Researched live on **2026-09-17**. These are proposed working agreements, not vendor policy. The 2026 releases are dated; several current documentation pages expose no publication/update date, noted below. ## 1. RULES Keep the seven imperative sentences in the shared file; the examples and rationale are review annotations. ### `finish-authorized-work` **Rule:** Complete requested, reversible work without reconfirmation, asking only when missing information materially changes the result or an action exceeds existing authorization. **Why:** Prevents unnecessary permission loops. **Good:** Fix the requested bug and run the relevant checks. **Bad:** End with “Shall I implement the fix?” **Backing:** [Astra guide][a], [Fable guide][f]. **matters most for: both** — Both guides describe premature pauses for already-requested work. ### `keep-requested-scope` **Rule:** Implement every requested behavior, make targeted edits, and report unrelated improvements without implementing them. **Why:** Keeps the change complete and reviewable. **Good:** Fix pagination; mention the unrelated caching defect separately. **Bad:** Rewrite the data layer while fixing pagination. **Backing:** [Fable guide][f]. **matters most for: Fable** — Its guide identifies unsolicited extensions and whole-file rewrites. ### `explain-real-blockers` **Rule:** When an instruction blocks progress, identify its file and exact requirement, explain the conflict, and finish independent authorized work. **Why:** Makes permission decisions auditable. **Good:** “The deployment rule requires approval; the patch and checks are ready.” **Bad:** “Policy prevents further work,” without identifying any policy. **Backing:** [Astra guide][a], [Claude instruction docs][m]. **matters most for: Astra** — OpenAI specifically warns that conflicting skill instructions can halt work. ### `bound-verification` **Rule:** Run required checks and checks relevant to the change, then stop unless failures, further edits, or unresolved risks justify more testing. **Why:** Establishes a measurable completion boundary. **Good:** Run the affected regression test and required lint check; report results. **Bad:** Repeatedly rerun the full suite after an unchanged, successful result. **Backing:** [Astra guide][a], [Astra instruction-design guidance][b]. **matters most for: Astra** — Its documented thoroughness can produce excessive verification. ### `make-progress-visible` **Rule:** State the initial action, report meaningful findings or blockers during extended work, and finish with a standalone account of changes, check results, and unfinished items. **Why:** Users need visibility beyond tool output. **Good:** “Fixed empty-page handling; regression test passed; integration tests could not run without the service.” **Bad:** Several minutes of silence followed by “Done.” **Backing:** [Claude prompting guidance][g], [Fable changes][n]. **matters most for: Fable** — Anthropic documents fewer intermediate updates and provides explicit progress-display support. ### `ground-code-claims` **Rule:** Read the relevant code before describing its behavior, and distinguish observed results from assumptions in the final report. **Why:** Makes conclusions inspectable. **Good:** “The opened handler retries twice; production behavior remains unverified.” **Bad:** “Production is fixed,” based only on an edited file. **Backing:** [Claude prompting guidance][g], [Astra instruction-design guidance][b]. **matters most for: both** — The guidance supports relevant investigation without compulsory whole-repository exploration. ### `verify-changing-facts` **Rule:** Search unfamiliar or fast-changing product names exactly as supplied, open authoritative sources, and disclose when current verification is unavailable. **Why:** Familiarity can conceal stale information. **Good:** Verify the named SDK version before recommending its API. **Bad:** Substitute a remembered predecessor and answer confidently. **Backing:** [Fable guide][f]. **matters most for: Fable** — At low effort, it may answer from memory instead of searching. ## 2. MODEL-SPECIFIC NOTES - **Astra adapter:** Request concise, outcome-first prose and selective formatting; OpenAI documents detailed, heavily formatted defaults. Applying aggressive anti-formatting language to Fable can suppress useful structure. [Astra guide][a], [Claude prompting guidance][g] - **Fable adapter:** Encourage paragraph breaks and useful lists; preserve progress narration. Its writing can be dense and its updates sparse. These corrections do not address Astra’s documented verbosity tendency. [Fable guide][f], [Claude prompting guidance][g] - **Delegation:** Astra may need encouragement to delegate. Fable’s documented tool-loop issue concerns batching independent calls; these are different problems and should not become a shared “always parallelize” rule. [Astra guide][a], [Fable guide][f] - **Loading:** Maintain the shared rules in `AGENTS.md`; use a small `CLAUDE.md` containing `@AGENTS.md`, followed by the Fable adapter. Codex discovers `AGENTS.md`; Claude Code documents this import explicitly. Keep Codex-specific instructions in its own configuration scope. [Codex instructions][c], [Claude instructions][m] - **Harness settings:** Fable’s progress-display and preserved-thinking controls belong in integration configuration; Claude Code already preserves the required conversation prefix. They are not portable instructions for Astra. [Fable changes][n] ## 3. REJECTED - **“Never ask questions.”** Both models need an exception for consequential ambiguity. - **“Ask before every edit.”** Reintroduces documented unnecessary approval pauses. - **“Always add tests and run everything.”** Encourages excessive testing and unrelated test artifacts. - **“Read all project documentation first.”** OpenAI explicitly identifies this as excessive for small changes. - **“Never use Markdown.”** Fable already tends toward insufficient formatting. - **“The user is absent.”** Fable’s autonomous-work example uses this premise; a shared interactive guide cannot assume it truthfully. ## 4. SOURCES Every opened URL is listed below. **Undated** means no publication/update date was shown; retrieval today does not establish an update date. | Source | Date shown | |---|---| | [Astra prompting and behavior guide][a] | Undated | | [Rethinking skills and prompts for GPT-6 Astra][b] | 2026-09-11 | | [OpenAI API changelog][o] | Astra release entry: 2026-09-03 | | [Codex AGENTS.md original URL][r] | Undated; redirects to next source | | [Codex AGENTS.md canonical documentation][c] | Undated | | [Fable 5.1 overview][v] | Model release: 2026-09-01; page undated | | [Fable 5.1 prompting guide][f] | Undated | | [What’s new in Fable 5.1][n] | Undated | | [Fable 5.1 announcement][l] | September 2026 | | [Claude prompting best practices][g] | Undated; explicitly includes Fable 5.1 | | [Claude Code CLAUDE.md and AGENTS.md documentation][m] | Undated | [a]: https://developers.openai.com/api/docs/guides/latest-model [b]: https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra [o]: https://developers.openai.com/api/docs/changelog [r]: https://developers.openai.com/codex/guides/agents-md [c]: https://learn.chatgpt.com/docs/agent-configuration/agents-md [v]: https://platform.claude.com/docs/en/models/fable-5-1/overview [f]: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1 [n]: https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1 [l]: https://www.anthropic.com/claude-fable-and-mythos-5-1 [g]: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices [m]: https://code.claude.com/docs/en/memory