Open source · Markdown · Obsidian

Keep knowledge.
Keep its evidence.

Capture a source, check the key claims, and bring what you learn back into your work.

The website explains it. Your vault holds it.

These are installable Agent skills. This is a showcase, not a hosted vault for your private files.

Skill
Instructions and resources loaded for a task. A skill guides the workflow; it is not a change to model weights.
Vault
Your original sources, maintained knowledge and audience-specific outputs. Existing conventions can stay.
This website
A public example of the method, results and installation. Buttons switch prepared example content.

What happens to an unsupported number?

This setup-test example is entirely fictional. Choose a step to see the input, screening, evidence and result. This page does not run AI or live research.

Raw/meeting.md

Original input

“Three testers paused during setup. Terminology increased completion time by 45%. I find the interface unfriendly.”

Keep the source unchanged, even when a later conclusion differs.

Intake preview

Select claims

Prioritize: the origin of 45% and the claim that terminology caused delay. Preserve as opinion: the author’s interface preference. Skip: background unrelated to this task.

Select by relevance to the task and its important premises.

Raw/test-log.md

Inspect evidence

Fictional test log: three pauses across five tests. No terminology control group or timing result supports 45%. Status: observation supported; cause and percentage unresolved.

Missing evidence is not disproof. Repeated copies are not independent verification.

Wiki/setup-pauses.md

Maintain a note

Three of five fictional testers paused during setup. Terminology is one possible explanation; causality and effect size remain unresolved. Sources: meeting note + test log. Preserve 45% as an attributed unchecked claim.

Maintain a bounded conclusion linked to its original evidence.

knowledge-retrieve

Retrieve it

Question: can we confirm a 45% increase due to terminology? Answer: no. Inspected records support pauses but provide no timing or control evidence for 45%. Source: Wiki/setup-pauses.md → Raw/test-log.md.

Return supporting sources and gaps together; do not invent an answer.

knowledge-impact

Trace the impact

Proposed: replace “terminology caused 45% delay” in Outputs/team-brief.md with the bounded observation. Preserve: original meeting quote and dated snapshots. State: proposal only; no files edited.

A changed conclusion does not authorize editing every dependent file.

Keep Raw unchanged, revise Wiki with evidence, and propose updates to Outputs.

Choose for the task you have.

Use $knowledge-vault if you are unsure. It routes only to installed modules and reports missing capabilities.

Capture this item

$knowledge-capture

Select a link, screenshot or text. Preserve its source and extraction limits.

Find existing knowledge

$knowledge-retrieve

Answer from inspected notes, with sources and missing evidence.

Trace a change

$knowledge-impact

List affected notes, decisions and outputs before editing.

Choose the first batch

$knowledge-triage

Inventory a batch and identify duplicate candidates without ingesting everything.

Inspect the vault

$knowledge-health

Audit links, evidence gaps and conflicting state without automatic cleanup.

Explore the original 11 core modules

Start with one installation profile.

Clone or download the repository, then run from its root. The command changes with your selection. Replace the target with your Codex skills directory.

python -X utf8 scripts/install_codex_skills.py <codex-skills-directory> --profile lightweight --dry-run
python -X utf8 scripts/install_codex_skills.py <codex-skills-directory> --profile lightweight
python -X utf8 scripts/install_codex_skills.py <codex-skills-directory> --profile lightweight --verify

A first request

“Use $knowledge-capture for this screenshot and keep the original. Then use $knowledge-intake to select the key factual premises and preview the intake.”

Use --dry-run to inspect the plan. Writes require authorization, and existing permission remains valid within its scope.

Repository and full documentation

Capabilities and their limits.

Does it verify every sentence?

No. It first selects key premises. The default is three claim groups and two search rounds each; unresolved premises stay explicit and originals remain unchanged.

Does it crawl my saved social posts?

No background account scanning. You select one link, screenshot or text. Access barriers are respected. Computer use does not guarantee avoiding platform detection.

Is this an Obsidian plugin?

No. It is a Markdown-first Codex skill suite; Obsidian properties, links and Bases are optional enhancements.

What has actually been verified?

Automated checks cover structure, links, installation consistency, failure recovery and evaluation manifests. Live model behavior still needs evaluation; this fictional example is not an execution result.

Read the evaluation protocol

Where do the methods come from?

Core modules draw on provenance, atomic notes, human-centered design and reflective practice. These explain design choices, not proven effectiveness of the complete system.

Foundations and limits

Want to share a usage result?

Share feedback: include the task, observed result and corrections needed. Leave private source material out.