absorb-sources

Generated source view for the actual executable productivity/absorb-sources skill. The durable routing article is Agent Operations Skills. Source: skills/productivity/absorb-sources/SKILL.md

Runtime Source

Field Value
Category productivity
Origin unknown
Slug absorb-sources
Source slug absorb-sources
Family Agent Operations Skills
Source skills/productivity/absorb-sources/SKILL.md

Bundled Resources

These files are part of the executable skill folder and must be preserved with the skill source.

File Role
agents/openai.yaml Agent UI metadata

Description

Absorb saved sources (Discord, X, Hacker News, articles, repos, media, notes, URLs) into Kevin Wiki: expand each one, find the pages it changes with kb, write cited changes. Use when Kevin asks to save, absorb or research sources.

Skill Source

---
name: absorb-sources
description: >-
  Absorb saved sources (Discord, X, Hacker News, articles, repos, media, notes, URLs) into Kevin
  Wiki: expand each one, find the pages it changes with kb, write cited changes. Use when Kevin
  asks to save, absorb or research sources.
---

# Absorb Sources

Turn source material into changes on the wiki pages it concerns. A queue row, a
report mention, an engagement score or a topical page is not proof of learning.

## Run the loop

1. Read the source in full: thread and author replies, linked articles,
   repository README, docs, code and releases, attachments, images and video.
   Mark anything you could not open.
2. Correlate platforms before interpreting them. Treat Discord as Kevin's intent
   when it points at an X post, a Hacker News thread, a site or a repository,
   and keep both source identities.
3. Find the owner pages with `kb <words>` and open them with `kb get <name>`.
   Start with the smallest set of pages that answers the question. Widen to
   neighbors and raw evidence only when the source names them.
4. Extract every useful claim, preference, task, example, question and
   counterexample. Compare each with the owner page's current text and decide:
   reinforce, refine, replace, contradict, create, or no change with a reason.
5. When a source recommends an agent platform, separate the portable context
   layer from the execution layer. Rules, skills, MCP contracts, source identity,
   approval policy and proof stay owned and portable across runtimes. Sandboxes,
   model access, fleet orchestration and hosted agent UX may be rented when they
   are replaceable and never become the only canonical writer. Record the
   source's incentives and test its claim against primary evidence.
6. Patch the smallest owner page. Create a page only when no page can own the
   idea. Update a workflow, skill, design rule or project page when the evidence
   changes how agents should work. Route a recurring procedure to
   `skill-creator`.
7. Cite each change inline with `[Source: ...]` and add a dated Timeline line.
   Bump `updated:` only on pages you checked against their sources.
8. Prove retrieval: `kb <words a future agent would type>` ranks the changed
   page in the top 3; add aliases when it does not. Then run the closeout in the
   wiki `AGENTS.md` and append `wiki/log.md`.

Read the whole semantic unit of the current source even while comparison
context stays narrow. A narrow focus limits unrelated corpus loading; it never
licenses skipping an article body, repository surface, thread or reply.

From another directory, do not edit the wiki. Queue each fact with
`kb note <page> "<fact>" --source "<url>"`; the weekly harvest writes it in.

## Agent conversations

Agent tasks are a source class. List them, then read the transcript the list
names:

```bash
npm run brain -- sessions index
npm run brain -- sessions list --provider codex --since 2026-07-01 --workspace Kevin-Wiki --query "workflow" --limit 50
```

Each row ends with the transcript path. Read completed turns, and the tool
output when it supports a claim. Titles and summaries are discovery metadata,
not evidence. Do not treat a running task as complete. For an authorized
export outside the default stores, add it first with `sessions index --path
<absolute-directory> --provider codex|cursor|claude`. Correlate the task's
workspace with its project page and current checkout before changing the page.

For X bookmarks, follow `skills/productivity/x-bookmark-absorb/SKILL.md` for
platform-specific artifact recovery, then return to this loop.

## Storage and publication

Keep exact private source bytes under `raw/` or the ignored `.brain/`. Never
make a public page depend on private source bytes.

Timeline

1 page links here