📜 From Delta to Playbook
The delta has already been measured—now it becomes action. This track turns the numbers into rules
within a single .md actionable, and then injects this playbook
to the model in every session (hook, skill, or CLAUDE.md) — with a plan B if you don't have your own data.
Track map
Detailed content
📝 Writing the Playbook
From the number to the rule: what a playbook is, the generator, ordering by signal strength, fixing instead of copying, transferable rules, and the one-line checklist.
A single .md actionable, capturing the delta as RULES (“think before acting”), not as an impression or video narrative.
Rules can be injected; impressions can’t. The playbook is the artifact that enters the model’s context.
1 file; imperative instructions; real numbers embedded; honest about what does NOT transfer.
The generator: reads the compare.json (--from-json) or measure in real time, and inject the actual NUMBERS into the rule text.
It’s what ensures the playbook speaks to YOUR data, not generic impressions.
--from-json reuses measurement; without it, measures the history; numbers embedded.
The rules are ordered by measured STRENGTH: think before acting (robust delta, +45 pts) first; read-before-edit and test-after are good practices.
The order communicates priority: the rule at the top is the biggest lever.
Strength ≠ video order; solid delta at the top; small sample → caution with the rest.
The honest playbook INVERTS Fable’s weak habits (overthinking simple tasks, verbosity, plans that turn into essays) instead of copying them.
Blindly copying carries over the flaws too; correcting means keeping only what helps.
Reasoning proportional to difficulty; act and keep summaries brief; lean plan.
Think before tackling non-trivial tasks; purposeful tool use (density); read before editing → 100%; close the loop with a test/build; canonical sequence; stop when you have enough.
These are the instructions the target model adopts—the heart of the playbook.
understand → plan → read → edit → test → report.
The final distillation in one line, to paste at the top of coding tasks: entender → plano curto → ler → editar → testar → relatar.
It’s the minimal reminder that fits into any prompt at no context cost.
think before non-trivial tasks · purposeful tool use · read-before-edit→100% · no verbosity.
💉 Injecting into the Model (and With No Data of Your Own)
Hook, skill, or CLAUDE.md; plan B with open traces on Hugging Face; what transfers and what doesn’t; and the ethics of closing the loop.
O inject-playbook.sh injects the playbook as additionalContext each session; config in settings.json; it's fail-open.
It’s the “always on, no thinking required” approach—the model starts with the right pace.
event SessionStart; additionalContext; fail-open (doesn't break the session).
The skill fable-mindset loads the playbook on demand and brings the scripts along — for when you prefer to invoke it explicitly.
You don't always want the playbook in EVERY session; the skill gives you control.
on demand; scripts included; explicit invocation.
Paste or link the playbook in the CLAUDE.md — it’s already auto-injected into every session. Minimal effort.
Versioned with the repo, with zero hook configuration.
hook = always active / skill = on demand / CLAUDE.md = minimal effort.
If you’ve barely interacted with Fable, use open traces on Hugging Face: Glint-Research/Fable-5-traces. O import_hf_traces.py (+ pip install datasets) runs the same exercise.
No one is left out for lack of history — the method runs on third-party data.
open dataset; defensive importer; same behavioral exercise.
You imitate the RHYTHM (think first, read before editing, close the loop); you don’t clone the model—the power is in the weights. The playbook improves execution; it doesn’t turn you into Fable.
Sets expectations: real gains in pace, with no magic promises.
pace transfers; weights don't; execution improves, identity doesn't change.
Run it on YOUR history; treat the corpus as sensitive (it contains your code/data); the method applies to Opus, Codex, and open source; iterate on the playbook with new data.
Closes the loop responsibly and keeps the playbook alive.
sensitive corpus; generalizes across models; continuous iteration.
The loop closes here
This course is part of the INEMA.CLUB community — research, education, and experiments with applied AI. From log to injected playbook, the method is yours.