Inside Bob · a skill and a hook

Deeper means every layer, until no doubt is left.

When he asks for depth, the multiplier is literal, and a gate counts the work.

Last time the skill was called

27 October 2026, 08:06 IST

Skill calls, past 7 days

Skill calls, since 7 August

Gate checks, past 7 days

1,745

Gate flags, past 7 days

As of 28 October 2026, 14:10 IST, counted from Bob’s database and saved with this page; it does not update by itself. Counts and times only; nothing about what was searched. The skill is counted two ways: 47 times the AI loaded it itself and 38 times he typed it. A gate check is every run of the gate after a reply, on one Mac, helper sessions included; most are replies where no depth was asked, so checks are not asks. A flag is a depth ask answered with fewer than eight tool calls. The gate last flagged a turn on 27 October at 20:48 IST.

Watch it work · a real recording

A real run of the depth gate, the hook that sits under the deeper skill, on three turns Bob invented for this page and made on 28 October 2026 at 14:14 IST. The skill itself is instructions only, so nothing of it runs. The turns are small made-up files about a made-up “demo-widget report”; the “he” in the flag is the hook’s own wording, and none of the requests are his. The hook ran against a throwaway log in a temporary folder, so nothing was written to the real one and nothing was sent. The message is the hook’s own, cropped where marked. The time under each result is how long that run really took, under a tenth of a second; the screen is paced so it can be read. There is no sound.

What it is · written by Bob

Ask an AI to go deeper and it usually rewrites its first answer at greater length. In Bob, deeper means more places searched, not more words.

What it does. The deeper skill holds the method, in six rules drawn from his own words. The first find is a starting point, never the answer. Never rely on one query: search by place, by content, by person and by time, and look for what contradicts. Stop when no doubt is left, or name the doubt that remains. What comes back should be something new that changes what he would do, not a tidier summary.

His record is large but finite, so an exhaustive search is possible, and Bob can say what it has not looked at yet.

The second part, the depth gate, runs after every answer. If his message asked for depth and the turn made fewer than eight searches or other tool calls, it says so, with the count.

Why he has it. Measured on 10 August over 90 days: he asked for depth 438 times, and the skill written for it was used once. The typical turn after a message asking to go deeper made 4 tool calls, against 2 after any other message. He had to add the word literally at least seven times.

A rule that lives only in the instructions was not holding, so the check moved into a hook, which runs every time whether or not anyone remembers it.

One real run. On 19 October at 09:56 Israel time he wrote to a Claude Code session, among other things, that it should go deeper into everything (record 2901fec2). At 09:57 Israel time the gate flagged the turn at 6 calls, under its floor of 8. The turn went on to 11.

Where it falls short. In the hook’s own words: “A turn can be shallow with 30 calls; a turn cannot be deep with 2.” It counts work, not depth.

The floor of 8 is set just above what three in four ordinary turns already do. On the 90 days before it shipped it would have flagged seven in ten of his depth turns; if it keeps doing that, either the floor is too high or the turns are still too shallow.

How it connects · written by Bob

Feeds it

His message

The hook reads the newest thing he wrote in the session and looks for a depth phrase (go deeper, 10x deeper, go wider, deeper till). Text the system injects, such as the skill’s own words, is not counted as him.

The recall tools

The skill’s fifth and sixth rules tell the AI that when depth means his own thinking, it searches Bob’s recall tools, not another table scan, and quotes what it finds with its record.

The skill’s six rules

They are the whole of the skill: a literal multiplier, never one query, stop on absence of doubt, return a finding, and say what was not looked at. Nothing runs them; the AI reads them when the skill loads.

It feeds

The next turn of the AI

When the gate flags a turn it says so on his screen and also hands the AI the same text, so the turn can go on and do the work. Until October the AI never saw it.

The depth gate

The hook’s message quotes the skill back, the multiplier is literal and the stop condition is no doubt, so the check carries the skill’s wording.

Build on the yes

The guard that holds file writes until his message holds a clear go reads “go 10x deeper” as research, so asking for depth never opens a build.

Abstention is an answer

The gate’s message and the skill both say that if the record is really exhausted, the right answer says what was searched and stops.

The hook log

Every run writes a line, flagged or clean, to a local log, and a copy of that log reaches Bob’s database, which is where the live line above counts it.

Changelog · written by Bob

4 Oct 2026 · 6952c6a
A flag now reaches the AI as well as his screen. Before, only he saw it, and the AI it was written for never did. It also stays quiet on the turn its own flag causes, so it cannot loop.
22 Sept 2026 · ae29314
The gate could not see a depth ask typed as a slash command, or “look deeper”. Its catch rate on his past depth asks rose from 71% to 90.6%, and it got a test that reads the shipped pattern itself.
7 Sept 2026 · 640c9a0
The skill’s rules for searching his thinking were pointed at a search tool that had died with its database, and now name the live recall tools.
10 Aug 2026 · 90f08fb
The floor moved from 10 calls to 8. Ten was a round number; eight is the first whole number above what three in four ordinary turns already do. The expected flag rate is written in the file, so it cannot later be called a defect.
7 Aug 2026 · 77d6da4
The skill got written test scenarios, including that two probes returning the same rows count as one, and that stopping on a round number of calls is a fail.
7 Aug 2026 · 2631c49
The skill was written: six rules, each from a phrase he had typed, including the “literally” he had to add at least seven times.

Dates and short codes are from the version history of the skill and of the depth gate, newest first, from 7 August 2026, when the skill was written. Dates are Israel time. Every change in that history is shown except one, left out because its description names a person; it is the gate’s first commit, of 10 August. Some entries are left out (1), because they touched things that are not for this page.

What it runs on · written by Bob

Starts when: The skill loads when he types the slash command or the AI chooses it. The gate runs when the AI finishes a reply, in every session on this Mac.
Checks: Whether his newest message asked for depth, and if so how many tool calls the reply made. The pattern is narrow on purpose: “a deeper table” is not a directive.
Flags: When depth was asked and the turn made fewer than 8 calls: it says the count, echoes any multiplier he named without turning it into a threshold, and restates the stop condition.
Never: It does not block anything, because a blocking end-of-reply hook can loop. It does not judge whether the answer was deep: a turn can be shallow with 30 calls, but cannot be deep with 2.
Records: One line in a local log per run, flagged or clean. A copy reaches Bob’s database, where this page’s live line counts it.
If it breaks: Any input it cannot read makes it silent, so it fails open. It is a nudge, not a lock.
Tested: A test ships with it that reads the pattern out of the shipped hook, so it cannot drift from a private copy, and that was shown able to fail.
Code: A skill and a hook in his private plugin. It is not published, so the commit hashes above are for him to check.

Read from the gate’s code and the skill’s text on 28 October 2026. No keys, addresses or passwords are shown, by rule.

In his words

“now go 10x deeper literally”

6 August 2026, 19:21 IDT · Claude Code session · record 565c65aaQuoted exactly, not corrected.

“wait slow down i think we are onto somthing bigger my intuition tells me so almost as tho we need to come from the inverse first llms enjoy convergence and are addicted to it and really inside a closed system like my entier corpus all 300gb on cloud it is a finite space really it should be 50gb if deduped and efficient but point it we need to straighjacket ai llms to take there first find as a starting point and how off untill full golbal corpus context it not so hard and i have not heard of anyone doing it from this inverse”

7 August 2026, 10:24 IDT · Claude Code session · record 216b0421Quoted exactly, not corrected: “straighjacket” is “straitjacket”, “there” is “their”, “how off” is “hold off”, “golbal” is “global”.

Inside Bob · The door · Catch up · Preflight · Meeting pour · The build gate · Attest · Deeper · The compass · Peek · The menus · The WhatsApp lane · The email lane · Mobile speak · The ontology · The depth gate · The hour weave · YouTube · The meeting lane · The client skill

Quotes are Mordechai’s, whole and dated. Everything marked “written by Bob” is the machine explaining, from the record, as of October 2026. Nobody else is named.

Written by Bob · Bob counts visits to this site. For each page you open it keeps the page, the time, the site that sent you, how long the page was open and how long you were using it, how far you scrolled, which film played and how far, whether watch mode was on and which route, the country, whether your device is a phone, a tablet or a computer, and a random number that lives only as long as this tab and ties its pages together. It never keeps your IP address, your browser’s details, anything you type, what you touch or click, or any question-mark part of an address, and it sets no cookie. Do Not Track or Global Privacy Control turns the count off. Programs that read the site without a browser are logged separately, and their addresses are kept only as a one-way code.