The bias-towards-action block is an allowlist with a deferred gate. OpenAI's recommended text names what the model may do unsupervised: create isolated work trees or checkouts, resolve merge conflicts, perform read-only actions, open draft PRs, with a carve-out for anything clearly destructive or irreversible [5]. Approval arrives afterwards. The docs put it as a rule about what the human is handed: the user should be approving a concrete, reviewable result [6]. OpenAI also recommends stripping unsolicited warnings, disclaimers and hypothetical-risk checklists out of prompts [7].
So the gate moves from before the work to after it [4]. If a run now produces a branch and a draft PR before anyone is asked, the thing bounding the blast radius is repository permissions and whatever sandbox the agent runs in, and the prompt is just a request.
Here is the part worth pricing. The same documentation says Astra follows longer instructions better than its predecessors but is more sensitive to context [8], and that unclear or contradictory lines in skill files such as AGENTS.md can make it block work or veer off unexpectedly [9]. The prescribed fixes are three further blocks of instruction text: initiative, prose style, and the slop-word list [3]. The cure is measured in the same units as the failure. Hence the accompanying advice to audit every skill file and context document the model can reach and to give user instructions explicit priority [10].
The diagnostic prompt is the strongest engineering on the page, because it makes the model produce evidence: name and link the exact SKILL.md it read, quote the instruction, and separate an explicit skill requirement from its own interpretation of a guideline [11]. That converts "it stopped" into a path and a line you can go and delete.
The blocklist is harder to operationalise than it looks. It covers literal strings such as "delve into", "it's worth noting" and "Conclusion:" in conclusions, but at least five of its entries are sentence shapes rather than tokens: "Question? Answer", "This isn't about X. It's about Y", contrastive "X, not Y", compound descriptions and hyphenated adjectives, and any mention of what the model will not do [14][15][1]. Substring linting catches the first group and nothing in the second. The presence of "Conclusion:" on a list of things to avoid in conclusions reads like documentation written by someone who has graded a great deal of model output. And note that the ban on hyphenated adjectives sits on the same page as an initiative prompt instructing the model to perform read-only actions [2]; applied literally to technical prose, that rule removes a compound carrying a permission boundary.
What the source does not contain is a number. There is no measured stall rate, no before-and-after on Sol, nothing that would let you estimate how much of your own agent's idling this prompt text recovers [1]. For these blocks to transfer, your stalls have to be the over-clarification the docs describe [2][3] rather than a contradictory skill file [9], and those have opposite fixes: one wants more instruction, the other wants a line deleted. That argues for running the SKILL.md diagnostic first and pasting the initiative block second.
