fablize

name

fablize

description

A harness that makes Opus (or any Claude model) behave like Fable — it enforces seeing a task through to the end, with evidence and verification, as procedure. Use when starting a multi-step task (2+ sequential stories), long autonomous work, debugging or root-cause investigation, building render/executable artifacts (HTML, SVG, games, charts), or when the user says "fablize", "see it through", "verify as you go", "split into goals".

fablize — run Opus like Fable

Principle: a harness cannot raise a model's ceiling. It makes the model go all the way to its own ceiling — by enforcing verification, completion, and investigation as procedure. When the capability ceiling is the blocker (open-ended creative detail, self-driven discovery), escalate (§4).

Apply only what the task signals (smallest matching discipline; overlap only when genuinely multi-category). When installed always-on, this routing is automatic.

0. First run — set up automatically (once)

Before doing the requested task, check whether fablize has been onboarded on this machine:

cat ~/.fablize/progress.json 2>/dev/null

• If the file exists — skip onboarding, go straight to the task.
• If it is missing — onboard once with a single AskUserQuestion. Phrase the question and options in the user's current conversation language (detect it from recent messages — Korean, English, Japanese, etc.).
  • Question (meaning, translate to the user's language): "Set up fablize?"
  • Options (meaning, translate): "Local — this project only (recommended)" / "Global — all projects" / "Skip".
  • On Local/Global — run setup (it injects the block and writes progress.json; it never stars anything by itself), then continue with the task:

    bash ${CLAUDE_PLUGIN_ROOT}/setup/setup.sh <local|global>
    

    After setup, run bash ${CLAUDE_PLUGIN_ROOT}/setup/setup.sh ask — if it prints STAR_ASK <lang>, ask once via AskUserQuestion (in the user's language) whether they'd like to give the repo an optional GitHub ⭐, then run setup.sh star yes or setup.sh star no accordingly. If it prints nothing, skip silently.
  • On Skip — record it so it won't ask again, then continue:

    mkdir -p ~/.fablize && printf '{"setup_done":false,"skipped":true}' > ~/.fablize/progress.json
    

This means the user can just run /fablize (or trigger it) without running setup first — the first run onboards itself, once, with one question.

1. Multi-story loop (2+ sequential stories)

Decompose into sequential stories and complete one at a time, producing evidence as you go. Self-contained — no external goal system required. Run from the repo root; state persists in ./.fablize/ (resume with status even across sessions).

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/goals.py create --brief "<summary>" \
  --goal "title::verifiable objective" --goal "title::..."   # the last goal must be a verification story
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/goals.py next         # activate a story + handoff
# ... work that story only ...
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/goals.py checkpoint --id G001 --status complete --evidence "<concrete evidence>"
# the final story is a verification gate: --verify-cmd "<command>" --verify-evidence "<result>" are required
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/goals.py status       # first command when resuming

Rules: complete requires non-empty evidence; the final goal cannot complete without a verify command and its result (the engine refuses). If blocked, record --status blocked and report. Single-step tasks skip this loop.

2. Deep investigation (debugging / unknown cause / review)

Read and follow ${CLAUDE_PLUGIN_ROOT}/packs/investigation-protocol.txt: reproduce first → form 3+ competing hypotheses → gather evidence per hypothesis → trace the full causal chain (removing the symptom is not removing the defect) → verify before and after → report the hypotheses you rejected. For reviews, report everything including low-confidence findings and filter in a separate step.

3. Verification grounding (render/executable artifacts — always)

For artifacts whose correctness only shows when run (HTML, SVG, games, UI, charts), follow ${CLAUDE_PLUGIN_ROOT}/packs/verification-grounding-pack.txt: run it in the real renderer → observe the actual output → fix what the observation reveals → re-run. A static parse confirms well-formed, not correct.

3-1. Working style (always)

Lead with the outcome. Stay within the requested scope (no incidental refactors or abstractions). Ground every completion claim in a tool result from this session. Confirm before destructive or hard-to-reverse actions.

4. At the capability ceiling (escalate)

Signals you have hit the model's ceiling: stuck on the same problem 2+ times; open-ended creation where detail itself is the value; deep review that needs out-of-spec discovery. These are capability, not procedure, and a harness cannot fill them. In order: (1) adaptive thinking already scales with difficulty — recommend /effort xhigh to the user to push the current model to its ceiling; (2) reactive effort delegation — if the blocker is a bounded, hard slice (not the whole task), delegate just that slice to a background Workflow with effort:'max' (model inherited): package the evidence (symptoms, attempts, failure point, repro, the specific sub-question) as the agent() prompt, force a structured return via schema, then resume with its result as authoritative. This is the only real per-task effort knob in a normal session — the Agent tool exposes model but no effort; only Workflow/Agent SDK do. Opt-in, and not yet proven on real work (the shadow layer in docs/MEASUREMENT_PROTOCOL.md measures whether it helps): use it for a genuinely stuck slice, not routinely, and never trigger it from risk/deep classification alone — that over-escalates simple high-risk tasks (false-escalate); (3) if still short, hand off to a stronger model in a fresh session with the same evidence package; (4) otherwise report the limit honestly and name where a human must step in.

Install (always-on, optional)

Run once: bash ${CLAUDE_PLUGIN_ROOT}/setup/setup.sh → choose local (recommended) or global. Uninstall: bash ${CLAUDE_PLUGIN_ROOT}/setup/uninstall.sh. The UserPromptSubmit router hook registers automatically when the plugin is installed.