This page collects what the official texts of Anthropic say about false completion, sycophancy, and the end of a turn. Read it before you change the turn gate of the OpenSpec change claudetto-0-1-0, or a prompt of claudetto that tells the model how to report. The research was done on 2026-10-16, for Claude Code 2.1.301. Failure Gaps of 0.1.0 maps these facts to the gaps of claudetto.

The Opus 5.5 prompting guide has the section "Unattended agentic runs". It gives these rules:

• "Treat a text-only end of turn as a report rather than as proof the task is done."
• A separate, smaller model checks a completion condition at each end of turn.
• The harness stops after 2 or 3 automatic continuations.
• "If something the model started is still running, such as a background command or a subagent, don't treat the task as done yet."

The Sonnet 5.5 prompting guide has the section "Verification on coding tasks".

The best practices have no guidance on sycophancy. For a preamble, they say to "strip [it] in post-processing". Claude Code shows the reply before a Stop hook runs, so a plugin cannot strip a preamble (D58 in Engine Facts).

The files are in .claude/skills/claude-api/shared/. Each line number is of the copy in this repository on 2026-10-16.

Topic	Text	Source
Apology	"Don't add apologies or preambles, don't be overly self-critical"	model-migration.md:1123-1128 (Opus 5), kept for Opus 5.5 at model-migration.md:2051
Promise of work	Check the last paragraph, and when it promises work that is not done ("I'll..."), do that work now with tool calls	model-migration.md:1502 and model-migration.md:1817
Done with no check	The fix paragraph for Sonnet 5.5 and Haiku 5.5 at low effort	model-migration.md:2207 and model-migration.md:2369
Claim audit	The claim audit of Fable 5.1	model-migration.md:1484-1486
Whole task	The clause to finish the whole task	model-migration.md:1091
Verifier	"fresh-context verifier sub-agents tend to outperform self-critique"	model-migration.md:1518
Prohibition lists	A list of what not to do anchors the failure that it forbids	prompt-audit.md:100
Rules with no check	Make each such rule a hook, an allowlist, or a schema check, or delete it	prompt-audit.md:119
Pressure words	Pressure language and ALL-CAPS	prompt-audit.md:60-75
What to keep	The keep list of a prompt audit	prompt-audit.md:187-201
Risky commands	Move a risky bash action to a tool that the harness can gate	agent-design.md:28-31

The effort facts are these:

• Opus 5.5 has medium as its default effort.
• Sonnet 5.5 and Haiku 5.5 skip checks at low effort (model-migration.md:2183 and model-migration.md:2336).
• Haiku 5.5 keeps its rules best at high effort (model-migration.md:2338).
• Haiku 5.5 can drop a rule when the user argues against it (model-migration.md:2371).

The skill also names mechanisms of the API only: system messages in the middle of a conversation, clear_at, task budgets, and the advisor tool. These files do not show that Claude Code gives a plugin access to any of them.

live-sources.md has no URL for the prompting guides or the system cards.

The cards were read with pdftotext. Each page number is the number printed on the page.

The Opus 5.5 system card gives these facts:

• It reports a "modest countervailing increase in susceptibility to user pressure" (§6.1.2, p93-94).
• The audit is mostly in English, so it measures less misalignment in other languages (§6.4.11, p122).
• §6.4.3 (p110-111) defines the metrics of sycophancy and of false completion.
• Fewer than 1% of the sessions have a flagged destructive action. Most of the decrease comes from requests for permission (§6.5.2, p126-127).
• §6.2.1 and §6.2.2 give the rates of reward hacking.
• The model disclosed its changes in 96.9% of the cases (§6.5.4.4).
• The card names no feature of Claude Code as a mitigation.

The Sonnet 5.5 system card gives these facts:

• Sonnet 5.5 is better than Sonnet 5 on instruction following and on reckless actions (§6.2.2, p65).
• Opus 5.5 is better on most honesty metrics (§6.2.3).
• The card has no evaluation of reward hacking.

The Haiku 5.5 system card shows that Haiku 5.5 makes more false completion claims than Opus 5.5, Sonnet 5.5, and Mythos 5.1 (§6.2.3, p65). So a Haiku judge can classify a reply, but it cannot verify work.

• The hooks guide gives the Stop input and the hook types (D58 in Engine Facts).
• The tools reference gives the time limits of a background Bash command (D59 in Engine Facts).
• Checkpoints do not track an edit that Bash makes.
• The auto mode classifier and the permission modes are in Auto Mode.

• No text gives a mechanism that stops the user from seeing an opener or an apology. A Stop hook sees the reply after the user sees it.
• No text shows that a word list or a regular expression on the reply is a fix. claudetto does not match words of natural language, because a user can write in each language (task group 18).
• The texts give the rules of a judge, but they give no rate of false blocks for a judge in Claude Code. Stop Prompt Hooks has the only measured rates, from dotclaude 0.28.0.

• Failure Gaps of 0.1.0: the next page to read.
• Stop Prompt Hooks
• Engine Facts
• Evidence