What your instruction files are actually telling me to do
Private deliverable. Enter the hub password to continue.
That's not right. Try again.
What your instruction files are actually telling me to do
Anthropic's Fable 5.2 guide has a section called "Keep caution from piling up." This is what had piled up.
You said yes to job 3 from the Fable 5.2 page: audit the rules themselves. Three
passes ran over your trio (claude.md, annette.md, soul.md), the 58 files in
~/memory/topics/, and 154 CLAUDE.md and AGENTS.md files across your projects.
Nothing was changed. Every finding below names the file and the line, and the project-file and reference checks were verified against the disk rather than assumed.
The short version: the rules you have written are good. The problem is the ones you replaced. A deleted rule does not disappear, it just stops being in the file you edited, and it keeps giving orders from the seven files you did not.
Start here. These five cost you something every week
1. Your API keys are in a repo, and the repo is pushed
topics/api-keys-services.md and topics/cloudflare-deployment.md hold live keys
in plain text: Anthropic, OpenAI, two OpenRouter keys, the Cloudflare global key,
the Workers and Registrar tokens, and a database password. ~/memory is a git repo
that pushes to annettethompson/charlie-memory, both files are tracked, and they
appear in 14 commits of history.
I checked before telling you: that repo is private. So this is contained to
your own GitHub account rather than out in the open, and it is not an emergency.
It is the same shape as the YouTubeIntelligenceFeed key that has been waiting
since July.
What fixes it, in your own rules' terms: the keys already live in Windows
environment variables, so the memory files should name the variable and nothing
else. Move the values to a gitignored .creds/ file, replace them in the two
topic files with the variable names, and rotate the Cloudflare global key and the
two OpenRouter keys, which are the ones with real reach. Say the word and I will
do the move and hand you the rotation list.
2. Seven projects still forbid mentioning your separation
You un-redacted the marriage ending on 2026-07-09, and your own public wording is "my marriage of 34 years ended." These files never got the memo:
DivorceRecovery/CLAUDE.md, lines 37 and 48, which is the divorce recovery bookNotTooOldForAI/CLAUDE.md:86AnnetteJobSearch/CLAUDE.md:43BoneVoyageRevival/CLAUDE.md, lines 42, 112 and 220AjijicRelocationGuide/CLAUDE.md:87Vitalency/CLAUDE.md:166BobsHealth/CLAUDE.md:17- plus the worktree copies under three of those projects
A session writing your divorce-recovery book is being told it may not mention your divorce. The fix is to delete the separation clause from each and keep only the redaction that is still real: other people's health information and confidential-legal matters. The "don't name the kids" wording is separate and should stay.
3. One project would put handoffs back on the hub
MenopauseKnowledgeBase/CLAUDE.md:13 says handoffs "DO get published to the hub"
and that the hub URL is what you get. That is the 2026-08-01 order you reversed on
2026-08-08, after it buried your real deliverables under 52 handoff pages that then
had to be deleted. topics/menopause-empire-project-map.md:85 says the same thing
and calls the correct rule dead.
Two files, and the next session that reads either one starts the mess again.
4. The Chrome rule disagrees with itself
annette.md lines 228 to 230 say connecting to your real Chrome over the debugging
port is dead: Chrome ignores the flag against your default profile and a copied
profile comes up logged out, so logged-in work goes through the Chrome extension.
claude.md says the opposite, in so many words: connect to your real Chrome over
CDP. topics/browser-automation.md is named after the dead pattern, and your
auto-memory repeats it.
This fires on every task that needs a site you are signed in to, and the wrong branch wastes a few minutes relaunching Chrome before failing.
5. The variable that bills you is set right now
Your 2026-10-20 audit deliberately stored the Anthropic key as
ANTHROPIC_DIRECT_API_KEY, because headless claude -p runs bill to
ANTHROPIC_API_KEY whenever that variable exists, instead of to your Max plan.
ANTHROPIC_API_KEY exists in your Windows environment, non-empty, sitting right
next to the other one. It is not loaded in this session's shell, so a fresh shell
is where it would take effect. Every headless run started from a new terminal
since it came back has had the chance to bill your API credits rather than the
plan you already pay for.
topics/llm-routing.md:33 is probably why it came back: it names
ANTHROPIC_API_KEY as the variable to set, and lines 154 and 156 repeat it in
sample code.
Two fixes, and the second one is yours to approve because it changes your
environment rather than a file: correct the three lines in llm-routing.md, and
remove ANTHROPIC_API_KEY from your user environment so only the DIRECT one
remains. Say the word on the second and I will do it, no typing for you.
Rules you deleted that are still giving orders
Grouped by rule. Every one of these was overturned by a decision in claude.md.
| The dead rule | Where it still lives | What it does if followed |
|---|---|---|
| The separation is redacted | 7 project files plus worktrees | Silences your own public story, see above |
| Handoffs go to the hub | MenopauseKnowledgeBase/CLAUDE.md:13, topics/menopause-empire-project-map.md:85 |
Rebuilds the 52-page mess |
| Use WebFetch | topics/web-deployment-rules.md:11, topics/web-search-routing.md:35 |
Sends me to a tool that failed 8 of 8 times |
ANTHROPIC_API_KEY |
topics/llm-routing.md:33, :154, :156 |
Bills headless runs to credits |
The validator has a darkbg check, use skipChecks |
topics/web-deployment-rules.md:148, topics/visual-taste-profile.md:286 |
Adds a skip for a check that no longer exists |
| Never white text on dark, plus sticky bars | topics/web-deployment-rules.md:157-158 |
Both of your deleted design rules in one retired block |
| No dark backgrounds, pink accent | TrainerBooked/AGENTS.md:27, CRM worktree, 2 TrainerBooked worktrees |
Contradicts the same project's own CLAUDE.md |
| Resend is the email rail | BoneVoyageRevival/CLAUDE.md:58, :75, :184, SocialMediaPlanner:41, NutriForge:58, topics/project-log.md:105, ROUTINES.md:213 |
Builds on the rail SES replaced in August |
| D1 needs the Cloudflare global key, error 10000 | MensGroups:45, Postiz:44, topics/api-keys-services.md:109, :154, topics/cloudflare-deployment.md:10, topics/tooling-gotchas.md:357, :766 |
Makes me hunt for a key file, then report a permissions block you fixed in August |
| File it in Master-TODO.md | RecipeMemoir:91, RecipeMemoir-auditfix:91, AjijicRelocationGuide:66, SmartStrongAlive/coaching:75, :77, :122, plus 3 memory files |
Writes to a file that does not exist |
| Brain OS session start | topics/brain-os-rules.md:5, :9, OWNER_PREFERENCES.md:14 |
A live session-start instruction for a system retired in July |
| The EA startup routine | topics/ea-startup-routine.md:3-29 |
Marked disabled, still written as binding, including "ask Annette which lane" |
| Llama 70B or DeepSeek V3.1 as the default worker | AIMusicCreation:151-154, FBGroupIntel:147-151, YouTubeIntelligenceFeed:68-70, Vitalency:165, :167, topics/llm-routing.md:46-62 |
Routes work to models you replaced in July |
| No sticky, no exceptions | MenopausePracticeSites:66, MedSpasBoulder:67, BobExecCounseling:82, TrainerBooked:27 |
Would flag your own beta feedback widget as a violation |
| Node 22 LTS | topics/dev-tools.md:9 |
Wrong version, 24.15.0 is installed |
Three more that are marked retired but whose bodies still read as live instructions:
feedback_master-todo-pinned-top-section.md (lines 9 to 24),
feedback_bang-prefix-is-claude-code-only.md, and
topics/desktop-files.md (lines 17 to 35). A banner at the top does not stop a
model reading the rule underneath it.
Where two files tell me opposite things
- The review panel.
annette.md:234-240says you trimmed it to three LLMs on 2026-07-30 and that it retires the Gemini, Codex and Grok trio, with the exact three pending your confirmation.claude.mdstill says run all three.topics/adversarial-review-rules.md:7says retired, and its own body at lines 9 to 17 says "never run fewer than all three."OWNER_PREFERENCES.md:38says all three too. This one needs you: name the three models and I will make every file agree. - Which model orchestrates.
claude.mdsays Opus 4.8 at high effort, your order of 2026-07-13.annette.md:234says Opus 5 coordinates.topics/annette-tooling-inventory.md:35says Opus 5 in one line andclaude-opus-4-8three lines later. All of it predates Opus 5.5 and Fable 5.2. - Which Opus the spend rule covers.
soul.md:54saysclaude-opus-4-8,claude.mdsaysclaude-opus-4-7. It is a money rule, so it should say one thing. - Fetching pages.
claude.mdsays never WebFetch.topics/web-search-routing.md:35says use WebFetch first. - Who runs commands.
claude.mdsays I run them, never you.topics/aws-agent-rules.md:69says wait for you to be at the screen andtopics/tooling-gotchas.md:699says ask you to click. Those may be the genuine sign-in exceptions, but they are not written as exceptions. - Localhost. Zero-tolerance rule 3 says never hand you a localhost URL.
Jarvis/CLAUDE.md:47says to tell you to openhttp://localhost:3107and hold space to talk, with the same pattern at lines 35 and 43. - Your domain count.
annette.md:218says a 16-domain Cloudflare portfolio.claude.mdand your auto-memory say 25, and email routing is live on all 25. - The budget tier.
soul.md:50and:108define the cheap tier as "Llama-tier" and gate anything above it.claude.mddefines the gate as premium models at or above $1 per million input, with newest DeepSeek as the default. Two definitions of one spend rule. - ROUTINES.md against itself. The banner says everything is stopped except the RecipeMemoir nightly job. Rows 47, 48, 49, 51 and 66 still read as live daily jobs, with no stopped flag on the row. Anyone reading by row gets it wrong.
The rules that stop work
This is the part Anthropic's guide is specifically about. Current models follow an absolute literally, so a broad "never" turns into a request for approval on ordinary work, and a gate written for one risky case fires on everything.
Observation mode never ended. MyExecAssistant/CLAUDE.md:57-59 and :166-167
declare a two-week log-only period in which I propose and never act. Phase 1 began
on 2026-05-04, so that window closed in May. The _observation/ folder has one
file in it, from July. It sits in direct conflict with your standing "run
continuously, never pause to ask," in the project you work in most.
A question with nowhere to land. The same file, line 51, says every new skill
runs DeepSeek against Opus for its first ten invocations and that I ask you "was
DeepSeek good enough?" each time. The _model-routing/ folder it writes to does
not exist, and claude.md itself calls that tracker fiction. Ten interruptions per
skill, filed nowhere.
Delegate everything. claude.md says any multi-tool task goes to a background
agent and that a chain of fast calls is not exempt.
topics/background-agent-rules.md:66 calls it the sequence trap and says the
entire task goes to an agent if it needs more than one tool call. Five project
files repeat it, two of them with "NO EXCEPTIONS FOR SEQUENCES," and your
new-project template bakes it into everything new. Anthropic's guide names standing
delegate instructions as a thing to review on this model, because subagents are
often the largest line in a session's cost and the model already delegates when it
helps. Read literally, a one-line fix waits on a background agent.
The 30-minute re-read. Stated in claude.md, annette.md:3, soul.md:69 and
topics/memory-rules.md:16, which adds "re-read both files before composing your
response." The trio is 705 lines. A clock-based reload with no trigger is a large
recurring cost, and the staleness signals you already list (an identity question,
"I already told you," public-facing copy) do the same job better.
Review gates on everything. topics/adversarial-review-rules.md:19-23 makes
the panel mandatory before any first run, at every phase completion, after more
than two bug fixes, and before any PR, with "never skip them" at line 17 and a
750-line learnings file that line 54 calls mandatory reading first.
Per-item approval that should be per-batch. BoneVoyageRevival:60 and :193
want your approval on each post. FBGroupIntel:157 makes every output a draft
forever. MyExecAssistant:38 is one sentence that lumps a pre-approved social post
in with moving money.
Sixty files say read these first. About 60 project files open with "before
responding to anything in this project, read" and then @import the same files,
so the trio loads twice.
One more worth a look: AIMusicCreation/CLAUDE.md:160 has an undated two-week
proposal period before auto-publish, so nothing can ever graduate.
Things that point at nothing
Checked on disk, not assumed.
C:\Users\annet\memory\AGENTS.mddoes not exist, and three projects@importit:MenopausePracticeGrowth/AGENTS.md:23,SmartStrongAlive/AGENTS.md:22,TrainerBooked/AGENTS.md:3. The import fails silently, so the shared agent rules those three expect have never loaded.SiteShipperis gone from ClaudeProjects.AnnetteThompson/CLAUDE.md:17still gives its manual deploy command and env path. The engine lives in thedeploy-siteskill now.MenopauseContentRepurposer/CLAUDE.md:3and its AGENTS.md point at.claude/memory/annette.md. The file is at~/memory/annette.md.SESSION-CONTEXT.mdis cited but missing inMedSpasBoulder,LambdaMOO,PhotoFinder,ExerciseBuddyFinderandMenopausePracticeGrowth.- Missing scripts and docs:
VerityAgentic/scripts/invite_client.py,MenopauseKnowledgeBase/scripts/resolve_channel_ids.py,NotTooOldForAI'sbrand.md,editorial-calendar.mdandsubstack-settings.md,chromeextensions/DECONSTRUCTION-PLAYBOOK.md, andWatchLaterSorter'srun_daily.bat, which is archived while a Task Scheduler line still calls it. - Missing projects and memory files:
LinkYieldAI,ReverseEngilneeringYouTubeChannel(also misspelled),user_charlie_voice_male.md, and two Bone Voyage memory files. AjijicRelocationGuide:54points into aclaude-obsidian-vaultfolder that is gone.
And in the trio itself:
- The Town handoff file is one folder off.
claude.md:234givesC:/Users/annet/MyExecAssistant/town-inbox.md. The real file is atC:\Users\annet\ClaudeProjects\MyExecAssistant\town-inbox.md. A session that takes the path literally reads nothing, finds no items from Plum, and tells you there are none. There is one waiting. - A skill that does not exist.
claude.md:174says to load anaws-secrets-managerskill before any secrets work, and never to callget-secret-value. There is no such skill in~/.claude/skills, in the archive, in~/.agents/skills, or in the plugin cache, and the AWS plugin is not enabled. So the instruction is "load X, and until you do, do not do the work," with no X. soul.md:85links../executive-assistant-principles.md, which from~/memory/resolves one level too high. The file is at~/memory/executive-assistant-principles.md.- The Typst brand package is not installed.
claude.md:177describes@local/verity:1.0.0as the pattern to copy. There is no local Typst package directory on this machine, noTYPST_PACKAGE_PATH, and VerityAgentic does not import it. Typst itself is installed and working, so this is a missing package, not a missing engine. - Three tools listed as global are not on PATH:
yarn,tscandtsx(claude.md:168). Node 24.15.0, Python 3.13.7,pnpm,typstandghare all exactly where that line says. active-blocks.jsonis at~/.claude/state/, not in MyExecAssistant (claude.md:158)./morninghas no local skill or command. Only the plugin version exists, so the verbatim ONE Thing question your rules attach to it is not actually wired to anything you can type.
Your project list in annette.md has drifted too. MosconiAlly (line 167) is
archived under ClaudeSupport\_archive. Of the active list at lines 198 to 211,
StressReduction, MexicoEmploymentLaw, YouTubeHiddenGemMining and
PodcastEpisodeFactory are archived, SiteShipper has moved into
ClaudeSupport\_tools and the deploy-site skill, MenopausePracticeSites lives
under a misspelled MenopauePracticeGrowth folder, and BackOfficeStack is not on
the disk at all. Line 260 names pyttsx3 and the Azure speech SDK as the TTS
engines; neither is installed, and edge-tts 7.2.8 is.
What checked out clean
Worth saying, because it is most of what was tested. All six recursive-delete deny
rules are present in settings.json and the ask list is empty, so the backstop
your 2026-08-03 order relies on is real. settings.json is BOM-free with its hooks
intact, and both the handoff-push and hub-first hooks are registered where your
rules say. Every script flag your files quote is correct, on all eight scripts.
Every environment variable they name is set. Every site URL returns 200. The
ResearchLibrary is laid out exactly as described, with the old fulltext/ name
gone.
Two housekeeping notes
53 worktree copies. There are 53 CLAUDE.md and AGENTS.md files under
.claude/worktrees and _worktrees, snapshots from before the July cleanups. They
keep every dead rule above alive for anything opened inside them. Worth pruning the
worktrees that no longer have live branches.
Your two topic indexes disagree. claude.md has one at the bottom and
annette.md has another. About 22 files in topics/ appear in neither, including
all seven book-pattern files and the six annette-* files you split out today. One
index should point at the other rather than both trying to be complete.
And 29 em dashes across 18 memory files, which is trivial to fix and worth doing because those files get quoted into prompts.
What I would do, in four passes
- The five at the top. Keys out of the repo, the separation clause out of seven projects, the handoff rule out of two files, Chrome settled one way, and the API key variable name corrected. This is the pass that changes what happens tomorrow.
- The dead rules table. Mechanical, about 40 edits, no judgment needed once you have said yes.
- The gates. Observation mode deleted, the A/B question dropped, delegation softened to "when it saves the main context," and the 30-minute clock replaced by the triggers you already wrote. This one changes how a session behaves, so it is worth doing deliberately and watching for a few days.
- The references and the worktrees. Fix the paths, prune the snapshots.
One decision is genuinely yours and blocks part of pass 2: which three models are in the review panel. Everything else I can do and show you as a diff.
How this was checked, and what it missed
Three passes ran in parallel: the trio by hand, the topics folder, and the project
files. Project-file findings and every dead reference were verified against the
disk. Two large topic files, tooling-gotchas.md and
adversarial-review-learnings.md, were searched rather than read end to end, so
their stale facts are under-sampled. Most root-level memory files were skimmed. No
file was changed, and no key value was read, quoted or copied anywhere.
Published to Annette's hub. Rebuilt from the source markdown, so edit the source and rerun rather than editing this page.