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

ingest-all

Skill0 starsChanged 7 days agoMITView on GitHub

Compile every uncompiled source in raw/inbox in one pass, planning the whole batch first, approving once, then writing. Use when the user says "ingest all", "compile everything", "clear the inbox", or when several sources are waiting and compiling them one at a time is just friction.

The file

We split this skill into the rules it gives an agent

10 rules. About 0 are common, 6 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

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

    line 16
  2. Distinctive:

    Before planning anything, run:

    line 20
  3. Distinctive:

    Not every source carries the same risk, so not every source needs the same approval.

    line 45
  4. Very specific:

    Held sources stay in raw/inbox/: Do not move them, do not mark them, do not write a holding file.

    line 68
  5. Distinctive:

    Say what you did: Every source that wrote itself gets one line in the report, and every source held gets one line saying which row of the …

    line 73
  6. Very specific:

    1. List what is waiting: Every file in raw/inbox/ with status: uncompiled, oldest first.

    line 91
  7. Very specific:

    2. Read them all, and orient once: Read index.md and the existing pages that any of them plausibly touch.

    line 95
  8. Distinctive:

    4. Wait for approval: Do not write before it. One approval for the batch is the point.

    line 127
  9. Distinctive:

    6. Stop mid-batch and ask: if any of these happen, rather than pushing through:

    line 133
  10. Very specific:

    If two of eight actually matter, /ingest those two properly first.

    line 168

≈ “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

ingest-all

Same work as /ingest, done across the whole inbox with one approval instead of one per source. Everything in ../ingest/SKILL.md still applies, and the rules about citation, provenance, contradictions and loops are not relaxed because there are more files.

What changes is only where the human checkpoint sits: once, over the whole batch, instead of between every source.

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

Pull and drain first

Before planning anything, run:

git -C vault pull --rebase
python scripts/telegram_capture.py --once

/ingest-all compiles what is in raw/inbox/, and two things put files there that this copy of the vault has not seen. The scheduled jobs on GitHub commit to the vault there, and with the Telegram relay set, every message is committed there as it is sent. Without the pull, "compile everything" silently means "compile everything this copy happened to have", so a note sent from your phone an hour ago is missed, and the run reports success anyway.

The drain is for when the relay is off, since Telegram messages then land in the inbox only when that script runs. With the relay set it says a webhook is set and exits. --weekly already drains before writing the brief, for the same reason. Draining twice costs nothing: the script tracks an offset, so a second run in the same minute fetches nothing.

Unattended, skip both. scripts/autopilot.py --capture drains immediately before starting this skill, and on GitHub the vault was checked out moments earlier. A headless run cannot get permission to run either command anyway, and trying only put a warning in every nightly log that messages might be missing when they were not.

What writes itself, and what waits

Not every source carries the same risk, so not every source needs the same approval. Plan the whole batch first, as always, then split it.

The test is which store the plan writes to, which is the two-store distinction in CLAUDE.md applied one step earlier, at the decision of whether to write rather than only at what may be written.

The plan writesStakesThen
wiki/ onlyrebuildable from raw/ at any timewrite it
a loop carrying a datea nudge fires, or fails to firehold
a change to a date already recordedthe same, and it overrides something you sethold
anything in mem/nothing can reconstruct ithold, you may only propose there anyway
a claim contradicting a page in the vaultboth must be kept and the conflict recordedhold
more than one reading of what the source meansa guess becomes a fact once writtenhold

A wrong wiki/ page costs a regeneration. A wrong date costs a reminder that does not arrive, or one that arrives about the wrong thing. That asymmetry is the whole rule.

The mixed case is the common one. An article saved is knowledge; "I should read this" is a commitment. Write the source page, hold the loop. The reader keeps what was learned and still decides what is owed.

Held sources stay in raw/inbox/. Do not move them, do not mark them, do not write a holding file. Whatever remains in the inbox after a pass is by definition what waited, which is what the brief counts and reports. Re-planning them on the next pass is intended: the vault has changed since, and a source that was ambiguous last night may not be tonight.

Say what you did. Every source that wrote itself gets one line in the report, and every source held gets one line saying which row of the table above caught it. Nothing enters the vault unseen, and nothing waits without a stated reason.

Promotion

Compiling the whole inbox at once is the only moment the compiler sees across sources, so it is the right moment to promote. After writing every source, run python scripts/synthesis.py and create the entity and concept pages it lists.

This is where the wiki stops being a pile of source summaries. A vault with fifty source pages and no entity pages answers "what do I know about X" by re-reading fifty pages and learning nothing that persists.

Procedure

1. List what is waiting. Every file in raw/inbox/ with status: uncompiled, oldest first. Say how many. If more than 10, do the oldest 10 and say the rest are still queued, since a batch too large to read the plan for is a batch nobody actually approves.

2. Read them all, and orient once. Read index.md and the existing pages that any of them plausibly touch. Doing this once for the batch is the main saving over running /ingest repeatedly.

3. Plan the entire batch before writing anything. One block per source:

1. 2026-08-30-collect-boardgames.md
   pages   wiki/sources/… (new), wiki/concepts/boardgames (new)
   loops   "decide which boardgame to buy" — overlaps #3 below
   flags   none

2. 2026-08-30-start-learning-to-cook.md
   pages   wiki/sources/… (new)
   loops   "start learning to cook" — due end of September
   flags   none

3. …

Total: 8 sources, ~14 pages, 5 loops, 1 contradiction.
Proceed, or shall I hold any of these back?

Look across the batch, not just within each source. This is the thing one-at-a-time compiling cannot do:

  • Duplicate loops. Three notes about boardgames should produce one loop, not three. Say which you are merging.
  • Entity promotion. A name appearing across several sources in the batch may cross the three-mention threshold now. Promote it, and say that you are.
  • Sources that contradict each other, not only ones contradicting the existing wiki.

4. Wait for approval. Do not write before it. One approval for the batch is the point; zero approvals is not.

5. Compile in order, applying /ingest's rules to each. Report a line per source as you go, so a long batch shows progress rather than going silent.

6. Stop mid-batch and ask if any of these happen, rather than pushing through:

  • a source would touch more than 15 pages on its own
  • a source is unreadable, or its attachment: is missing
  • a source contradicts something you already wrote earlier in this same batch
  • the total is heading past ~40 pages

Report what is done, what stopped you, and what remains. A half-finished batch you were told about beats a finished one that quietly went wrong.

7. Update index.md once at the end, not per source.

8. Append one log entry per source, because the log is a per-source record and batching must not collapse it, or /unsource loses the trail:

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

Then one batch line:

## [YYYY-MM-DD] ingest-all | N sources, N pages, N loops

9. Report in five lines or fewer: sources compiled, pages created and updated, loops opened, loops merged, contradictions flagged, anything held back.

When not to use this

The first ten sources of a new vault. Compile those one at a time, because you are still learning whether the compiler reads your material the way you would, and per-source approval is how you find that out. Batch once you trust it.

When the inbox is a mixed bag of important and trivial. A batch plan encourages skimming. If two of eight actually matter, /ingest those two properly first.

Rules

Everything in /ingest holds. Emphatically:

  • Every claim cites a source and a locator. More files is not a reason to get loose.
  • Never write mem/. Propose at the end; the human decides.
  • raw/ stays immutable apart from status:.
  • Be conservative about loops. A batch is where false loops multiply fastest, and brief precision is the metric this project lives on. When unsure, list it rather than filing it.

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.