The file boot/operating-system.md is 225 KB, 4,982 lines, 24 chapters. It loads in full at every session, whether the task is writing an outreach email or fixing a button on a mini-site.
Measured on 27 July sessions, the preamble is ~112,000 tokens and the working context tops out around 167,000 before it is rebuilt — so roughly two thirds of it is spent before the conversation starts. Around 450 of the 4,982 boot lines are needed at all times.
Measured effect over July: one context loss every 3.5 messages. On 16 July, 38 messages and 18 context losses. Each time, the conversation is rebuilt from a summary — what was said thirty minutes earlier comes back compressed, and some of the nuance is gone.
The clearest example. Chapters 13, 14 and 20 — the solitude of leadership, flow conditions, the personal side of relationships — come to 905 lines, 18% of the file. They serve exactly one skill, /coaching. They still load on every email sent to a candidate.
The instinct is right. Requests like "log this as best practice", "make this template a prerequisite for every search", "add to it, don't replace it" are exactly how to use a system that accumulates. All 23 skills were built that way.
One thing is missing, and it is mechanical: writing the file is not enough for it to be read back. Of the four requests of this kind in the last ten days, three produced a file that nothing reloads.
| Request | File produced | Read back by |
|---|---|---|
| PDF button, mini-site best practice (17/07) | wiki/protocols/minisite-print-pdf-button.md | nothing |
| Personalised candidate site template (23/07) | no file created | nothing |
| "Candidates' reach" approach (23/07) | diary/2026-07-23-pattern-candidates-reach.md | nothing |
| Career patterns | wiki/career-patterns/index.md | /analyse-profil |
The last one works, for a specific reason: the /analyse-profil skill declares that file as its source. That is the only mechanism that brings knowledge back. It already exists — it has simply been wired once.
What it costs. The "candidates' reach" pattern was logged on 23 July. On the 24th it had to be pointed to by hand, then the same two points had to be corrected again — the opening hook, the cultural dimension raised at intake — before being logged a second time. The same lesson was taught two days running.
The split comes down to three categories. The question to ask about any piece of content is simple: is this true all the time, or only when I am doing one specific thing?
| Level | Loads when | Holds |
|---|---|---|
| Boot | always | The invariants: founding beliefs, red lines, decision framework, behavioural rules. What has to hold true even in a three-message conversation. Roughly 450 lines. |
| Skill | when the skill is called | The method behind one specific job: assessing a candidate, preparing a client meeting, writing an email. Only one skill uses it. |
| Shared pattern | when a skill declares it | What several skills need: relationship principles, decision filters, tone of voice. Written once, referenced by each skill that uses it. |
The file already started the job. Every chapter of the operating system ends with an OS Mapping section: a condensed YAML digest of that chapter. The short version already exists, chapter by chapter. That is what stays in boot; the long version moves into the skill.
Content used by a single skill
| Chapter | Lines | Destination |
|---|---|---|
| 13 Solitude of leadership · 14 Flow states · 20 Emotional & relational self | 905 | /coaching |
| 24 Recruitment Brain V2 — the 14 operational engines | 790 | /delivery |
| 23 Candidate Evaluation Engine — scorecard, decision formula, question bank | 373 | /assessment |
| 17 Client Reading & Diagnosis — placeability grid, rule of 3 O | 342 | intake / /delivery |
| 16 Commercial Philosophy — selling without selling | 200 | /redaction |
| 19 Client Meeting Method — the 7 preparation principles | 144 | a /call-prep skill, to be created |
Chapter 23 makes the point on its own: the description of the /assessment skill says it produces assessments "according to the methodology". The methodology sits in boot, not in the skill that applies it.
Content used by several skills
| Chapter | Lines | Becomes | Referenced by |
|---|---|---|---|
| 12 Relationships & Trust — the 15 principles | 260 | patterns/relationship-principles | /redaction, /network, /followup, /coaching |
| 2 Decision Framework + 7 Systematic questions | 71 | patterns/decision-filters | /coaching, /delivery, /assessment |
| 24.0 Bias Monitor Engine | ~40 | patterns/bias-monitor | every skill that evaluates a person |
| Yoni's tone of voice (currently scattered) | — | patterns/tone-of-voice | /redaction, /candidat, /send, /dashboard-client |
What stays in boot: quick-reference card, architecture, founding beliefs, decision framework, energy model, Tribe OS, systematic questions, red lines, conflict resolution, behavioural rules. Around 450 lines out of 4,982.
A real example: writing an outreach email for the STRAREX role.
- 4,982 lines loaded, including the solitude of leadership and flow conditions
- ~112,000 tokens spent before the first exchange — two thirds of the usable context
- The outreach pattern logged the day before is not reloaded — it has to be pointed to by hand
- Context saturates after a few exchanges and earlier corrections are lost
- ~450 lines of boot, then
/redactionloads what it needs — most of the context stays free for the work - Commercial philosophy, tone of voice and relationship principles arrive with the skill
- The outreach pattern is declared as a source of the skill: it comes back unprompted
- Context holds for the length of a real working session
The same content stays available. What changes is when it arrives.
When a best practice is established, there are two moves, not one.
1. Write the file — this is already being done, and done well.
2. Declare that file as a source in the skill that will need to read it back — this is the step that is missing.
Without the second move the file exists but nothing goes looking for it. Three cases out of four in the last ten days.
In practice, the wording to use: "log this as best practice and attach it to /redaction". The difference is four words, and it changes the outcome: the first produces a note, the second produces something the system keeps.
/coaching lines out of boot/coaching picks it up when it is called.OS Mapping digest of each chapter stays in boot, the long version moves into the matching skill./analyse-profil, which already works this way.Two points found in the June and July sessions are not about how the brain is organised. They are fixes on our side.
| Issue | Finding | Status |
|---|---|---|
| Word documents not read | 15 distinct documents uploaded to the webchat, none usable — all of them intakes (client intakes, executive profiles, competency grids). Until this is fixed, retyping the content by hand is the only option available. | Fix specified, no new dependency required |
| Client dashboards | The dashboard shows an automatic refresh but is not connected to any source: the data only updates when an agent regenerates it. The HireSweet sync was asked for three times in one week — the request was legitimate. | Either a scheduled sync, or the misleading refresh comes out |