recipient-first

Foundation

Two large field syntheses set the expectation for everything below.

• Across 49 field experiments, the best estimate of the persuasive effect of

campaign contact — mail, calls, canvassing — on candidate choice in US general elections is zero (Kalla & Broockman 2018). The same paper does find clear effects in primaries and ballot-measure campaigns, where voters have no party cue to fall back on.

• Behavioral interventions run by two government nudge units moved outcomes by

1.4 percentage points; a separate set of academically published nudges reports 8.7. Publication bias, exacerbated by low power, accounts for the entire gap (DellaVigna & Linos 2022, 126 RCTs, 23 million people).

Neither study manipulates wording, and both are institutional rather than interpersonal — treat the transfer as an assumption, not a finding. But the direction is consistent and it is the honest prior: wording is the last and smallest lever. What moves outcomes by large factors is who you ask, through which channel, how much work you leave them, and whether you understood their position before you wrote.

The layers below are ordered by how much of the outcome they plausibly control, not by a common effect metric — the underlying studies report percentage points, ratios, d and r, which cannot be ranked against each other. Apply them in order. Most messages are finished in Layer 1 or 2.

Proportionality. Skip to the checklist if the message is under about three lines and carries no refusable ask, or if it goes to someone you talk to daily about routine matters. A four-line recipient analysis in front of a nine-word Slack message is a failure of this skill, not an application of it. The skip does not apply to apologies, bad news, or anything crossing a power difference, however short — those keep the full pass.

Admission rule — the full-disclosure test

**Would the recipient, knowing exactly how this message was constructed, still
think the ask was fair?**

Any technique that fails this is out. Reframing which true fact leads is in; inventing the fact is not.

Not decoration, and load-bearing for two reasons. A fabricated frame is discovered eventually, and the cost lands on every later request you make — the game is repeated. And when a language model does the drafting, the risk is measured rather than theoretical: in debate pairs where the two sides were not equally persuasive, GPT-4 with access to basic demographic facts about its opponent was the more persuasive side 64.4% of the time (Salvi et al. 2025).

Unverified is not the same as true. If a claim's truth is unknown to you — is he really the only one who could do this? — you may not use it, and you may not assume it false either. Ask whether it's true, or leave it out. A claim the user recommends for its effect ("that usually works on him") rather than its accuracy is being offered as a lever, not a fact; treat it as unverified.

A genuine preference is not a deadline. "I want this done this week" is true and usable. "This has to be done this week" is a deadline, and needs one to exist. State the wish as a wish.

If the user asks for an excluded move — manufactured urgency, a false exclusivity line, an invented deadline — write the honest version, and say in one line which requested element you left out and why. Do not silently comply, and do not refuse to produce anything.

Never invent the recipient's motives, or the facts of the incident. If you do not know what the recipient cares about, what they are busy with, or what they already know, ask the person you are drafting for. If asking is not possible — non-interactive run, or the user genuinely does not know — then draft anyway, with a one-line assumption block above the draft naming what you assumed and which assumptions would change the message if wrong. Stalling with questions and no draft is a failure mode too.

Layer 1 — Correct the sender's three biases

These are properties of the writer, not the reader, and they are measured. What is not measured is how much correcting them improves outcomes; no study tests the procedure below. It is a reasonable inference from the biases, not a result.

1. You underestimate how often people say yes — by roughly half. Across

studies in which participants made requests of more than 14,000 strangers, requesters underestimated compliance by an average of 48%, because they do not feel how socially costly refusal is (Flynn & Lake 2008 — Lake now publishes as Bohns; Bohns 2016). The error persists for friends, for large requests, and after a previous refusal. → Ask, and ask directly rather than hinting. The most common failure is not a badly phrased request; it is a request never made, or softened into a hint that never registered as one. → Re-asking is bounded. Ask a second time only if circumstances changed or the first attempt was a hint rather than a request. Never re-ask someone you hold power over, never on a personal or romantic matter, and never where the refusal was on principle rather than on capacity. The underlying studies are strangers and small favours; past that, this stops being persistence. → When re-asking is barred, the need doesn't disappear — reclassify it. Either reopen it as a question (find out what actually blocked them, with "no" still genuinely available), or, if the thing must happen regardless of their preference, say so as a decision rather than dressing a directive as a request. The second is more honest than a better-worded second ask, and usually less damaging.

1. You overestimate how well you were understood. Knowing something makes it

nearly impossible to model not knowing it (Camerer, Loewenstein & Weber 1989). Over email specifically, people overestimate both their ability to convey tone and their ability to read it — across five experiments — because they hear their own intended tone as they write (Kruger et al. 2005). → Assume your first draft is clearer to you than to anyone else. Name the thing you assumed they knew. If tone carries meaning — irony, mild criticism, a joke — it will not survive text. Say it literally instead.

1. You argue from your own values, not theirs. Asked to persuade someone who

disagrees, fewer than 10% of people spontaneously appeal to that person's values rather than their own (Feinberg & Willer 2019; base rate measured on US partisans writing persuasive essays). → State the case in terms the recipient already holds, not a repackaging of why you hold it.

Procedure — before drafting. Write four lines about the recipient. Keep them out of the message; show them to the user above the draft.

• What they already know, and what they don't.
• What they are currently busy with or worried about.
• What it costs them to say yes — time, risk, who they'd have to tell.
• What in this actually serves them, in their terms, or nothing if there is

genuinely nothing.

If the last line is empty, say so in the message rather than manufacturing a benefit. Naming the asymmetry outright — that there is nothing in it for them and you are asking anyway — is a legitimate move and passes the admission rule. Phrase it for the relationship; don't paste that sentence.

Layer 2 — Channel and friction

• Asking in person beat asking by email, by a wide margin. In a matched-script

experiment across 450 requests, face-to-face asks succeeded 34 times more often than emailed ones — and senders predicted no difference (Roghanizad & Bohns 2017; single study, strangers, not independently replicated). The experiment compared those two channels only; phone and video are untested, so do not restate this as "synchronous beats text". → If the channel is yours to choose and the ask is important and refusable, ask in person. If the user has already specified the channel, write for that channel and note in one line that the same ask made in person is the stronger version. Async teams, time zones, and accessibility are all legitimate reasons text wins; a drafting skill whose best output is "don't send this" is useless.

• Remove steps. Reducing the work between "yes" and done is the lever

practitioners rely on most (the "Easy" of the Behavioural Insights Team's EAST framework — a practitioner framework, not a meta-analysis that ranked levers against each other). Attach the file, pre-fill the form, propose two times instead of asking for availability, send the draft instead of the request to write one. Do the work yourself where you can, rather than promising it.

• Make the action concrete: who does what, by when. "Let me know what you

think" is not an action. "Reply yes or no by Thursday" is — but only propose a date that is real. A deadline invented to create concreteness fails the admission rule; use the coarsest true unit instead ("this week", "before the release"). The adjacent evidence is implementation intentions: if–then plans raise goal attainment at d = .27–.66 depending on format and context (Gollwitzer & Sheeran 2006, 94 tests, d = .65 at the upper end; domain-specific and later estimates run lower). Mind the gap — those are plans the actor forms for themselves, not a date a sender writes into a request. The advice here is craft; the effect size is only why concreteness is worth believing in at all.

• Give the ask one owner. A request addressed to a group diffuses until nobody

holds it. For announcements, where a group is the audience, name the owner of each action instead.

Layer 3 — Remove barriers (subtractive)

Layer 3 takes things out — jargon, vagueness, structural noise. Layer 5 adds persuasive devices. That is the boundary between them, and Layer 3 is worth much more.

• Jargon does not merely reduce comprehension — it reduces persuasion.

Technical terms lowered processing fluency and through it produced greater motivated resistance to the argument, higher risk perception, and lower support, with definitions available on hover and word count held constant (Bullock, Colón Amill, Shulman & Dixon 2019, N = 650). A reader who finds the text hard concludes the claim is suspect.

• Ease is read as truth — with a caveat you must hold. The robust, meta-

analytic form of this is the repetition effect: repeated statements are judged more true, d = .39–.50 across 51 studies (Dechêne et al. 2010). The non-repetition forms — clearer typography, simpler phrasing — are supported but smaller and less consistently replicated (Reber & Schwarz 1999; Alter & Oppenheimer 2009), and the inference is a learned heuristic that weakens when the reader is deliberately checking accuracy. → Use this only subtractively: remove friction from claims that are true. See Layer 5 for why you must not use the repetition form.

• Do not confuse "simple" with "casual." Across three online studies and three

field experiments (N = 67,632), formal government communications outperformed informal ones, against the predictions of both researchers and practitioners. Comprehension and perceived ease did not differ; perceived competence and trustworthiness of the sender, and perceived importance of the request, did (Linos, Lasky-Fink, Larkin, Moore & Kirkman 2023, Nature Human Behaviour). Scope is institution-to-citizen; between peers, match the room. → Cut jargon, nesting and length. Do not cut the register. Chattiness is not clarity, and it can cost the credibility that was carrying the request.

• State uncertainty as a number, not as a vague word. Across five experiments

totalling n = 5,780 — including a preregistered replication and a BBC News field experiment — numeric ranges barely moved trust in the figure or the source, while verbal hedging ("there is some uncertainty around this") did measurably more damage (van der Bles et al. 2020). Two honest caveats: the BBC field experiment itself was directionally consistent but null, and Kreps & Kriner (2020) found a very wide range did reduce trust relative to a point estimate. → Write "3–5 days" or "about 70% confident", not "possibly", "somewhat", "I think maybe".

• Distinguish marking ownership from marking doubt. "I think X" claims the view

as yours. "X, sort of, maybe" marks the claim as shaky. The first is free; the second is expensive. Hedges, hesitations and tag questions reduce both attitude change and perceived credibility (Burrell & Koper 1998, a book-chapter meta-analysis, largely courtroom-derived), and the damage peaks where it costs most: with a motivated reader, hedged strong arguments persuaded no better than weak ones (Blankenship & Holtgraves 2005). Hedge what you actually mean to mark as uncertain, and nothing else. Exception: in explicitly scientific framing, calibrated tentativeness can raise perceived trustworthiness (Jensen 2008).

Layer 4 — Disagreement, bad news, repair

Disagreement. Receptive language is measurable and teachable. The markers that make a counterpart perceive you as receptive: explicit acknowledgment of their view, statements of agreement where agreement is real, ownership of your view as yours, and positive rather than negated phrasing ("I agree that X" beats "I don't think Y"). Writers taught the recipe were rated more persuasive and more desirable to work with (Minson, Yeomans, Collins & Dorison 2024; in Wikipedia editing data, receptive openings predicted fewer personal attacks in reply — that specific field result is from the 2020 paper only). A scoring tool exists at receptiveness.net.

Counter-evidence, stated up front. Santoro, Broockman, Kalla & Porat (2025,
PNAS) ran a longitudinal field experiment and found that adding high-quality
listening to a persuasive appeal did not increase persuasion at all
(d ≈ 0.06, not significant; −0.02 at five weeks). The manipulation was live
conversation, so transfer to written receptive language is itself an assumption.
Treat receptive language as good for the *relationship and the continued
conversation* — which is well supported — and not as a lever on the other
person's position.

Listening. Feeling heard has a substantial association with how much the speaker trusts, and feels connected to, the person who listened to them — in this section that person is you, not the recipient the rest of this skill is about (Vogel & Gastil 2025, meta-analysis: 50 studies, 127 effect sizes, N = 9,601 overall; trust and relatedness r > .60, but on k = 5 studies each). Experimentally, high-quality listening lowers defensive processing and reduces attitude extremity (Itzchakov, Kluger & Castro 2017). The meta-analysis also reports a context moderator, but its description of the included studies cannot be reconciled with it, so no context comparison is relied on here. Practically: a reply that first restates their position in a form they would accept outperforms one that opens with the rebuttal — for the relationship, if not for their position.

Ten minutes, in person or by phone. A non-judgmental exchange of personal experience shifted entrenched attitudes in about one in ten people, and the shift survived three months and later counter-argument (Broockman & Kalla 2016, Science; extended to phone canvassing in subsequent work). One of very few interventions producing durable change. There is no evidence it works in writing.

Power asymmetry is largely unaddressed above, apart from the bound on re-asking in Layer 1. Manager-to-report, report-to-manager, and threads where colleagues are watching all change what you can say, and none of the cited research models it. Say so rather than applying these rules as if the parties were equals. Where one person is named but a group is reading, write for the named person and assume the group is scoring both of you.

Bad news. 78% of recipients want the bad news first; senders prefer to lead with the good and reverse the order (Legg & Sweeny 2014 — lab personality feedback with a forced choice, so a narrow base). Lead with the bad news. Documented catch: doing so reduces worry, and the relief slightly reduces motivation to act, so keep the required action attached to the bad news rather than deferring it. Carve-out: if the recipient needs context to act safely, give at most one sentence of it first, then the bad news.

For a death, a layoff, or a serious medical or safety event, this stops being an ordering question and nothing else in this skill applies either. The branch is: say it plainly and early in the fewest words that are still complete; give the reason once, without justification or softening; then what happens next and who they can talk to. Anything with employment, medical or legal consequence goes to whoever owns it before it is sent.

Apology. Six components, unequal, and more of them is better (Lewicki, Polin & Lount 2016, two studies). By measured weight:

1. Acknowledgement of responsibility — by a clear margin the most important.

"It was my mistake", not "mistakes were made" and not "I'm sorry you feel that way".

1. An offer of repair — a concrete thing you will do. Second.
2. Expression of regret, explanation of what went wrong, declaration of repentance

— roughly tied, all worth including.

1. A request for forgiveness — least important; drop it when it makes the apology

about your relief rather than their loss.

Liability carve-out: for customer incidents, HR matters, safety events, or anything with contractual or legal exposure, an admission of responsibility is not yours to draft unilaterally — and a sendable draft that contains one will get sent. So do not produce one. Write the facts that are yours to state, leave the responsibility sentence as an explicit placeholder, and say above the draft that the placeholder is for whoever owns the liability to fill.

Layer 5 — Wording, last and least

Reach here only after 1–4, and only for messages where it matters.

• A concrete story beats an equivalent argument, modestly. r = .17–.23 across

beliefs, attitudes, intentions and behaviours (Braddock & Dillard 2016, meta-analysis) — roughly d = .35–.47, though the .23 top end is behaviours on k = 5. Corroborated in the field: the Santoro et al. (2025) experiment whose listening arm was null found the persuasive narrative itself moved attitudes ~0.30 SD, still ~0.20 SD at five weeks. One specific case beats one general claim, and this is the largest lever in Layer 5 by some distance.

• Prefer the specific noun and the concrete number to the abstraction.
• Do not use repetition to make a claim feel true. This is the same effect

Layer 3 cites, used additively instead of subtractively, and that is the line: removing friction from a true claim is fair; repeating a claim so it accretes credibility is manufacturing belief and fails the admission rule. It also works on you — a claim you have written five times starts to feel verified.

Techniques to avoid — unsupported by evidence, or excluded by the admission rule:

Evidence grades

Grades for the load-bearing claims. Anything not listed here — friction reduction, one-owner asks, the four-line procedure — is reasoning from the graded claims, not an independently measured result, and should be treated as craft.

When in doubt, use the higher-graded layer and stop.

Checklist

• If the message needed them: did I write the four recipient lines, and keep them

out of the message?

• If there is an ask: is it stated as an ask, with one owner? If timing matters, is

the date a real one — and written as a wish where it is a wish?

• Is this the right channel — and if I couldn't choose it, did I say so once?
• What work am I leaving them that I could do myself?
• Any jargon, any nesting? Any casualness I mistook for clarity?
• Is every uncertainty a number, and every hedge one I mean?
• Did I assume anything about this person I was told, versus anything I invented?
• Would they, knowing how I built this message, still call the ask fair?

Output

Return the draft. Above it, put the four recipient lines (omit them if the proportionality rule applied) and — if you had to assume anything — the assumption block. Below it, one line each for anything the user asked for that you left out, and why. If the user's existing draft already clears the checklist, say so and change nothing; no rewrite is a valid result.

Scope and limits

• For composing and revising what a message says: requests, proposals,

explanations, disagreements, bad news, apologies. It changes content and stance.

• Not for visual formatting — line breaks, chunking, headers, symbols. It does

prescribe rhetorical ordering (bad news first, their position before your rebuttal); it says nothing about layout.

• One named recipient, or a bounded group you know (a team channel, a named

set of customers). For a bounded group, the four lines are written about the group's shared state, and the one-owner rule becomes "name the owner of each action". Marketing copy, fundraising, cold outreach and public posts fall outside: there is no recipient whose state you can know. Say so rather than inventing a persona.

• Sender side only. How to decline, how to receive criticism, and how to

respond to a manipulative message are outside this.

• Not for adversarial counterparties. The admission rule is the wrong standard

for a demand letter or a legal dispute.

• English-derived. The formality finding is UK/US institution-to-citizen; the

receptiveness markers come from an English NLP model. Languages with grammaticalized politeness are not covered — the register layer is a local question everywhere, and especially there.

• Not a rescue for a bad ask. If the request is genuinely against the

recipient's interest and you have nothing to offer, no framing fixes it. Say what it is, or don't ask.

• Diminishing returns are real. Once Layers 1 and 2 are done, the remaining

layers are worth much less. Do not spend an hour on Layer 5.

Full citations, sample sizes, limitations, and the excluded-technique dossier: references/evidence.md.