Skip to content

Plan Mode for Knowledge Artifacts

For PRDs, strategy memos, and architectural briefs, freezing direction in a reviewed plan before prose is generated compounds the gains plan mode delivers for code.

Plan mode's value on knowledge artifacts is higher than on code, not lower. Prose errors do not fail a test suite — they hide in reviewer fatigue, anchor downstream readers on the first draft, and only surface after someone has already built on a wrong assumption. Hannah Stulberg, who runs her PM workflow through this pattern, states the thesis plainly: "The plan is not overhead. The plan is the work" (Gupta × Stulberg).

Why knowledge artifacts suffer more

Two mechanisms compound for prose:

For code, a failing test catches misread intent. For a PRD, no such gate exists — the reader is the gate, and the reader is fatigued.

Activating plan mode

Shift+Tab cycles Normal → Auto-Accept → Plan. The status bar reads ⏸ plan mode on when active. Other entry points are claude --permission-mode plan, the /plan command, or "defaultMode": "plan" in .claude/settings.json (Claude Code common workflows). Press Ctrl+G to open the pending plan in your editor before you accept it. This is the concrete edit-before-drafting step.

The plan-mode posture

For knowledge artifacts, the plan-mode gate replaces first-draft generation with three disciplined moves:

  • Parallel research subagents. Fan out to subagents where "each agent writes to a temp file" and "the orchestrating agent compiles the final document from temp files" (Gupta × Stulberg). The plan lists the subagent tasks. Execution happens after you accept.
  • AskUserQuestion to push back on framing. Plan mode uses AskUserQuestion to gather requirements before it proposes a plan (common workflows). Stulberg inverts this: the agent surfaces gaps and challenges the framing, acting as a thinking partner rather than a producer (Gupta × Stulberg).
  • Self-verification clauses in the plan. Commit the plan to require cited sources and a named audience before any prose is generated. Plan mode is a prompt wrapper, not a sandbox. System reminders enforce the read-only constraint, not technical gating (Ronacher, 2025).

When not to use it

Skip plan mode when the artifact is shorter than the plan would be, or when write-to-think iteration is the point:

  • Short or templated artifacts. Standup updates and fixed-template postmortems. Anthropic's caveat: if you "could describe the diff in one sentence, skip the plan" (best practices).
  • High-iteration exploratory drafts. When the author is working out what they think, plan mode freezes direction too early.
  • Strong outline discipline. A plain markdown plan file on disk captures most of the same benefit. Armin Ronacher prefers this, but notes that plan mode's end-of-plan user prompt is a touch his approach does not replicate (Ronacher, 2025).
  • Known friction modes. Automatic plan-mode entry lacks an opt-out switch (#30042). Separate bugs document it breaking CLI message-routing (#29725) and re-entering after auto-compaction (#29956).

When the recommendation backfires

Plan-mode-for-PRDs underperforms a write-to-think workflow under three specific conditions:

  • Fast-moving discovery. An over-specified plan freezes direction before thinking has happened. Every accepted plan becomes a constraint on the learning the next iteration would have surfaced (the plan-first loop documents the same backfire).
  • Low-fidelity orchestrator synthesis. Parallel subagents pay off only when the orchestrator faithfully composes their outputs. Contradictory findings that are silently resolved produce a confidently-wrong synthesis. Single-agent reading would have surfaced the same findings as ambiguity.
  • Plan-file drift after acceptance. A plan committed at hour zero becomes stale as evidence shifts and scope contracts. Without supersession discipline (plan files as resumable artifacts), the accepted plan anchors the draft against a reality the author has already moved past.

Example

Take Stulberg's PRD workflow. Press Shift+Tab twice to enter plan mode. The agent then dispatches parallel research subagents across product/customers/ call notes, relevant metrics queries in analytics/, and prior PRDs in product/specs/. Each subagent writes a summary to a temp file. The orchestrator compiles the findings and uses AskUserQuestion to surface the tensions: gaps in the customer evidence, metrics that contradict the hypothesis, and prior decisions that constrain scope. Ctrl+G opens the proposed plan in the editor. The author edits the target audience, adds self-verification clauses ("cite a customer call for each user-need claim"), and accepts. Only then does the agent exit plan mode and draft the PRD (Gupta × Stulberg). The plan file is committed under product/specs/ and reused on the next PRD. The plan-as-artifact is covered in plan files as resumable artifacts.

Key Takeaways

  • On knowledge artifacts, plan mode's forcing function matters more than on code: there is no failing test, only a fatigued reader.
  • The mechanism is action-bias suppression plus reflect-before-produce — externally validated, not tool-specific magic.
  • The plan-mode posture is parallel research, AskUserQuestion as a pushback tool, and self-verification clauses — not a modal UI ritual.
  • Skip it for short, templated, or exploratory drafts; for strong-outline authors, a plain markdown plan file reproduces the benefit.
  • Plan mode is a prompt wrapper. The value comes from the decomposition, which is portable — the Claude-specific mode just automates the habit.
Established — multiple independent sources Last reviewed: 2026-06-13
Feedback