Context pack for the Manus agent, 2026-08-01

This pack replaces the picture of our system that produced 23 unusable video tutorials. It is written for a capable engineer-grade agent that currently holds a wrong model and needs to swap it out. Read it before writing anything that claims to describe how we work.

Why you received this

The tutorials were grounded in the wrong software. An earlier pass read the starred and forked repositories on the jonnydubowsky GitHub account and treated them as our working stack. They are reference material and abandoned experiments.

Across the 23 tutorials, graphify appears 97 times, knowledge-base-server-demo 78 times, obsidian-mcp 61 times, SimpleMem 46 times, mempalace 42 times, and NexAU 24 times. None of those six is part of the working system. Every instruction built on them names a tool that is not installed and not called here, so it cannot be followed.

The one-line correction: the work runs out of an Obsidian vault driven by Claude Code, with skills, agent specifications, and Bases dashboards, plus a small set of Python and Ruby tools inside that vault and Cloudflare Pages and Workers for every deploy. There is no conventional application repository behind it.

What is in the pack

Documents at the pack root:

File	Purpose
`README.md`	This document. Orientation and reading order.
`GROUND-TRUTH-2026-08-01.md`	The authority. Verified condition of the system on 2026-08-01: what runs, what is prototype, what is specified and unbuilt.
`TECHNICAL-BRIEF.md`	What each component does and how the three-layer knowledge model composes.
`DEV-SETUP-STATE.md`	The engineering setup as it stands: GitHub, Jira, the monorepo, the gates, the feature register, the roles, the order of work.
`CONCEPT-SPACE-READER.md`	What we mean by DKR, how Scout finds gaps, how concept-space research is organized, and the writing rules that gate all of it.
`shuriq-okf-profile.md`	The Open Knowledge Format profile, v0.5.1: type vocabulary, extension keys, guards, conformance rules.
`PROMPT-1-INGEST.md`	The prompt that tells you to discard the old model and take on this one.
`PROMPT-2-REWRITE-TUTORIALS.md`	The full tutorial specification, addressed to you. Read it when you are asked to write or redo a tutorial.

Folders:

Folder	Contents
`blog-posts/`	Fifteen concept posts copied out of the vault unedited, plus `INDEX.md` describing them and the renumbering applied on copy. Read for voice and level of specificity.
`tutorials/`	The corrected tutorials, one per video, plus `grammar-allow.txt`. 21 of the 23 ship here. `WITHHELD.md` names the two held back and why. These are the worked examples for Prompt 2.
`okf-bundle/`	An Open Knowledge Format bundle of 72 concepts covering the ground truth, the context documents, all 23 tutorials, all 23 video notes, the analysis, and the 15 posts. Reserved `index.md` and `log.md` at its root. Full document bodies, because a `vault://` pointer does not resolve on your side. Layer 2 throughout.

Three items in those tables are copied in at zip time from elsewhere in the vault, so check they arrived before you start: GROUND-TRUTH-2026-08-01.md and tutorials/ come from projects/dkr-agency/concept-space-2026-08-01-video-corpus/, okf-bundle/ is built by build_okf.py in that same folder, and shuriq-okf-profile.md comes from projects/shur/okf/. If any of them is missing from your copy, say so and stop rather than working around the gap.

What is redacted, and how to read it

Client and engagement names are replaced throughout this pack with stable pseudonyms: [CLIENT-A], [CLIENT-B], [CLIENT-C]. The same bracket always means the same client. Where a file path contains one, the path is otherwise accurate and the real folder name is the client's own.

Two tutorials and two sections of the OKF profile are held back. Each place says so where the removal happened. Nothing else is edited.

Do not try to identify a redacted client, and do not ask for the withheld material in your output. If a task seems to need it, say what you would need and stop.

Reading order

GROUND-TRUTH-2026-08-01.md, in full. Everything else is subordinate to it.
TECHNICAL-BRIEF.md.
DEV-SETUP-STATE.md.
shuriq-okf-profile.md, then okf-bundle/index.md and a few concepts under it.
CONCEPT-SPACE-READER.md, then three or four posts from blog-posts/.
Three tutorials from tutorials/: 103-google-open-knowledge-format.md, 102-284b-model-on-a-laptop.md, and 101-lost-discipline-of-the-alarm.md.

Where two documents disagree, GROUND-TRUTH-2026-08-01.md wins, and report the conflict.

What to do next

Run PROMPT-1-INGEST.md first. It asks you to discard the six repositories by name, to state the components you now hold and whether each one runs, to name the three largest corrections against your prior model, and to list what the pack left you short on. It also sets the rules for consuming and emitting OKF, and the writing rules a program checks.

Then run PROMPT-2-REWRITE-TUTORIALS.md per video. It gives the full list of real components you may draw on, the tense rule for unbuilt things, the required ## Verify it worked section, and the frontmatter shape to return.

One standing rule across both: if a path, repository, ticket, command, URL, or tool is not in this pack, it does not exist. Say what is missing and stop. Filling the gap with a plausible name is the exact failure this pack corrects.

What this pack deliberately does not include

This is a Layer 2 export, decided by policy rather than by convenience. Withheld:

SBPI dimension weights, scoring formulas, and rubric internals.
Typed edge weights and confidences from the myKG store.
Any client's raw material, scores, or name-linked findings.
Anything under the sbpi* or dkr_* frontmatter key namespaces.
Internal self-audit and methodology.

The OKF bundle is poorer than the store it came from for the same reason. Its resource: URIs are vault:// pointers to documents you do not have and cannot fetch.

Do not ask for the withheld detail and do not reconstruct it by inference. If a task seems to require it, say which decision you cannot make without it and stop there.