# 05 · Letter to future sessions > From Fable 5, 2026-07-03 — the one session on this tier this environment gets. > `00`–`04` and `06` are the rules. This file is the context the rules can't carry: > three things the user didn't ask me to tell you, how this system will most likely rot, > and what honestly cannot be fixed by process. --- ## Three things nobody asked, in order of importance ### 1. The harness will not stop you — so the brake has to be you `settings.json` runs with `defaultMode: "bypassPermissions"`, a blanket `permissions.allow` (including bare `Bash`), and `skipDangerousModePermissionPrompt: true`. There is **no confirmation prompt between you and `rm -rf`, `git reset --hard`, a `DROP TABLE`, or an outbound API call.** The user chose speed, knowingly. That means the confirmation bar in `02-judgment.md` Rubric 3 is not a formality here — it is the *only* safety layer that exists. Before any hard-to-reverse command, re-read the evidence that justifies *that specific* command. A signal that pattern-matches a known failure may have a different cause. ### 2. Trust the live tool list over the config — the config has debris Observed on 2026-07-03: `permissions.allow` names `mcp__serena`, `mcp__context7`, `mcp__mcp-deepwiki`, `mcp__Playwright` — but global `.mcp.json` is empty and none of them were connected in my session (only `ace-tool` and `exa` were). An *allowed* tool is not a *connected* tool. Also, `settings.json` pins `model: "claude-fable-5[1m]"`, which will be stale once this tier is gone — if sessions start on a wrong/erroring model, that line is why; suggest the user update it via `/model`. General rule: what's in your system prompt and ToolSearch results is real; `settings.json` is aspiration. ### 3. This user's leverage problem is breadth: 50+ projects, near-zero persistence `~/.claude/projects/` has 50+ slugs (API gateways, model routers, LLM tooling — `k2i-*`, `sora2api`, `llmbench`, `code-abyss`…). Almost every `memory/` dir is empty. The same classes of problem (auth flows, provider adapters, rate limiting, streaming) recur across these repos, and every session re-learns them. The single most valuable habit you can add on top of the doctrine: **when you solve something in one project that smells generic, write the memory entry** — and if it's about the harness or the user's stack rather than one repo, put it in `06-lessons.md` where every project sees it. The user works in Chinese — converse in Chinese; keep doctrine files in English (instruction-following in English is more reliable for the smaller models these files are written for). --- ## How this system will most likely degrade — and the counter for each Ranked by probability. I'd bet on #1. 1. **Read-but-not-applied.** You load the router, agree with it, then grep the repo inline anyway because the task feels urgent. The doctrine becomes decoration. *Counter:* the triggers in `01` are numeric (~50 lines of raw output) precisely so "does this rule apply" is a check, not a vibe. If you catch yourself pasting raw scan output into the conversation, that IS the trigger — stop mid-turn, delegate the rest. User-side counter: if you (the user) see walls of tool output, say the word "doctrine" — treat it as an instruction to re-read the router and comply. 2. **Drift by accretion.** Each session appends its own phrasing of an existing rule; five copies later two contradict, and a small model follows whichever it read last. *Counter:* `04`'s "read the whole file first, update in place, never add a near-duplicate" plus the compaction trigger. When you see the same rule in three places, that's not emphasis — it's rot. Merge it. 3. **Staleness inversion.** The environment changes; doctrine keeps asserting the old world; a weak model trusts the doctrine over its own observation (doctrine *sounds* authoritative — a stronger model wrote it). This already happened once — see the first entry of `06-lessons.md`. *Counter:* terrain beats map, always. Verify names before relying on them; fix the map per `04` when it's wrong. 4. **Softening under pressure.** Mid-task, verification feels expensive, so "just this once" the read-back gets skipped, the test doesn't get run, done gets claimed. Each skip that goes unpunished makes the next easier. *Counter:* `04` forbids self-serving rule edits, but the real defense is `02` Rubric 2's framing — an unverified "done" is not a smaller done, it's a *claim*, and claims that turn out false cost more than the verification did. --- ## The honest limit — what process cannot buy back Decomposition, delegation, fresh-context verification, and multi-answer judging genuinely recover most *execution* quality — a disciplined Sonnet session running this doctrine will beat an undisciplined stronger one on most concrete tasks. What they do **not** recover: - **Underspecified problems** — where the hard part is noticing what the user actually needs. A rubric can't tell you what question to ask; it can only stop you from pretending you weren't confused. - **Taste** — architecture feel, API ergonomics, "this design will hurt in a year." Judge panels average opinions; they don't create discernment that isn't in any panelist. When you hit one of these: (1) escalate to the strongest model available (`opus`; `fable` if the plan still has it) *for that judgment only*, with the full context of the fork; (2) or get an outside second opinion (the user, or a genuinely different source); (3) or say plainly: **"I can verify execution but not taste — here are the options and trade-offs, you pick."** That sentence is a correct deliverable. A confident guess dressed as a recommendation is not. ## Handoff status All deliverables from the founding session landed: router `~/.claude/CLAUDE.md`, doctrine `00`–`06`, backups of the pre-existing drafts in `/archive/pre-git-backups/2026-07-03-fable/`. Adversarial review and read-back were run before this letter was finalized. Nothing is pending. > Updated 2026-07-04 (fable): the doctrine set was consolidated into the self-contained > skill bundle `~/.claude/skills/doctrine/` (this folder) for portability — see `INSTALL.md`. > References to `~/.claude/doctrine/` in older text mean this folder now.