I was using Fable wrong, this is how I fixed it

Technological / Product Analysis & Key Claims (Fable 5.1 vs “Astro/Astra”, Claude tools, agent workflows)

Output stability & “noise” reduction

• The speaker argues Fable 5.1 is less noisy than an alternative (“Astra/Astro”), based on a prior comparison (including a “quality output” graph).
• Benefit of less noisy output:
  • The model can run longer within one pass
  • It can return better at “come back when finished” behavior, especially for long tool-using runs.
• Tradeoff:
  • Fable still has “dips” (quality drops).
  • The speaker emphasizes reducing these dips is high value because they hurt more when the model is otherwise reliable over long runs.

Prompting / tuning guide derived from Anthropic’s Claude Fable 5.1 docs

• The video aims to distill learnings over months into actionable prompting techniques.
• Core theme: many behaviors can be steered via prompt instructions, not only via system/tool settings.

Agent cost / reasoning-effort levels

• The speaker references Anthropic’s prompting guide section: “consider all effort levels.”
• Recommendation (for their use case):
  • Default to High reasoning
  • Avoid Low/Medium because it often fails, leading to re-runs and extra tokens.
• Behavior differences described:
  • Low → max can waste tokens on complex prompts.
  • On simple tasks, higher effort doesn’t necessarily explode cost (Fable remains efficient enough).
• Comparative claims:
  • Astra adapts token usage better to task size (e.g., small worst-case token gaps in a benchmark).
  • Gemini 3.1 Pro can use very large reasoning tokens.

Subscription usage economics

• The speaker says Fable isn’t worth full API price, so they use it via subscription allotments.
• Approximate quota structure:
  • Half the weekly limit reserved for Fable
  • Other half used for other models (they call Opus “spiky” / problematic)
• Practical observation:
  • By week’s end, they often use most or all of the Fable half.

User-facing progress updates during long tool calls

• A key behavior change in Fable 5.1:
  • It writes fewer noisy user-facing updates during long tool-call turns.
• If users do want progress updates, the speaker recommends prompting for them explicitly:
  • Add a line in the system prompt like “brief updates while working”
  • Provide updates on:
    • what it found
    • what it did
    • what’s next
  • This helps when a reader sees only the last message.

Reasoning trace availability & history edits

• The speaker notes an API/system behavior change:
  • If you edit earlier messages in a thread, you may lose access to reasoning traces.
• Rationale given:
  • Mitigates distillation attacks (preventing extraction of hidden reasoning).

Formatting controls

• Writing density
  • Fable 5.1 formatting is improved, but it can be more dense (longer sentences, fewer paragraph breaks).
• “Mannered prose” anti-pattern
  • Claude may use metaphor/flourish instead of direct statements.
  • Suggested explicit instruction: “Please remove all mannered pros.”
• Formatting instruction best practice:
  • The speaker advises deleting/avoiding over-specific formatting constraints in Claude Markdown / agent MD (e.g., “don’t use bold”), since they can over-penalize formatting.
• New product detail:
  • An update in “Claude/Cloud code” now uses agent MD from the repo (not only Claude MD) in some conditions.
  • Result: outdated formatting instructions might suddenly affect outputs.

Steering citations / quotes

• The model may reproduce source text without marking it as quotes.
• Workaround:
  • Provide an example prompt demonstrating how the model should mark quotes.

---

“Most Useful” Operational Prompting Strategies (finish the task, keep going, agent skills)

Ensure it actually finishes

• For long asynchronous work, the model may describe next steps (“Next, I will…”) without doing them.
• Fix:
  • Add a nudge like “go do it”
  • Or ensure the prompt sets expectations that it must execute to completion, not just plan.

“Babysit” / monitoring via skills

• The speaker describes a skill-based workflow for long PR/review cycles:
  • A “babysit PR” skill monitors new comments/checks.
  • Only acts on findings newer than the latest push.
  • Verifies bot findings against source/CI and resolves review comments.
  • Avoids scope creep and filler comments.
• The goal:
  • Keep working until everything is green, then stop.

End-state control is central

• The speaker repeatedly emphasizes prompting should specify a stop condition, such as:
  • “Wait until everything is green and merge.”
  • “File PR and tell me when it’s up.”
• Multi-path endings:
  • “If not happy, report; if happy, continue for an hour/merge without further intervention.”

---

Real-World Workflows / Examples Mentioned

T3 code + screenshot/image rendering fix

• Example: diagnosing screenshot/image rendering failures in T3 code.
• Approach:
  • Use thread/tool-call history to identify failures.
• Outcome:
  • Merge a single PR addressing root causes.

Large “YOLO” experiment: runtime replacement

• The speaker ran an experiment on LakeBed (their sandbox/runtime project).
• Approach:
  • Consider a risky upstream change
  • Ask the model to assess readiness
  • Use staging-style testing
• “Derisking” steps proposed:
  • shadow mode
  • synthetic traffic / counters
  • structured handling for refresh failures
• Outcome:
  • The model merged changes, created additional PRs as needed, soaked/staged them, and iterated until confidence was achieved.
• Evidence cited for correctness:
  • 1,400 receds in 160 steps, with no wrong answers (as support for correctness during incremental changes).

Verification via other agents/models

• Recommended workflow:
  • Use Fable to write code and propose changes.
  • Optionally delegate verification to another model (e.g., having Fable ask Astra for review/testing when appropriate).
• For computer-use tasks:
  • Use a “codec”/computer-use tool setup on-device.
  • Fable can handle image-based verification, then ask Codeex/Astra to confirm.

---

Safety, Correctness, and Sandboxing Philosophy

• The speaker frames improving agent reliability as enabling “bigger” autonomous cycles, such as:
  • merge + monitor + respond to reviews
  • staging testing + shadowing
  • “stop when done” end states
• Explicit warning:
  • This is not for every production scenario yet—it depends on having staging/QA and robust test systems.

---

Sponsor-Related Technological Concept (Localization + Agent PR/CI Workflow)

Localization platform (General Translation)

• Key emphasis:
  • Localization is necessary for global markets (“85%” not English claim).
• Claimed agent-friendly workflow:
  • Agents can write translatable code by integrating with source-code level workflows.
  • A glossary enforces consistent terminology across app surfaces.
  • Workflow:
    • write localized changes
    • file PR
    • CI/localization pipeline kicks in

---

Source Speakers / Main Credited Sources (as requested)

Main speaker / source

• The YouTube video narrator (author of “Fable versus Astra” and the prompting guide summary).
• Explicitly mentions:
  • Anthropic Claude docs
  • their own T3 Code setup

External sources referenced

• Anthropic
  • Claude Fable 5.1
  • Claude prompting documentation
• Sentry
  • bug tracing
  • “Sentry MCP” for agents
• General Translation
  • localization sponsor