Your Claude Code keyword codes

The Codes

Eight phrases that change how Claude works. What each one does, when to say it, and what it spends.

8 codes · 2 modes · 4 helper agents · first used Jul 11, 2026

✦

Start here

All 8 at a glance

Code Blackmode1 to 2 agentsProve it before saying it: a plan, two sources, one critic, pasted evidence.say code blackCode Greyone shot1 agentOne critic reads one change and says ship, modify or reconsider.say code greysanity check thisCode Orangeone shot3 agentsWide web research on cheap agents, then one ranked answer.say code orangethoroughly researchCode Whiteone shotno agentsParks this topic on disk, then says /clear or /compact.say code whiteswitching topicsCode Blueone shotno agentsSaves memory, commits and pushes. No questions asked.say code bluesave thisCode Purpleone shotno agentsTurns the work into one finished, polished md, html, chat or email piece.say code purplemake it prettyCode Pinkmodeno agentsSame message, fewer words. Stays on until you say pink off.say code pinkconciseCode Elephantone shotno agentsClaude takes a screenshot of your screen and looks.say code elephantcode elephant maincode elephant all

Try these once

Four codes nobody has said yet. Tick one off after you try it.

Say code grey before your next deploy or one off script.Not used yet. One critic, under 250 words, about a minute, and nothing runs until you say ok.
Say code orange the next time you want the web combed.Not used yet. It names the agent count first and uses the cheap model, instead of a turbo black fan out.
Say code white when you switch topics.Not used yet. Long sessions were the costliest zone: 62% of turns ran past 150k of context.
Say code elephant when something on your side screen needs a look.New Sep 27. Or press Win+PrtScn and it shows up with your next prompt.

Section two

Which one do I want?

Each code turns one dial. The first stop is the least, the last is the most.

RigorHow hard Claude proves things

code greyone critic, one verdict code blackplan, two sources, critic, evidence turboup to 4 researchers and 1 critic ultracodea Workflow, agent count said first

ResearchHow wide Claude searches

code orange3 cheap researchers, one ranked answer as many as possiblemore than 4, the number said first

SavingWhat gets written down

code whitememory and handoff, then says clear or compact code bluememory, handoff, commit and push

PolishHow finished the result looks

code purplea polished md, html, chat or email piece make this a pagea Hub page built to the standard

LengthHow long the reply is

code pinkfewer words until pink off one linea hard cap on a single reply

EyesWhat Claude can see

code elephantyour side screen mainthe ultrawide allthe whole desktop

Section three

Each code in full

What it does, every phrase that fires it, what it never does and what it spends. Tap a row.

Code BlackStays on until the task is genuinely done, not a one time step.moderigor (evidence burden)

Does

Rigor mode: plan first, prove every load bearing claim from two angles and a primary source, gate anything irreversible, run one bounded critic, and paste evidence before saying done.

Use it when

Being wrong is expensive: money, legal, a lease or tenant, taxes, a live site, family medical, or anything you will repeat to another person.

Say

code black be rigorous this matters verify everything don't guess triple check turbo code black code black turbo code black super turbo

On its own

When the stakes are plainly real even if you don't say it: money, legal, lease/tenant, taxes, live production site, family medical, or anything you will repeat to another person.

Versions

code black (plain) say code black0 or 1 researcher + 1 critic at the end. Research inline first; spawn only when isolation pays.
code black turbo say turbo spawn agentsUp to 4 researcher + 1 critic. Declare "N agents (model): why" BEFORE spawning.
ultracode say ultracode super turboWorkflow tool allowed; state the expected agent count up front. The only tier allowed to use the Workflow tool.
neutral rigor stance (default)Agents judge; the output is whatever is true. For "is this right", "what does this cost", "should I do X".
advocate stance say get this approved push back write the dispute my lawyerAgents play the other side first to rehearse their rebuttals. Cite checking gets stronger, not weaker. Claude says the switch out loud once. Also fires on any draft going to an HOA, insurer, contractor, permit office or opposing party.

The rules

Plan first (no plan, no writes)
Two independent angles on every load bearing question
Primary sources beat summaries; a price is load bearing (quote-range first)
Bounded critic, once
Gate irreversible work; one record first
E2E and visual proof: run it, fetch the PLAIN url, read the screenshot
Never say "I cannot" without checking the toolbox
Iron Law and honesty log (VERIFIED / UNVERIFIED / NOT CHECKED)
Durability: findings to disk early

Plus a 13 item ship checklist it copies into the reply and ticks off.

Never

Never claims done without fresh evidence pasted in the transcript (the Iron Law).
Never lets these words out without a recheck: "should", "probably", "seems to", "appears to", "Done!".
Never runs a parallel persona panel. One bounded critic, once.
Never takes a destructive or outward facing step without a gate. One record first before any bulk write; if it surprises, stop.
Never says "I cannot" without checking the toolbox.
Never spawns an agent without first stating agent, model and why.
Never trusts research time numbers at send time: anything going to a third party is rechecked when it is sent.

Not for

A quick sanity check on one change: use code grey
Formatting a deliverable: use code purple
Saving or committing: use code blue
Broad research with no decision attached: use code orange

Spends

0 or 1 researcher (sonnet) + 1 critic (opus) at the endResearch inline first; spawn only when isolation pays. The main thread keeps the judgment.

Born

before Jul 11, 2026. Made to kill half answers, unverified claims and premature "it's fixed". The Sep 12, 2026 v2 rewrite added spend rules because black windows were 56 to 71% of all output tokens; rigor is not fan out.

Used

Loaded 44 times before the Sep 12 rollout, 12 times since; you said it 135 times before and 46 since; its share of all output fell from 60.2% to 16.2%.

Source

~\.claude\skills\code-black\SKILL.md

Code GreyOne review, one verdict, then it hands back.one shotrigor, lite (evidence burden)

Does

One critic reads the change and returns SHIP, MODIFY or RECONSIDER in under 250 words. It never runs the change itself.

Use it when

A single file change, one hub page edit before deploy, a one shot script about to run, a dispatcher job add or edit, or "sanity check before I send this". If the worst case is "revert one commit and re-run", it's grey.

Say

code grey sanity check this sanity check before I run this before I send this catch anything dumb look this over look this over before I run/send it

Grey or black

Grey

Single-file write change
One hub page edit before deploy
One-shot script about to run
Dispatcher job add/edit (sb_dispatcher.py JOBS)
"Sanity check before I send this"

Black

Live-site deploy beyond one page (WP/Kinsta/CF)
Klaviyo/email send to real people
Tenant/lease, tax, or family-medical doc going out
Bulk write to WP/Woo/CF KV/Klaviyo
Anything with no rollback

Neither

Read-only query
Additive (new file, new section)
Doc edit
Dev-only / already reviewed
Boilerplate

Verdicts

SHIP MODIFY RECONSIDERSHIP waits for "ok, do it". MODIFY asks whether to apply the fix and re-review. RECONSIDER offers to escalate to code black.

Prints

CODE GREY REVIEW

Never

Never executes the change. SHIP waits for your "ok, do it".
Never uses more than one reviewer. If it needs more eyes it escalates to black instead of growing.
No implementation blueprint and no downstream audit; those are black's jobs.
Never goes over 250 words. If it will not fit: "this looks bigger than grey, want a code black review?"

Not for

Irreversible or multi system work: use code black
Formatting a deliverable: use code purple
Saving or committing: use code blue
Read only questions

Spends

1 critic (opus)Falls back to general-purpose on sonnet if the critic agent is missing. Target under 1 minute.

Born

Sep 12, 2026. Grey sized changes were getting either zero review or a five agent review that cost 50x what the change did. Before Sep 12 grey had 0 uses on the tower because it only existed on the work side.

Used

Loaded 0 times since it arrived Sep 12 (the goal is at least once).

Source

~\.claude\skills\code-grey\SKILL.md

Code OrangeOne research sweep with the agent count said out loud first.one shotbudget (research breadth)

Does

Wide net research on cheap models: states the agent budget first, fans out one researcher per angle, then gives one ranked answer with a source table and a verified vs unverified split.

Use it when

You want breadth first research with no decision riding on it.

Say

code orange thoroughly research deep dive the web find as much as you can research everything about go through the web and find as much as you can search the web and spawn agents

Versions

stated N say as many as possibleRaises N above 4; the number is still said before any spawn.

You get

ANSWER: ranked findings, most decision relevant first
SOURCE TABLE: url | supports | date | independent-of-others?
VERIFIED vs UNVERIFIED vs NOT CHECKED
One line: what would change the answer

Never

Never spawns before the budget line: "Budget: N researcher agents (sonnet) + inline synthesis."
Never one agent per website; one agent per ANGLE.
Never pastes raw researcher output; summaries only.
No synthesis agent; the merge happens inline.
Never decides about your money or systems: it says "this is black, not orange" and switches.

Not for

Decisions about your own systems or anything with stakes: use code black
A one fact lookup (just answer it)
Reviewing a change: use code grey

Spends

default 3 researchers (sonnet), synthesis inlineN=1 for a narrow question. More than 4 only if you asked for "as many as possible", and even then the number is said.

Born

Sep 12, 2026. "Thoroughly/exhaustively research" was the single most frequent uncoded ask (96 turns), usually served by unbudgeted code black turbo fan outs at 22x the cost of a main thread turn.

Used

Loaded 0 times since it arrived Sep 12 (the goal is at least once).

Source

~\.claude\skills\code-orange\SKILL.md

Code WhiteSaves this topic, prints where it stands, tells you what to press next.one shotpersistence

Does

Writes the current topic's state to memory and the handoff, prints a short where we stand block, then tells you whether to /clear (new topic) or /compact (same topic, long tail).

Use it when

You are switching topics, or a long session is about to wander into something new.

Say

code white switching topics checkpoint park this wrap this topic new topic checkpoint this

On its own

Claude proposes it unprompted when the session is deep into its context and the topic is changing.

Versions

recommends /clearWhen the next work is a different topic.
recommends /compact <focus>When the same topic continues but the tail is long.

Prints

CODE WHITE CHECKPOINT (Done / Open / Pointers / Recommend)

Never

Never runs /clear or /compact itself; they are your call.
No new research, no fixes, no polish. Anything broken goes under Open, not chased.
Never leaves anything existing only in the conversation. Disk first, always.
Keeps the whole output under about 25 lines.

Not for

An ordinary progress question on an active task
Saving or pushing code: use code blue
Producing a document: use code purple

Spends

0 agents

Born

Sep 12, 2026. 62% of main thread turns ran above 150k context and 30% above 400k, the measured worst efficiency zone; 20 of 24 sessions never compacted, they just grew to the 1M ceiling.

Used

Loaded 0 times since it arrived Sep 12 (the goal is at least once).

Source

~\.claude\skills\code-white\SKILL.md

Code BlueRuns every step in order with no questions, then reports.one shotpersistence

Does

Save and ship: updates memory and the handoff, checks nothing secret is staged, commits every touched local repo, and pushes where a remote exists.

Use it when

You are wrapping up a work session and want everything saved and committed.

Say

code blue save this wrap up commit and push ship it

Prints

CODE BLUE COMPLETE

Steps

Review what changed this session
Update memory and handoff files
Commit each touched repo after the secrets check
Push only where a remote exists
Verify and report

Never

Never asks questions; it just executes.
Never commits SSNs, .env files, credentials or evidence binaries.
Never leaves a token in a remote URL (inject, push, strip in the same block).
Never pushes to an account that is not yours.
Never pings Discord unless you asked for a notification.
Does not clear the session either; code white only checkpoints and tells you which one to press.

Not for

Checkpointing context without shipping code: use code white
A polished document: use code purple
Reviewing a change: use code grey

Spends

0 agentsIf nothing changed on disk it still updates memory; that IS the ship.

Born

Jul 26, 2026. The session may be ending; durability is the whole point.

Used

Loaded 9 times before the Sep 12 rollout, 4 times since; you said it 28 times before and 12 since.

Source

~\.claude\skills\code-blue\SKILL.md

Code PurpleBuilds one finished, formatted piece and reports where it is.one shotoutput polish

Does

Turns the work into a polished deliverable (md, html, chat paste or email) with TL;DR first, aligned tables, and every bell and whistle.

Use it when

You want something formatted to read, keep or hand to someone.

Say

code purple make it pretty format it clean this up write me up make this a page make this pretty write me up X

Versions

html format say make this a pageLoads hub-page-standard and follows it exactly.

Prints

CODE PURPLE COMPLETE

Asks first

Format: md, html, chat, email. Audience: self, family, external (default family).

Always adds

Hyperlink nearly everything
Hover tooltips on jargon, metaphor first
Sticky TOC for 3+ sections
Back to top on anything longer than two screens
Copy buttons on anything meant to be pasted
Cross links between deliverables, and a Hub card added locally; it goes live only when you ask for a deploy
Canonical UPDATED datetime pill

Never

Never publishes or deploys (CF Pages, artifacts) without being asked; writing the file is the deliverable.
Never emails anyone but your own addresses without an explicit named recipient.
No dash punctuation anywhere in deliverables: no em dashes, no en dashes, no spaced hyphens (your global voice rule).
Never calls html done without a headless render and a screenshot actually read.
Pretty never means verbose; the deep dive goes in a second file.

Not for

Verifying or reviewing a change: use code grey
Verifying or reviewing a change with stakes: use code black
Saving or committing: use code blue
A plain chat answer that needs no formatting

Spends

0 agents

Born

Jul 26, 2026. You asked for the same finishing touches over and over, so they became non negotiable: a deliverable without them is a draft.

Used

Loaded 5 times before the Sep 12 rollout, 2 times since; you said it 21 times before and 3 since.

Source

~\.claude\skills\code-purple\SKILL.md

Code PinkStays on for the rest of the task until "pink off".modeoutput length

Does

Concise mode: answer first, cut every sentence the reader would not miss, nothing they already know. It changes length only; rigor and any other code keep running and just report short.

Use it when

You want the same message in fewer words, in replies or in drafts for other people.

Say

code pink concise just the answer short version tl;dr less words pink the draft pink off

Cut test

If this sentence disappeared, would the reader lose a decision, a fact they need, or an action?

Dials

`code pink`	concise mode for the rest of the task
`code pink: N lines / one line / one sentence`	hard cap on the reply
`pink the draft / pink that`	trim one specific piece, mode stays as it was
`bullets only / table only / TL;DR only`	shape constraint for that reply
`pink off / full detail / walk me through it`	lift the mode
`no metaphor / plain`	skip the metaphor for that reply
`explain it like the water tower`	a full extended metaphor (several sentences) is allowed for that reply

Metaphors

One metaphor of about 12 words folded into the answer line, then .. and the precise fact. It replaces a sentence of jargon, never adds one, and must survive a one breath break test.

Always on

Any reply longer than one screen ends with a Code pink summary block: 3 to 5 lines with the outcome, what you must decide, and what is still running. Your direct questions get answered there too. This one is always on, pink mode or not. Sep 27, 2026: answers in the middle of a long report got buried and you had to ask twice.

Never

Never reduces rigor; code black still verifies.
Never drops a required part of a deliverable; purple keeps its sections.
Never hides an uncertainty: "Unverified: ..." still gets its one line.
No preamble, no recap of what you said, no closing offer, no options you did not ask for.
Never a second metaphor to rescue the first.

Not for

Skipping verification (code black still verifies)
Dropping required parts of a deliverable (code purple keeps its sections)

Spends

0 agents

Born

Sep 17, 2026, 9:54 PM Denver. You trimmed Claude's reply to a recruiter's "please confirm receipt" email. You cut "confirmed and received, thank you" (replying IS the acknowledgement) and turned "Friday, Sep 18" into "tomorrow" (she sent the invite, she knows the date).

Used

Loaded 4 times since it arrived Sep 17.

Source

~\.claude\skills\code-pink\SKILL.md

Code ElephantSay the phrase and Claude takes a screenshot and looks. Separately, any Win+PrtScn you take shows up on your next prompt by itself.one shoteyes (screen input)

Does

Claude screenshots your side screen (or the ultrawide, or the whole desktop) and reads it before answering. Win+PrtScn shots and files dropped in the Claude inbox surface on their own on your next prompt.

Use it when

You want Claude to see what is on your screen without drag and drop.

Say

code elephant code elephant main code elephant all look at my screen look at monitor 2 screenshot my side screen look at my screen / monitor 2 / what I have up

Versions

code elephant say code elephantThe non primary 2560x1440 side screen.
code elephant main say code elephant mainThe primary 5120x1440 ultrawide (where the terminal lives).
code elephant all say code elephant allThe whole 7680x1440 virtual desktop.

By itself

Win+PrtScn (whole desktop, no dialog)	a monitor2 crop of the side screen; the full original is kept and named in one extra line
Snagit / Win+Shift+S then save / any file	the file itself (tiled if wider than 2560, JPEG copy if over 3 MB)

At most 10 files per prompt; the rest follow. Files younger than 2 seconds wait for the next prompt. With several sessions open, the first session to get a prompt claims the shot.

Off switch

Everything off now: create ~\claude-kit\hooks\DISABLED (delete it to re-arm).
Just the watcher: remove its UserPromptSubmit entry from claude-kit\settings.snippets.json, then run python sync.py hooks --apply.

Never

Never captures while a .env, password manager, banking, tax or family medical window is visible; when unsure it asks one line first.
Never claims to have seen something it did not Read.
Never deletes inbox files; you clear Pictures\ClaudeInbox\read by hand.
Never blocks your prompt: the hook fails open and stays silent on any error.
Runs on this Windows tower only.

Not for

Capturing while a .env, banking or family medical window is up
A region (Win+Shift+S then Alt+V is better)

Spends

0 agents

Born

Sep 27, 2026. You cannot drag an image into the Claude Code CLI, so this gives Claude its own eyes on your screens. A race test the same day found two sessions could both claim one screenshot, so claiming now holds an OS lock.

Used

Too new to appear in the Sep 27, 2026, 3:35 PM scorecard.

Source

~\.claude\skills\code-elephant\SKILL.md

Section four

Codes together

Say several codes in one message and Claude runs the look only ones first and anything you cannot undo (a commit, a push, a send, a deploy) last, then gives you one reply. Example: code white and code blue together means the checkpoint notes are written first, then the commit and push.

Easy to mix up

This	Is not this
`wrap up` code bluecommits and pushes	`wrap this topic` code whitewrites a checkpoint, commits nothing
`spawn agents` code blacka black turbo run	`search the web and spawn agents` code orangean orange sweep; if it is unclear, Claude asks one question first
`code black turbo` code blackup to 4 researchers and 1 critic	`ultracode` code blackthe Workflow tool, many agents, count said up front
`code grey` code greyone change where the worst case is a revert	`code black` code blackanything with no rollback or real stakes

Pink with the others

With	What pink changes, and what it leaves alone
code black	Changes the report (one line verdicts, checklist collapsed to "all checked" or just the failures). Does NOT change the plan, the two angles, the critic, the E2E run or the honesty log on disk.
code purple	Prose inside sections is minimal, tables over sentences. Required sections, UPDATED pill, hyperlinks and tooltips stay.
code orange	The answer block is the ranked table plus one line. The source table stays.
code white	Nothing; the block is already fixed format.
code blue	Nothing; the block is already fixed format.
code grey	Nothing; already capped at 250 words.

How the rest fit

Black may call orange as its research phase; orange switches to black the moment a real decision on your money or systems is riding on it.
Grey escalates to black instead of adding reviewers.
Blue ships; white recommends /clear or /compact, you run it. Blue never clears the session.
Purple html defers to hub-page-standard.
quote-range runs inside black, turbo, orange and ultracode passes whenever a dollar figure is involved, and the price lookup goes first.

Section five

Helpers and keys

The codes spend these four helper agents. Cheap tiers: scout (haiku) find, researcher (sonnet) gather, verifier (sonnet) exercise, critic (opus) judge.

Agent	Model	Job	Turns	What it does
scout	haiku	find	10	Cheap read only finder: where does X live, grep big trees, inventory sweeps. Locates and quotes, never recommends.
researcher	sonnet	gather	25	Research worker for orange sweeps and turbo fan outs. One bounded question, two sources for anything load bearing, distilled summary.
verifier	sonnet	exercise	15	Runs the thing for real (script, headless render and screenshot, plain live URL) and returns evidence. Never fixes anything.
critic	opus	judge	20	Bounded reviewer: the single grey review and black's final pass. Flags only what affects correctness; "no blocking issues" is a valid answer.

Inside a Workflow, the tier agent types inherit these turn caps and can die before returning on big questions; use the default agent type with a model override there (Sep 27, 2026 lesson).

Keys and switches

Keys	Does	Part of
Win+PrtScn	Saves the whole desktop; a side screen crop surfaces to Claude on your next prompt.	code-elephant (automatic half)
Win+Shift+S then Alt+V	Drag a region, then paste it straight into the Claude Code prompt. Nothing saved to disk; best for a small region or anything sensitive nearby.	code-elephant (zero build fallback)
Save or drop a file in Pictures\ClaudeInbox\unread	The file surfaces on your next prompt (Snagit, saved Win+Shift+S shots, PDFs).	code-elephant (automatic half)
"you should have my screenshot"	If Claude did not get it, another session claimed it: look in Pictures\ClaudeInbox\read (newest) or the code_elephant log.	code-elephant
/clear or /compact	Code white recommends one; you press it.	code-white
Create claude-kit\hooks\DISABLED	Panic button: every kit hook goes quiet on the next prompt. Delete the file to re-arm.	claude-kit (all hooks)

Section six

How they came to be

Jul 11, 2026

Earliest recorded code black (turbo) use on the tower

Jul 17, 2026

Earliest recorded code blue (FireSim repo)

Jul 20, 2026

"code black super turbo" ships the Bible Atlas and Lineage pages

Jul 26, 2026

Code blue and code purple adapted for the personal machine

Jul 27, 2026

Codes become global skills after a session said code black was not a protocol it had

Sep 5, 2026

hub-page-standard established

Sep 12, 2026

v2 rollout: black rewritten; grey, orange, white added; four agent tiers; hooks

Sep 17, 2026

Code pink born from the recruiter email trim

Sep 21, 2026

Black adopts the advocate stance and send time cite check

Sep 24, 2026

quote-range created; black makes price the first research task

Sep 27, 2026

Code elephant v1; metaphor rule added to pink and black

Sep 27, 2026

Bottom summary rule: every long reply ends with a code pink summary

Usage before and after the Sep 12 rolloutFrom the evidence scorecard.

Skill loads in 23 sessions before the Sep 12, 2026 rollout and 19 after, counted Sep 27, 2026, 3:35 PM Denver.

Code	Before	After
code black	44	12
code blue	9	4
code purple	5	2
code pink	did not exist yet	4
hub-page-standard	2	4
quote-range	did not exist yet	1
code grey	did not exist yet	0
code white	did not exist yet	0
code orange	did not exist yet	0

Code black windows went from 60.2% of all output to 16.2%, and subagents from 21.1% to 5.4%. Rigor stayed; the fan out shrank.

Everything else

The fine print

Rules that are always on (9)They apply whichever code is running, or none.

Saying the phrase is what fires the code. No other setup. ~\.claude\CLAUDE.md
Say several codes in one message and the look only ones run first, anything you cannot undo runs last, and you get one reply. ~\.claude\CLAUDE.md
If the code word is unclear or two codes fight, Claude asks one short question before spending any agents. ~\.claude\CLAUDE.md
Claude checks its toolbox before ever saying it cannot do something on the tower. ~\.claude\CLAUDE.md
Every helper agent is announced with its name, its model and one line of why. ~\.claude\CLAUDE.md
When the topic changes, Claude offers a code white on its own. ~\.claude\CLAUDE.md
Every code says which other codes it is not, so two codes do not fire on the same ask. ~\claude-kit\README.md (The color wheel)
Each color turns one dial: how hard to prove things, how many agents, what gets saved, how polished, how long. Elephant (Sep 27) adds a sixth thing the README does not list yet: eyes on the screen. ~\claude-kit\README.md (The color wheel)
Any long reply ends with a 3 to 5 line code pink summary, and your direct questions get answered there. ~\.claude\CLAUDE.md and ~\.claude\skills\code-pink\SKILL.md (Bottom summary)

Other skills you builtNot codes, but codes lean on them.

Quote Range (Sep 24, 2026): Honest cost estimates: price lookup first, a gross P10 / P50 / P90 range with a source on each, gross before net, rebates at the floor, and a ledger that grades every estimate when the real quote lands. A Sep 24, 2026 postmortem found bottom up builds ran about 1.9x low on the main job every time. Dollar figures, the ledger and your caps are kept out of the skill folder on purpose; they live in local memory only.
Hub page standard (Sep 5, 2026): One layout, one palette, one theme key (hub-theme), one UPDATED pill, one gate, one deploy path, one verification script for every Hub page. Code purple defers to it for html. Established from the Supplement Stack build after you said every page looked different and was hard to follow. The Bible app keeps its own colors; everything else conforms. Tower only; excluded from the kit sync on purpose.
Synced Anthropic skills: Anthropic provided skills that sync into ~/.claude/skills/synced. They are not keyword codes. docs, docx, google-workspace, import-memory, morning, pdf, pptx, skill-creator, xlsx.

Loose ends in the sources (11)Where two files disagree. None of them breaks a code.

medium CLAUDE.md, code-black skillCLAUDE.md says: code black turbo / ultracode ... same, with a declared fan-out (cap 4 researchers + 1 critic; Workflow only on "ultracode")The code-black skill says: turbo = up to 4 researcher + 1 critic; ultracode = Workflow tool allowed, state the expected agent count up front (no cap of 4)Why it matters: One CLAUDE.md row covers both tiers, so it reads as if ultracode is also capped at 4. The claude-kit README row does the same ("stated N (cap 4)").
medium code-black skill, memory: home_server.md, sickbeard.mdCLAUDE.md says: Workflow only on "ultracode"The code-black skill says: The same thing.What you actually did: Sep 13, 2026: a 13 agent "code black turbo Workflow"; Sep 23, 2026: "turbo code black review" run as a Workflow with 8 sonnet reviewers. Practice treats turbo as allowed to use Workflow and to exceed 4.Why it matters: The written rule and your own practice disagree; decide which one wins, then fix the other.
low code-black skill, code-orange skillThe files say: code-black lists "spawn agents" as a turbo trigger; code-orange lists "search the web and spawn agents" as an orange triggerWhy it matters: Same words can fire two codes. The ambiguous code word rule (ask one question) is the tie breaker.
low code-blue skill, code-white skillThe files say: "wrap up" fires code blue (commit and push); "wrap this topic" fires code white (checkpoint, no commit)Why it matters: Near identical phrases do different things. Say "wrap up" only when you want a commit and push.
low code-blue skillThe files say: The description says blue runs outside two named repo clones, while Step 1 and Step 4 of the same file list one of those clones as a commit and push target.Why it matters: An inconsistency inside the code-blue skill itself, not a CLAUDE.md conflict.
low Claude memory: code_black_keyword.md, Claude memory: MEMORY.mdThe files say: code_black_keyword.md's description lists six codes (black/grey/orange/white/blue/purple); MEMORY.md lists eight (adds pink, elephant).Why it matters: The memory pointer undercounts the codes.
low README.md, CLAUDE.mdCLAUDE.md says: nine rows including code elephantREADME.md says: README color wheel has no elephant row, and its layout block says the kit holds hub-page-standard, which the kit excludes on purpose (commit 0ab1c81).Why it matters: The README is the only place the "five dimensions" idea lives, and it is one code behind.
low CLAUDE.mdCLAUDE.md says: Updated 2026-09-12The same file shows: Rows for pink (Sep 17 and 18) and elephant (Sep 27) and the Sep 20 env line were added after that date.Why it matters: Header date is stale.
info CLAUDE.md, every code SKILL.md when_to_useCLAUDE.md says: Two or three headline phrases per codeEvery code SKILL.md when_to_use says: Four to eight phrases per code (full lists under each code's triggers.skill_when_to_use). "super turbo" and "spawn agents" appear only in the code-black body table.Why it matters: Not a conflict. The Say line under each code shows the full skill lists, since those are what Claude matches on.
info code-pink skillThe files say: Pink's With the other codes table covers black, purple, orange, white, blue and grey but not elephant (added later).Why it matters: Gap only.
low code-purple skill, CLAUDE.mdCLAUDE.md says: Zero dash punctuation: no em or en dashes, no spaced hyphens, no double hyphens.The code-purple skill says: code-purple's style line bans only em dashes.Why it matters: A deliverable could pass purple's own check with an en dash or a spaced hyphen. Your voice rule is the stricter one; the skill line could say all dash punctuation.

How this page stays currentOne data file, one build, one drift check.

Everything above comes from ~\Documents\Claude\Projects\Codes\codes.json, built from the skill files. When a skill changes, python _toolkit\check_drift.py in that folder names the file that moved and any new code, then python _toolkit\build_codes.py rebuilds the page.

Sources (31)Every file the data was read from.

~\.claude\CLAUDE.md
~\.claude\skills\code-black\SKILL.md
~\.claude\skills\code-grey\SKILL.md
~\.claude\skills\code-orange\SKILL.md
~\.claude\skills\code-white\SKILL.md
~\.claude\skills\code-blue\SKILL.md
~\.claude\skills\code-purple\SKILL.md
~\.claude\skills\code-pink\SKILL.md
~\.claude\skills\code-elephant\SKILL.md
~\.claude\skills\quote-range\SKILL.md
~\.claude\skills\hub-page-standard\SKILL.md
~\.claude\agents\critic.md
~\.claude\agents\researcher.md
~\.claude\agents\verifier.md
~\.claude\agents\scout.md
~\claude-kit\README.md
~\claude-kit\CODE_SYNC_PROTOCOL.md
~\claude-kit\REVERT.md
~\claude-kit\hooks\code_elephant_watch.py
~\claude-kit\hooks\code_elephant_capture.py
~\claude-kit\hooks\local\code_elephant.json
~\claude-kit\metrics\scorecard.md
Claude memory: MEMORY.md
Claude memory: code_black_keyword.md
Claude memory: feedback_code_black_is_full_diligence.md
Claude memory: feedback_quote_ranges.md
Claude memory: feedback_use_metaphors.md
Claude memory: feedback_workflow_agent_caps.md
Claude memory: evidence_scorecard.md
Claude memory: personal_operating_rules.md
~\.claude\CLAUDE.md (bottom summary rule, read Sep 27, 2026, by 7:04 PM Denver)