agentleFS
playingpossummplayingpossumm/loose-ends/.claude/skills/ingest/SKILL.md

ingest

Skill0 starsChanged 34 days agoMITView on GitHub

Compile one uncompiled source from raw/inbox into the wiki and loops. Extracts claims, writes and updates pages, links them, extracts open loops, flags contradictions, rebuilds index.md, and appends to log.md. Use when the user says "ingest", "compile", "process my inbox", or asks what is waiting to be compiled.

90 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use daily-arxiv, exp-eval and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • LambenthanSkill

    daily-arxiv

    每日从 arXiv 拉取新论文,过滤相关性,auto-ingest 高优先级论文,检测 SOTA 更新

    Also in 5 projects that use this

  • LambenthanSkill

    exp-eval

    实验判决门:Review LLM 独立评判实验结果 → 4 种判决路径 → 自动更新 claims confidence、ideas status、graph edges

    Also in 5 projects that use this

  • phanghonghaoSkill

    poster

    HTML/CSS A4 海报生成。从沙龙/活动 Markdown 文件生成科技风 A4 海报 PDF/PNG。Use when user mentions: poster, 海报, 海报生成, make poster, generate poster, 活动海报, 沙龙海报.

    Also in 5 projects that use this

  • skyllwtSkill

    exp-pilot-eval

    Pilot result evaluation — read pilot results, apply success criteria, update idea page (pilot_result, failure_reason if failed), generate PILOT_VERDICT_REPORT. Called by /ideate Phase 5 after /exp-pilot-run.

    Also in 4 projects that use this

  • skyllwtSkill

    exp-pilot-run

    Pilot experiment execution — read Pilot Spec YAML, write pilot code, run experiment(Confirm with the user before operation and require the applicant to conduct manual inspection), return results. Called by /ideate Phase 5. Does NOT modify wiki pages or judge pass/fail.

    Also in 4 projects that use this

  • +6 more from Lambenthan
  • +1 more from skyllwt

The file

Show its rules and text

We split this skill into the rules it gives an agent

15 rules. 0 are common, 8 are distinctive, and 7 are very specific to this project.

  • 0

    Common

    Most of the 74 edited versions we read keep these.

  • 8

    Distinctive

    This author's own choices. Worth a look.

  • 7

    Very specific

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

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

    When run unattended, apply the triage in ../ingest-all/SKILL.md: a source whose plan touches only wiki/ may be written without asking.

    line 12
  2. Distinctive:

    Read CLAUDE.md first if you have not this session.

    line 19
  3. Distinctive:

    Check the frontmatter before you start:

    line 29
  4. Very specific:

    3. Orient before writing: Read index.md, then every existing page plausibly related.

    line 38
  5. Distinctive:

    4. Extract, and show your work: Before writing anything, list for the user:

    line 42
  6. Distinctive:

    Stop here if the plan touches more than 15 pages.

    line 49
  7. Distinctive:

    6. Contradictions: Never overwrite. Keep both claims with dates and sources, mark the section, and log it.

    line 68
  8. Very specific:

    Write each to loops/open/<id>.md with status: open, surfaced: 0, a citation back to where it came from, and due: if one is implied.

    line 80
  9. Very specific:

    Every loop also needs title:, the name of the thing in title case, which is what the brief and the nudge print.

    line 83
  10. Distinctive:

    Write it the way the brief is written, since it reaches the same reader in the same week: absolute dates but never the loop's own due date, …

    line 92
  11. Very specific:

    Add nudge: morning or nudge: evening only when the nature of the item disagrees with its date.

    line 97
  12. Very specific:

    Every page you create needs a one-line summary: in its frontmatter, because that is what the index prints.

    line 108
  13. Distinctive:

    Every claim cites a source and a locator: No exceptions, including claims from images, which are cited as the image.

    line 128
  14. Very specific:

    Never write mem/: If the source implies something about the user's goals, preferences, or projects, propose it at the end and let them …

    line 130
  15. Distinctive:

    If the source is low quality or you cannot verify what it claims, say so plainly.

    line 134

“Common” is measured: kept by at least 6 in 10 of the 74 edited versions of this file we could read. “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

ingest

Compile one source at a time. This is the irreversible step in the system, since everything else can be redone from raw/, but a bad compile spreads across many pages before anyone notices. Work carefully and stay inside the caps.

When run unattended, apply the triage in ../ingest-all/SKILL.md: a source whose plan touches only wiki/ may be written without asking; anything touching a date, a loop, mem/, or an existing claim waits for the reader. When run with a person present, show the plan and ask as before, because the triage exists to make unattended compiling safe, not to stop you approving things yourself.

Read CLAUDE.md first if you have not this session. Its rules override anything here.

Procedure

1. Pick the source. Named by the user, or the oldest status: uncompiled file in raw/inbox/. If several are waiting, say how many and compile one.

2. Read it fully, including anything it points at.

Check the frontmatter before you start:

FieldWhat it means for you
attachment:The markdown is only a stub. Open the file it names, because that is the real source. A Telegram-captured PDF or photo has its text in the attachment, not the page.
origin: telegram (forwarded from X)X sent this, the user relayed it. Attribute claims to X, not to the user. This distinction matters and is easy to lose.
kind: imageView the image itself, not only any transcription beside it.
sent: vs captured:When they said it vs when it was filed. Use sent: for anything time-sensitive, since a forwarded message can arrive days late.

3. Orient before writing. Read index.md, then every existing page plausibly related. You cannot flag a contradiction against a page you did not read, so this step is what makes step 6 work, and skipping it is the most common way this system silently degrades.

4. Extract, and show your work. Before writing anything, list for the user:

  • key claims, each with its locator in the source
  • entities mentioned (people, tools, papers, projects, concepts)
  • open loops, see below
  • contradictions with existing pages
  • which pages you intend to create or update

Stop here if the plan touches more than 15 pages. Ask.

5. Write.

  • wiki/sources/<id>.md, one page for this source, covering what it is, what it claims and why it matters.
  • wiki/entities/ and wiki/concepts/, created or updated. Run python scripts/synthesis.py and promote whatever it lists. You compile one source at a time and cannot see mention counts across the other fifty; the script counts them for you, from names already in mem/ and from the category: you write on every source. Promote at 3 independent sources, or whenever the user asks.
    • An entity page answers "what does the vault know about this person, project or tool", with every claim cited to the source it came from.
    • A concept page keeps ## What sources say and ## Current view separate, and is the only place a contradiction between two sources becomes useful rather than merely recorded.
  • On every page touched, add this source's id to sources: and bump updated:. This is what makes /unsource possible later. Do not skip it.
  • Concept pages keep ## What sources say and ## Current view separate. Never merge them.

6. Contradictions. Never overwrite. Keep both claims with dates and sources, mark the section, and log it. Resolution belongs to the human.

7. Loops, the part that matters most. Scan for anything stated but unresolved:

SignalExample
stated intent"I want to learn X", "I should read this"
commitment"I'll send them the thing"
dated fact implying actiona birthday, a deadline, a renewal
a source captured for a purpose that never happeneda PDF filed under "to read"

Write each to loops/open/<id>.md with status: open, surfaced: 0, a citation back to where it came from, and due: if one is implied. Dated facts go to loops/dates/.

Every loop also needs title:, the name of the thing in title case, which is what the brief and the nudge print. The # heading may describe the loop; the title names it. "Foundation — start go-to-market planning (Monday)" is a heading, "Foundation GTM Planning" is a title.

And summary:, one or two sentences addressed to the user that stand alone in an email. The nudge prints it verbatim and prints nothing without it. Do not describe the loop ("the user wants to read this"); say the thing itself ("Richard Cook, 18 principles, about 20 minutes. Unread.").

Write it the way the brief is written, since it reaches the same reader in the same week: absolute dates but never the loop's own due date, which the nudge prints beside the title anyway; no em dashes; no closing clause explaining why the item matters. The frontmatter contract in CLAUDE.md has the full rule and the examples.

Add nudge: morning or nudge: evening only when the nature of the item disagrees with its date. Reading is an evening act whatever the deadline; anything needing an office to be open is a morning one.

Be conservative. A false loop costs the user trust in the brief, and brief precision is the metric this project lives or dies by. When unsure, list it for the user rather than filing it.

8. Rebuild index.md by running python scripts/build_index.py. Do not edit it by hand, since it is generated from the vault and the next run overwrites anything you write there.

Every page you create needs a one-line summary: in its frontmatter, because that is what the index prints. Write it for someone scanning ninety entries for the one they want: name the subject and what the page settles, not what kind of page it is.

9. Append to log.md, keeping the prefix format exactly so grep "^## \[" works:

## [YYYY-MM-DD] ingest | <title>

10. File the raw source from raw/inbox/ into the right raw/ subfolder and set status: compiled.

11. Report in five lines or fewer: what compiled, pages created, pages updated, loops opened, contradictions flagged.

Rules

  • One source per run. Batch compilation hides mistakes.
  • Max 15 pages touched. Over that, stop and ask.
  • Every claim cites a source and a locator. No exceptions, including claims from images, which are cited as the image.
  • Never write mem/. If the source implies something about the user's goals, preferences, or projects, propose it at the end and let them decide.
  • raw/ is immutable apart from the status: field.
  • Refuse credentials, keys, and health records; say you refused.
  • If the source is low quality or you cannot verify what it claims, say so plainly. A bad source compiled confidently is worse than one not compiled at all.

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.