agentleFS
open-science-pillarsopen-science-pillars/core/skills/consult-knowledge/SKILL.md

consult-knowledge

Skill0 starsChanged 53 days agoApache-2.0View on GitHub

Consult installed knowledge bundles before acting on a dataset: how to find every bundle, which directories to search, how to cite a match and voice its status, which concept wins on conflict.

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

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

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

    When two concepts disagree, the provider bundle wins over a plugin-local concept (a bundle installed from a provider plugin such as …

    line 108
  2. Distinctive:

    Never carry a concept's fact into a skill body.

    line 124
  3. Distinctive:

    Never resolve a disagreement between concepts silently.

    line 125
  4. Distinctive:

    Never treat an unread concept as consulted.

    line 126

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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

consult-knowledge

The one consult convention every Open Science Pillars skill and agent follows before it acts on a dataset, a quantity, or a method. Skills carry procedure and hard refusals; facts about data live in knowledge concepts and are read fresh per run. A concept added or corrected since the last run is found this way and changes behavior with no skill edit.

A skill that cannot load this file carries the five-word form instead, verbatim: "Consult installed knowledge concepts first."

Where to look

Bundles arrive as installed plugins. The installer's record of what is installed is the list of places to search, never a remembered path and never a cache directory walked by hand (a cache keeps superseded versions for a while; the record names the live one). Where a shell is available, claude plugin list --json prints the record: one entry per plugin with its version, installPath, enabled flag, and an errors field when a dependency is missing or out of range.

The record is not the whole list. It is written by the installer, so it does not know this session: a plugin loaded from a directory for development is absent from it, and a plugin the session turned off (or on) does not have its enabled flag changed to match. So search the running skill's own root first, whatever the record says. That root is in ${CLAUDE_PLUGIN_ROOT} for every skill the session loaded, install or development directory alike, and it counts as a bundle root; a concept the skill names by bundle path resolves there before anywhere else. A search that skips it reads a stale installed version of the very bundle the session is running, finds the concept missing, and says so: the failure looks like a knowledge gap and is not one.

A bundle root is a directory holding an index.md and log.md with concepts in typed directories: <installPath>/knowledge/ when the index is there (core, the domain plugins), or each directory one level below it that carries its own index (a provider plugin ships one bundle per provider: knowledge/podaac/, knowledge/esdis/). Search every root of every enabled plugin, plus any bundle a project's local config names. A plugin whose record carries errors is disabled by the installer; say that its bundle was not searched and quote the error, which names the install command that clears it.

Where the record cannot be read (a surface without a shell), search the bundles reachable from the running plugin's own root and the local config, and say which bundles were searched. Say the same when a concept is found under ${CLAUDE_PLUGIN_ROOT} but not in the record: name the version the record lists, so the reader knows the concept is newer than the install.

A skill or agent that names a concept by bundle path (knowledge/podaac/gotchas/<name>.md) means the concept under that path in whichever installed bundle root carries it; resolve the path against every root above, and cite the concept as found there.

Under each bundle root, glob the directories that exist:

directorywhat it answers
datasets/product identity, access, versions, native grid, the Uncertainty section
fields/per-variable facts: names, units, sign, staggering, DOIs
gotchas/traps: mechanism, wrong-result mode, the correct approach
recipes/validated analysis patterns with expected ranges
computations/attested computations: sanctioned code, inputs, receipts, the reference numbers recipes and findings quote
conventions/cross-cutting practice: calendars, fill values, CF, sanity ranges
validity-domains/where a product's claims hold and where they do not
findings/scientific claims with receipts, signed or draft
connectors/endpoint facts for interactive services

references/ holds files concepts cite (code, receipts, masks); it is reached through the citing concept, not searched on its own.

Grep by product name and ShortName, variable, quantity, method, region, and the check in play. Read every match in full. Where a concept exists, never work from a remembered number, list, or rule.

What to do with a match

Restate what each concept changes about the plan, at the step it constrains, and cite it by bundle path (knowledge/gotchas/ecco-native-vs-regridded.md) with its status. Numbers come from the concept, or from the computation receipt it cites, at the precision written there. A concept that shaped a choice is cited; a concept that is cited was read.

Voicing status

Concept status is stable unless the frontmatter says otherwise.

  • stable: state it plainly. For a high-severity claim (a wrong-result mode, a number a conclusion rests on) say how it was verified: a human: verified event is human-reviewed; a process-only event is machine-confirmed; none is unverified.
  • draft: consultable, voiced as unverified ("a draft concept, not yet reviewed, says ..."). Never presented as settled.
  • deprecated: never cited as current; follow superseded_by.
  • disputed: present: state the dispute and its link beside the claim.
  • past stale_after: say the sweep is due and the fact may be outdated.

Precedence

When two concepts disagree, the provider bundle wins over a plugin-local concept (a bundle installed from a provider plugin such as nasa-daac-knowledge is the provider tier, and a transitional copy of its text ranks with it); stable wins over draft; within a tier the later verified event wins. Say that they disagree. A concept wins over anything a skill or an agent remembers, including this file.

When nothing matches

Say so ("no installed concept covers X"), proceed on general method labeled as such, and never invent a concept, a path, a number, or a citation. The gap is ingest material: name it in the Provenance note so the steward can open a concept issue.

Must NOT

  • Never carry a concept's fact into a skill body; the skill points here.
  • Never resolve a disagreement between concepts silently.
  • Never treat an unread concept as consulted.

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.