Fable Mode

Fable Mode — the five-gate task loop, as a callable skill

This is the invokable form of the method documented in [docs/03-the-five-gates.md](../../docs/03-the-five-gates.md) — read that for the full essay; this file is the execution checklist. It changes how you run the current task; it produces no files of its own.

A hard task is anything where the first idea might be wrong. For a one-file edit or a simple lookup, skip the gates and just do the work.

Run the gates in order — a gate must pass before the next opens

Gate 1 — Scope. Write done in 1–2 sentences + the check that proves it. Load standing rules (MEMORY.md, RULES.md). Name the 1–3 load-bearing unknowns. Ambiguity that changes the build → one question at the biggest gap; otherwise state your default and proceed.

Gate 2 — Evidence. Open the real file/API/page before designing against it; training memory is only a hypothesis generator. Recalled workspace memory is dated evidence — recency-gate it. Cheapest probe of the biggest unknown first. One item end-to-end before scaling to N.

Gate 3 — Adversarial. Try to kill your answer — and test the killing case, don't imagine it. Steelman the existing thing before changing it. Re-decide after every tool result. Two failed attempts at the same fix = wrong diagnosis: stop patching, test the assumption underneath. "Already solid" is a legitimate finding.

Gate 4 — Verify. "It ran" ≠ verified. Verify at the layer of the claim with evidence you didn't generate (live page, screenshot, diff, count). Sample the tails. A suspiciously clean result = broken verification until explained.

Gate 5 — Report. Lead with the answer. Separate verified from assumed, out loud, with specifics. Report observations, not intentions. Never state as fact what you haven't verified this session.

When stalled, name your gate

Stalling or surprised by a result? Say which gate you're at and re-run it. The smell list (three attempts at one fix, "should work", building on unopened data, a too-clean sweep) is in the doc — any one smell means stop and step back a gate.

Stacking

This skill sequences the task; it doesn't replace task-specific tools. Content that ships still passes both content-gate/ gates; research-backed claims still go through deep-research; the session still closes with session-close. Fable mode decides when to reach for them.

Escalation rule

If a task keeps failing under this discipline, that's the signal to bring the founder a calibrated blocker report (or try a stronger model) — not to loosen the process. Keep the discipline either way.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tkhribech
• Source: tkhribech/claude-founder-os
• License: MIT

Install and usage instructions live in the source repository linked above.