AI that doesn't call it done until it checked.
The Fable Method is a discipline for AI coding agents: scope the work, ground every claim in real evidence, attack its own answer, verify before declaring done, and report straight. Fewer hallucinations. Less “done” that wasn't.
Five checkpoints · works on Opus or Sonnet · no install, no API key
AI's most expensive habit is confidence.
It says “done.” It didn't check. It states a fact it never verified. On a demo that's a shrug — in your business it's a wrong number in a customer's inbox.
One file. A discipline, not a dependency.
A single skill file. No install, no API key, no framework to learn. Drop it in and your agent starts working like an engineer who measures twice.
The Five Checkpoints.
Five checkpoints, run in order. Each one must pass before the next opens. Scroll through them — the rail lights up as you go.
- Checkpoint 1: Scope before work. You decide what "done" means up front — no moving goalposts, no half-finished work called finished.
- Checkpoint 2: Evidence before reasoning. Decisions come from real files and credible primary sources — not confident guessing, not some random blog.
- Checkpoint 3: Reason adversarially. It attacks its own answer before you have to — so mistakes get caught in-house, not in production.
- Checkpoint 4: Verify before declaring done. "It ran" isn't "it's right." Measure twice, ship once — the work is proven before it is called done.
- Checkpoint 5: Report calibrated. Straight answers — what is proven vs assumed is said out loud, and a real problem is never softened to sound agreeable.
Scope before work
State what "done" looks like before touching anything.
Define done + the check that proves it, persist the standard, separate known from assumed.
You decide what "done" means up front — no moving goalposts, no half-finished work called finished.
- Define "done" and the exact check that proves it.
- For multi-round work, write the standard down and test every round against it.
- Separate what is known from what is assumed — name the load-bearing unknowns.
Evidence before reasoning
Never design from memory — open the real thing.
Ground in real files and primary sources; push one item end-to-end before scaling.
Decisions come from real files and credible primary sources — not confident guessing, not some random blog.
- Open the real file, API, or data — memory is only a hypothesis.
- Prefer primary, authoritative sources over blogs and AI summaries; corroborate weak ones.
- Push one item through the whole pipeline before scaling to all of them.
Reason adversarially
Before committing, try to kill your own answer.
Red-team your emerging answer, steelman what survives; two failed fixes = wrong diagnosis.
It attacks its own answer before you have to — so mistakes get caught in-house, not in production.
- Attack your own answer as a hostile reviewer, then steelman what survives.
- Respect why the existing thing was built that way before changing it.
- Two failed attempts at the same fix means the diagnosis is wrong.
Verify before declaring done
"It ran" is not verification.
Check at the claim's layer with independent evidence; trust no delegate's "done."
"It ran" isn't "it's right." Measure twice, ship once — the work is proven before it is called done.
- Verify at the layer of the claim — "it ran" isn't "it's right."
- Use evidence you didn't generate; re-fetch web facts from a credible source.
- Never accept a subagent's "done" — re-run the check yourself.
Report calibrated
The report is part of the work.
Lead with the answer; state verified vs assumed plainly; never soften a real problem.
Straight answers — what is proven vs assumed is said out loud, and a real problem is never softened to sound agreeable.
- Lead with the answer, then the evidence behind it.
- Say what is verified and what is assumed, out loud.
- Never soften a real problem to be agreeable.
Measure twice, ship once.
Every checkpoint exists to catch the same failure: work that looks done but was never checked. Here's the shape of the difference.
Unverified "done" that slips through · per 10 claims
Illustrative — the Fable Method is a working discipline, not a benchmark. The shape is the point, not the exact numbers.
It brings a team.
The method doesn't just tell the agent to be careful — it wires in real tools at the exact checkpoint where they pay off.
Checks any web-based claim — versions, pricing, API changes — against the live page, then re-confirms it at verify time.
Facts get checked against the real web, not the AI’s memory.
Pulls current library and API docs before coding against them, so it never relies on stale training data.
It reads the up-to-date manual before it builds.
Fans out broad investigation — file scans, call-path tracing, gap-finding — in parallel while the main thread keeps momentum.
It sends a research team ahead so nothing gets missed.
Drives a real browser to check rendered state, fill forms, and test flows — screenshots become verification evidence.
It actually clicks through the site to confirm it works.
Runs a different AI model family in parallel to review the same change — two lineages catch different classes of bug.
A second, independent expert double-checks the work.
Spawns parallel worker and checker agents at scale, with adversarial verification, when a task genuinely warrants it.
It splits a big job across a team that checks each other.
The habits that make it hold up.
Habits that are always on
Small disciplines, every checkpoint, every task.
- Convert relative to absolute — "tomorrow" becomes a date, "latest" becomes a version.
- Surface constraints and trade-offs before they bite.
- Chase the cheapest probe of the biggest unknown first.
- Sort actions by reversibility — reversible: just do it; irreversible: confirm first.
- Unblock yourself before escalating; bundle the questions the user truly owns.
- Anything mechanical repeating 3+ times gets a script, not per-instance reasoning.
- Preserve by default — touch only what the task requires.
Smells that mean it skipped a step
Any one of these: stop, and go back to that checkpoint.
- Building something without opening the real data it depends on. (Checkpoint 2)
- You just said "should work" about something you can test right now. (Checkpoint 4)
- You’re on attempt three of the same fix. (Checkpoint 3)
- Your last three actions came from the plan with no check against results. (Checkpoint 3)
- About to report done, and the evidence is your intention, not an observation. (Checkpoint 4)
- A result came back suspiciously clean and you moved on. (Checkpoint 4)
- You can’t say in one sentence what "done" looks like. (Checkpoint 1)
- You marked a delegate’s work done on the strength of their own report. (Checkpoint 4)
- Your only source is a blog or AI summary, not the primary source. (Checkpoint 2)
Built for two kinds of people.
If you build with Claude Code
- Drop one file into ~/.claude/skills/ — no install, no API key, no dependency.
- Works on Opus or Sonnet; triggers on "fable mode" or automatically on hard tasks.
- Turns a fast-but-sloppy agent into one that grounds, verifies, and reports honestly.
- Stacks with your existing skills and tools — it’s discipline, not a framework.
If you run a business on AI
- AI that doesn’t hallucinate facts into a customer’s inbox or your dispatch board.
- "Measure twice, ship once" — work is proven before anyone calls it done.
- Straight reporting: what’s confirmed vs assumed, never dressed up.
- Boring tech that works — the discipline behind systems that actually hold up.
Get the skill.
Free. One file. Drop it into Claude Code and let your agent work like it actually checked.
What's in the download
fable-mode/SKILL.mdThe full method — five checkpoints, standing habits, and the team — ready to drop into ~/.claude/skills/.
INSTALL.md30-second setup and how to trigger it.
README.mdPlain-English overview of what it does and who it’s for.
Install in 30 seconds
- 1Unzip and drop the fable-mode folder into ~/.claude/skills/
- 2Restart Claude Code (or reload your skills).
- 3Say "fable mode" — or let it trigger itself on any task worth doing right.
Free · one file · works on Opus or Sonnet · no API key