Stakeholder Clone
Reconstruct a specific real person's judgment from their artifacts, then consult that clone to preview how they will react to your work — before you submit it. The source calls it "AI time travel": build a system that gives you the feedback the real stakeholder would give, ahead of the finalized version, so you can act on it while there is still time to change the work. You clone the person by having the model ingest their available corpus — public content (blog posts, talks, podcasts) or private artifacts you legitimately hold (a manager's Slack messages, emails, docs) — and then ask the reconstruction to evaluate your draft as that person.
The source packages three variants: "ask the board" (clone 4–5 thought leaders in your vertical from their public content — his example roster is Alex Hormozi, Mark Cuban, Andrej Karpathy — as a strategic sounding board); "internal focus group" (have the model interview to clone the end user of what you are building, for anything customer-facing); and "clone your manager" (for day-to-day work, reconstruct the person who will actually approve it). He reports an Anthropic growth lead cloned his manager from her blog posts, Slack, and emails to get her feedback before submitting.
What distinguishes this from its neighbours is two axes at once. Unlike the Adversarial Planning Council, the personas are specific named real people reconstructed for fidelity, not engineered adversaries built for opposition — and the target is your near-finished output, pre-submission, not an idea pre-build. Unlike Advisor Mode (a smarter model consulted when stuck), the clone supplies not more capability but a particular person's taste and priors. The honest caveats are real: a clone is a reconstruction of a judgment, not the judgment — it inherits whatever the corpus reveals and confabulates the rest — and cloning a real individual from their private messages raises consent/provenance questions the source does not address.
Claims
- A specific real stakeholder's likely reaction can be previewed by reconstructing their judgment from their artifacts and consulting the clone before you submit. best practice — context: any strategic, customer-facing, or approval-gated deliverable where a particular person's reaction is the bar and their corpus is available; weakest when no representative corpus exists (the clone then confabulates) or when you need genuine disagreement rather than one person's taste. The value is feedback timing — moving it upstream of submission — not a guarantee the clone is faithful.
- Clone for fidelity to a named individual, not for engineered opposition — the point is whose judgment you are simulating. principle — durable distinction: the informative signal is a specific person's priors and taste, which is a different instrument from a devil's-advocate panel built to attack an idea. Getting a real reviewer's reaction early is orthogonal to manufacturing dissent.
- Match the clone to the deliverable's actual gatekeeper: thought-leader board for strategy, end- user clone for customer-facing work, manager clone for day-to-day approval. best practice — context: choosing whose reaction actually decides the work's fate; a mismatch (polishing for a board when your manager approves it) previews the wrong judgment.
- A clone is a reconstruction of a judgment, not the judgment — it is bounded by the corpus and fills gaps by confabulation. principle — durable and load-bearing caveat: the same limit that makes any persona sub-agent "a strong hypothesis, not proof" (see Adversarial Planning Council); trust the clone as a cheap early read, verify against the real person for anything consequential.
- The video states an Anthropic growth lead cloned his manager from her blog posts, Slack messages, and emails to get feedback before submitting. observation — attributed anecdote, groundable and check-worthy; recorded as the source's claim, not adjudicated.
Related
- Adversarial Planning Council — the closest neighbour and the sharp contrast: a council builds engineered adversaries to critique an idea before the build; a stakeholder clone reconstructs a specific real person to react to output before submission. Same persona-sub-agent primitive, different intent and target.
- Advisor Mode — the other "consult something before finalizing" pattern; advisor mode pulls more capability from a smarter model, this pulls a particular person's taste.
- Curatorial-Voice Learning — the selection cousin: "who over what." Choosing whose judgment to clone is the same weighting of a specific person's voice over a generic standard.
- LLM-as-Judge — the clone is a judge specialized to one person's priors; the same calibration and independence concerns apply to trusting its verdict.
- Context Substrate — the corpus that makes a clone faithful is standing context; a clone is only as grounded as the artifacts fed into it.
- Evidence-Gated Completion — the output-side gate; the clone previews a human's reaction, evidence-gating proves the work is real. Complementary pre-submission checks.
- Distillate: You're the Problem, Not Claude — Six Fixes to 10x Output