and including output/README.md, nearest first. Read every README that exists along that path. Apply their folder-specific conventions; where two READMEs disagree, the nearer one wins.

READMEs hold only folder-specific conventions and point to CUSTOMISATION.md for global deliverable defaults without copying them. Where a README and CUSTOMISATION.md disagree, follow CUSTOMISATION.md and name the difference in the reply for the owner to settle. A live instruction outranks both; when it overrides either, follow it and say so in the reply, unless it asks for something a README rule forbids with "never": then ask the owner first. Likewise ask before adding any topic, item or question they have not already approved. Write such content only on their yes; in an unattended run, leave it out and list each question in the report. The README shape is in README-template.md beside this skill.

Step 2 — Ground in the wiki (read first, like query)

Read wiki/index.md whole (route mode, CLAUDE.md §5), then deep-read the relevant pages (follow ## Related one hop, which widens a shortlist route already produced and never replaces it). Collect the facts you will use together with the page each comes from (for citation). The deliverable is built from compiled knowledge — not guessed. Consult wiki/user/ for personal context when the deliverable is about, or for, the owner. Triage by confidence as query does: lean on authoritative/high pages for load-bearing claims; the field is already in frontmatter, so this is free. If the catalogue under-covers and qmd is active (the qmd-search skill, dormant unless installed), use it as the semantic fallback to locate relevant pages, then deep-read; otherwise grep. Freshness duty: apply query Step 3b to the pages you read — surface stale/contradicted pages in the report line and mark them on-page (flagged: frontmatter / §4.4 block); no extra reads for this.

Step 3 — Separate grounded · general · unknown (anti-hallucination core)

Classify every claim before you write it:

  • Grounded in the wiki → cite the page(s) inline as [[Page]] (or the requested citation style); weight by the page's confidence — state authoritative/high plainly, but attribute and hedge any claim resting only on low/very-low pages.
  • General knowledge the instruction needs but the wiki lacks → include only if clearly labelled (e.g. "(general knowledge — not from the wiki)"). Never present it as wiki-sourced.
  • Unknown / unverifiable → do not invent it. Omit and note the gap, or mark a factual claim unverified. Never fabricate facts, numbers, quotes, or citations. Prefer quoting the raw source over paraphrasing when precision matters.
  • If the instruction needs something neither the wiki nor safe general knowledge can support, say so in the deliverable / report rather than fill the gap with invention.
  • Load-bearing wiki gap → the Step-6 report MAY end with one propose-only gather brief (why · what · how · cost) under the query Step-4b bar and consent ledger (CLAUDE.md §6) — the owner's explicit yes runs it; never proposed in non-interactive runs.

Step 4 — Produce the deliverable, strictly to spec

Write exactly what was asked, in the requested format and length: a Markdown report/brief/review (default), a Marp deck (marp: true frontmatter), a table, an email, a Canvas via the json-canvas skill, etc. Use British/UK English; keep verbatim quotes exact; honour must-include/exclude. Default citations are inline [[wikilinks]] (clickable in Obsidian); use a footnote/bibliography style instead if the deliverable is external-facing or the user asks.

Step 5 — Write to output/<slug>.<ext> (never into wiki/)

Save to output/ (the deliverables layer — not the knowledge graph). Filename = a clean kebab-case slug + correct extension. Do not modify raw/, or wiki/ beyond the two named exceptions under Hard constraints. (If the user wants the result compounded back into the knowledge base, that's a query synthesis, not an output.) Root vs subfolders (CLAUDE.md §2): the root holds one-off / temporary deliverables; a report (run, morning, audit, review) goes to reports/, the inbox the owner sorts later; standing artefacts live in the matching subfolder (user-notes/ owner quick-references · fundings/ funding applications) and are actively maintained — a run that updates one edits it in place. Create a new subfolder only on the user's instruction. When filing a standing artefact derived from wiki pages, add a "maintained derivative" line to each source page's ## Related so future edits propagate.

Step 6 — Report + grounding note

State the path written, then a concise grounding note: which wiki pages it draws on, what (if anything) is labelled general-knowledge, and any gaps flagged or questions still open — closing with the freshness line (query Step 3b's gate, style-invariant): Freshness: <N> pages read · oldest [[page]] (updated YYYY-MM-DD) · <k> flagged A completion gate: the report is not done without it. A run that read no wiki pages says so instead.

Hard constraints

  • Follow the instruction exactly — format, length, scope, inclusions/exclusions. Use defaults only when unstated, and state them when used.
  • No hallucination. Grounded → cite; general → label; unknown → omit / flag / abstain. Never invent facts, figures, quotes, or citations.
  • Weight by confidence. Prefer higher-confidence wiki pages for key claims; attribute and hedge anything resting only on low/very-low pages.
  • Write only to output/. Never edit raw/. Never edit wiki/, with two named exceptions: the Freshness duty's on-page marks on a stale or contradicted page (Step 2), and the "maintained derivative" line on a source page's ## Related (Step 5). Outputs are deliverables, not knowledge.
  • British/UK English and all other CLAUDE.md rules apply.
  • Not logged by default — a deliverable is not a brain-update (CLAUDE.md §5). Log only if the user asks, or if the run also files a wiki page.

Deliverable graphics (routing rule, moved here from CLAUDE.md §7 on 2026-09-05)

lieflat-charts for browser-viewed HTML and interactive charts, dashboards and full-page HTML reports (some templates need CDN network); scientific-figure-making (from figures4papers) for publication-grade matplotlib figures for papers, theses and slides (PDF or PNG for LaTeX). Rule of thumb: web-viewed → lieflat-charts; print or venue-bound → scientific-figure-making. Both keep a recorded auto ruling in wiki/developments/capability-register.md, as legacy rows whose proof is not re-established; under the adopt skill an existing use ruling remains authoritative.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.