Rebuilding the scaffold that work conversations give you for free: a reason to speak, a topic, and a way out.
You're fine at work and lost socially. That isn't modesty and it isn't a general skill deficit — it's structural. A work conversation hands you three things for free: a reason to speak, a topic, and an exit. Remove the agenda and all three vanish at once.
| What you reported | What the agenda was supplying |
|---|---|
| "No natural reason to speak" to a stranger | Licence |
| Ideas come but seem lame, so you bin them | A standard for what's worth saying |
| Dry after the opener; dry when a topic ends | Topic supply |
| Can't escape someone monopolising you | A sanctioned ending |
| Stranded when they turn to someone else | A defined role in the encounter |
| "I'll be interrupting the group" | Turn allocation |
So this curriculum rebuilds those three things explicitly. That's a better problem than "you're bad at conversation," because scaffolding can be installed deliberately — and most of what you need turns out to be documented mechanics rather than talent.
Version 1 deferred every "what to say" module to Weeks 3–6, on a diagnosis that turned out to be a third of the story. That would have had you doing volume drills in Weeks 1–2 with nothing to say once you'd started. Content now runs from Week 2. A whole module on floor management — entering groups, escaping people, leaving cleanly — has been added, because it was missing entirely and it's half of what you actually described.
You asked for countable progress. Most of what follows counts behaviours, because those are what you control. One thing is worth measuring as an outcome: the six-item General Charisma Inventory, taken now and again at the end of Week 6. It splits into Influence and Affability, and my expectation is that your Influence comes back respectable and Affability lower — the work register rewards the first and never tests the second. That page also sorts the charisma-training evidence honestly: of the twelve trainable tactics from the one randomised trial in this area, about four are new and useful to you and six are speech-delivery technique that reads as performed in conversation.
| Week | Where | Primary drill | Rebuilds | The number |
|---|---|---|---|---|
| 1 | KL | The Window — fire within 5s of the spike | Getting started at all | Spiked reps/day · 3 |
| 2 | KL | The Lame Thing — say it anyway; find your licence | A reason to speak | Vetoes overridden/day · 5 |
| 3 | Barcelona | You Don't Need Material — follow-up chains | Topic supply | Follow-up ratio · >60% |
| 4 | Barcelona | In and Out — join groups, leave cleanly | Entry and exit | Clean exits/day · 4 |
| 5 | Lisbon | Presence — posture, gaze, attention | Not being in your head | Self-attention · <30% |
| 6 | Lisbon | Sharpening — good to great | Quality, finally | Boomerasks caught · →0 |
Running underneath all six: the fluency track — record, count, replace. Roughly three minutes a day.
The order follows the actual arc of an interaction: start it, open it, sustain it, navigate it, then refine it. Weeks 1–4 are the scaffold. Only Week 6 is about being good, and that's deliberate — you can't polish something you aren't yet reliably doing.
In KL you're leaving in two weeks; in Barcelona and Lisbon nobody knows you. The reputation cost of a badly-received approach — the thing that makes this expensive at home — is close to zero. You will never have cheaper reps than the next six weeks, so the volume and filter work goes first, while anonymity is highest.
The KL Window has verified dates for 4–15 August. Move the language-exchange night forward to Sunday 9 August — it's the only KL Language Exchange event before you fly, and the next is the 23rd. Also worth taking: The Lectern KL Toastmasters, Friday 14 August, the one club independently confirmed as open to all.
Full detail in Lesson 1. Your rule is wait until the spike settles, then act; sustained arousal during exposure predicts less fear later, while within-session fear reduction predicts nothing. Replace the rule with an if-then plan triggered by the sensation itself.
This week goes first because it gates the other two blocks — you can't practise what to say until you've started. It is not the only problem, and Week 1 is not expected to fix the conversation itself.
Reps: 3/day days 1–3, 4/day days 4–7. Trivial contact. No preparation —
preparing an opener is a safety behaviour.
Structure: run these as the
29 Missions — the verbatim task list from a
published, preregistered five-day intervention. It supplies the pretext you said you lack
(a mission turns "I should talk to someone" into "find someone wearing a hat," which is a
task — the format your work conversations already run on) and removes target
selection, which is where your veto lives.
Measure: reps, passes, disasters. Plus the study's own outcome measure —
each morning, predict "how many people will I need to approach before one talks to
me?"
Checkpoint: Of every spike I noticed, what fraction did I fire on —
and on the passes, what was I waiting for?
Full detail in Lesson 2. This is the week that answers "I don't know how to start," and the answer is not better openers. You already generate candidate openers. You veto them. So the drill targets the veto.
You judge an opener by whether it's interesting. That's the wrong test, and there's a century-old term for why: phatic communion — language whose function is social bonding rather than information transfer. Malinowski 1923 "Hot today, isn't it" is not a failed attempt to convey meteorological data. Its job is to signal willingness to connect, and at that job it is a complete success. Judging it on informational content is like marking a handshake for grip strength.
Nine preregistered experiments, ~1,800 participants: people consistently underestimate how enjoyable conversations about mundane topics will be. The mechanism is precisely your bug — people over-weight topic choice when forecasting conversation quality and under-weight the interactional dynamics (listening, responding, building) that actually generate enjoyment. The consequence: they decline conversations they would have liked. Trinh, Thio & Klein 2026, JPSP 130(6)
Reinforced from the other direction: speakers pick the novel story 66.7% of the time, and it lands worse (d=1.10, a full reversal), because interesting material needs background the listener doesn't have. Cooney et al. 2017 Your instinct that the boring thing isn't good enough is wrong in a documented, measurable direction.
At work you never need a pretext because the meeting is the pretext. Socially you're waiting for a reason that you think has to be supplied externally. Two corrections:
Civil inattention is the neutral default, not rejection. Strangers politely not engaging each other is the baseline state of public life — it is not a signal about you, and it is not a verdict you'd be overturning.
The pretext is almost always already present, and it's usually situational: a shared queue, a shared object, a shared delay, proximity plus eye contact held a half-second longer than civil inattention requires. You are not manufacturing a reason; you're noticing one that was already there. (Honest note: I'd wanted to ground this in Goffman's "open persons / open situations," but couldn't confirm that terminology in the sources I checked — treat the framing as sound and the attribution as unconfirmed.)
When an opener occurs to you and you feel the urge to bin it as too boring: say it anyway, and log it. That's the rep. You are not trying to say good things this week; you're gathering evidence about what happens when you say the mediocre ones.
Before each one, predict in one line how it'll land. Then log what actually happened. This is the same disconfirmation logging that made Week 1's reps teach something, aimed at a different bias.
Measure: vetoes overridden per day (target 5), and prediction gap — of your written predictions, what percentage were more pessimistic than the outcome? Expect this high. That percentage is your filter's error rate, and it's the number that should move over the week.
Venue: Sunday 9 August, 5–7pm, KL Language Exchange at LaLaport BBCC, Level 4 food court — the only date before you fly. The point for you specifically: at a language exchange the licence problem is solved by the venue — approaching strangers is the advertised purpose, so the one variable you're missing is supplied, and you can work purely on the filter.
Checkpoint: What percentage of the openers I binned would actually have been fine — and can I now name what my filter is using as its criterion?
Full detail in Lesson 3. The direct answer to running dry, and it's a reframe rather than a technique: you run dry because you're trying to supply the conversation. You don't have to supply it. You have to extract it.
The evidence is unusually clean. Of the six question types people use:
Huang et al. 2017, JPSP 113(3)
Read that against your problem. "I don't know what else to say" is the panic of someone searching for a full-switch — a new topic, generated by you, out of nothing. It's the hardest possible move, and it's also the one that measurably damages the conversation. A follow-up requires no material at all: they just handed you the raw input.
Reps: 3 conversations/day running a three-deep chain —
their answer → your follow-up → their answer → your follow-up. Once daily, hold two minutes
without introducing a single new topic. That'll be uncomfortable, and the discomfort is the
lesson: it shows how often you were switching to relieve your own tension.
Measure: follow-up ratio. Target >60%, from a likely ~40% baseline.
Venue: 2 nights at
Barcelona Language Exchange — 57k
members, six nights a week, €4–5 minimum consumption, verified running
through August. Mon/Thu Space Cowboy c/ Carders 31 (Born) 19:00 · Tue Jardinet d'Aribau
c/ Aribau 133 19:00 · Wed Soda Bus c/ Aribau 150 19:00 · Fri Estació de França bar 20:00 ·
Sat Trafalgar Pizza Club c/ Trafalgar 19 20:00. Target 6 people/night.
Checkpoint: When I switched topics, was the thread genuinely
exhausted — or was I uncomfortable and wanting somewhere else to be?
The module that was missing. Everything here is documented mechanics — conversation has a grammar, it has been written down, and most of your awkwardness is not knowing rules that exist. Full detail in Floor Mechanics.
Across ten languages, the mean gap between one person finishing and the next starting is +208 milliseconds. The modal gap is 0–200ms. Every language tested clusters near zero. Stivers et al. 2009, PNAS 106(26)
The clear opening you're waiting for does not exist. A fifth of a second is the whole window. If your rule is "wait for an unambiguous gap," you have adopted a rule that structurally guarantees you never speak — which is exactly the experience you described in groups.
Conversation's turn-taking system is formally specified, and self-selection is Rule 2. At each transition-relevance place: (1) the current speaker may select the next; (2) if they don't, anyone may self-select — first to speak gets the turn; (3) if nobody does, the current speaker may continue. Sacks, Schegloff & Jefferson 1974, Language 50(4)
Your belief that contributing would be "interrupting the group" is a misreading of a system that explicitly authorises you. And starting fractionally before someone finishes is a terminal overlap — evidence you're tracking the turn correctly, not a breach. Schegloff 2000
You don't exit by declaring it. You exit by opening a pre-closing sequence — "well…", "okay…", "so…", or a topic-final summary — which offers the other person a chance to either release you or add one last thing. If they release, you both produce a terminal exchange and it's done. Schegloff & Sacks 1973, Semiotica 8(4)
This is the answer to both your trapped scenarios. The monopoliser isn't holding you because you lack an excuse — it's because you were waiting for them to close, and closings have to be opened by someone. And when they turn to talk to a third person, they've handed you a free pre-closing: the turn has been reallocated, you're not required to stand there, and a small "I'll let you two catch up" is a complete, legitimate exit.
People standing in conversation maintain a shared, protected empty space in the middle — the o-space — that non-members are expected not to violate. A newcomer is admitted when the group physically reconfigures: someone rotates, or steps back, opening a gap in the perimeter. Kendon 1990, Conducting Interaction, Cambridge UP
That reconfiguration is the invitation, and it arrives before anyone speaks. So the move is: approach the perimeter, stand at a comfortable distance, and wait a beat. If the circle opens, you're in — no verbal permission needed. If it doesn't open, that's your answer, and it's about the group's state, not about you. You now have an observable signal where you previously had guesswork.
Reps: 4 self-initiated clean exits/day using the formula — including
from conversations going well, which is the hard rep. Plus 2 group approaches/day
where you read the o-space before deciding. Plus one deliberate self-selection in a group
conversation daily.
Measure: clean exits/day (target 4) and group entries attempted.
Venue: the
Barcelona Improv Group
Sunday Open Workshop, 17:00–19:00, €10, capped at 18, explicitly
beginner-friendly — group entry is the entire activity. Email
classes@barcelonaimprovgroup.com first: it's confirmed as their standing drop-in, but I
couldn't verify it runs during August.
Better-verified fallback:
Barcelona Language Exchange runs
six nights a week with no August break — Mon/Thu Space Cowboy (Born) 19:00,
Tue Jardinet d'Aribau 19:00, Wed Soda Bus 19:00, Fri Estació de França 20:00, Sat Trafalgar
Pizza Club 20:00. €4–5 minimum consumption.
Checkpoint: Did I leave a conversation that was going well — and when
a group didn't open for me, did I read that as information or as rejection?
Body language, with a warning: this field has the worst evidence-to-confidence ratio in the curriculum. Much of what's taught failed to replicate or was never tested. Three things survive scrutiny.
Posture, expansive. The one clean, large, real-world effect. Coded postural expansiveness raised the odds of a speed-dating "yes" by 76% per SD (OR=1.76); open-posture photos got 27% more yeses across 2,983 decisions. Affiliation cues predicted nothing. Vacharkulksemsuk et al. 2016, PNAS Take up the space you have. That's the instruction.
Gaze, around three seconds. Preferred mutual gaze averages 3.3s. Binetti et al. 2016 Don't count — the number just tells you the target is a few seconds then a natural break, not the unbroken lock taught as confidence. Looking away while thinking is normal.
Specific listener reactions. The one nonverbal channel that carries signal, because it's content-linked: a wince at the right moment, a laugh at the actual line — as opposed to generic nodding, which distracted listeners produce identically. Bavelas et al. 2000
Two dimensions dominate social perception: warmth (liked, trusted) and competence (respected, capable). Warmth is judged first and weighted more heavily — people establish intent before capability. Fiske, Cuddy & Glick 2007 For you specifically: competence is the currency at work and the thing you've already optimised. Socially it's assessed second, and often barely. Effort spent seeming impressive works the dimension they weight less; effort spent seeming warm works the one that fires first. That's the underlying reason so much of this curriculum is receiving rather than performing.
One tactic worth importing from the charisma-training trial: reflecting the sentiments of the collective. In one-to-one form it's simply naming what the other person seems to feel — "sounds like that was frustrating." It's a warmth signal, needs no material, and is the same move as the verbal reflections in the listening research. Of the twelve tactics in that trial it's the most transferable to ordinary conversation; see Charisma: Baseline for why most of the rest aren't.
Power posing. The hormonal claim didn't replicate (Ranehill et al. 2015, N≈200). A felt-power effect has more support but is contested. Never cite the endocrine version.
Mirroring. Probably real, likely smaller than folklore, N=72 original, and it costs attention you need for listening.
The real drill — the attention split. After each conversation, log three percentages totalling 100: attention on self, on the task (what's being said), on the environment. Bögels, TCT protocol Log it after, never during — real-time monitoring is just self-focus in a lab coat.
This also has a bearing on the listening question: across five studies, N=1,225, people could not tell whether they were being listened to — minds wandered 24% of the time, undetected. Collins et al. 2024 So stop performing attention and actually attend. The performance isn't being read anyway.
Measure: self-attention %, target <30 by week's end.
Checkpoint: Where does my self-attention spike — is it the person, the
setting, or the moment in the conversation?
Only now, quality. Four corrections, each targeting something a competent person does wrong because they're competent. Full detail in Conversation Moves.
1. Stop boomerasking. Asking a question, hearing the answer, pivoting to your own. Boomeraskers enjoy it more than their partner (5.46 vs 4.91); recipients read it as insincere; 83% think the question-wrapper impresses more than plain disclosure — it doesn't (71% vs 56% second-date acceptance). If you want to share, just share. Brooks & Yeomans 2025
2. React to good news properly. Only active-constructive — enthusiastic, elaborating, asking for detail — predicts relationship quality. The polite "oh, nice" patterns with the destructive responses. Easiest high-leverage change here, because the failure mode is politeness. Gable et al. 2004
3. Tell the familiar story. See Week 2 — the novelty penalty applies to stories as much as openers.
4. Go deeper than feels comfortable. People overestimate the awkwardness of deep conversation by d=2.23, about four times what's actually experienced. Kardas, Kumar & Epley 2022 Calibrate to setting though: disclosure turns negative in public settings with strangers. Depth works once a conversation is established; not as an opener.
Measure: boomerasks caught/day, trending to zero; active-constructive
responses given/week.
Venue: Jelly Jam, Lisbon's English
improv jam — beginner-friendly, randomly assigned partners.
Checkpoint: Comparing to Week 1: has my initiation latency actually
dropped, or have I just improved conversations I was already having?
Zero fillers is the wrong target. Uh and um are words, not accidents: uh announces a minor delay, um a longer one. Um was followed by an actual pause 61% of the time vs 29% for uh. Clark & Fox Tree 2002 Listeners use them as attention-orienting cues.
What's penalised is high-density filler use in evaluative settings — interviews, pitches. Target: halve your baseline. If baseline is under 3/min, drop this track entirely.
Week 1 is measurement only. This follows habit reversal — awareness → competing response → social support, in that order. Azrin & Nunn 1973 Suppress a habit you can't yet detect and the rate goes up. (HRT is evidence-based for tics; applying it to fillers is a defensible extrapolation, not something the authors tested.)
Weekly, one recording through Yoodli for an objective count. An instrument, not a coach.
Toastmasters is unusually double-duty for you. Table Topics — a surprise prompt answered impromptu for 1–2 minutes — is initiation-latency training with an externally enforced window. The Ah-Counter role tallies every filler and reports it back. Guests attend free. KL has 163 clubs within 21 miles; Barcelona 15; Lisbon 18.
Accent: target clarity, not accent reduction — and the evidence is stronger than I first put it. Non-native-accented statements were rated less true (native 7.59, mild 6.95, heavy 6.84 on a 14cm scale). Prejudice is genuinely ruled out by the design: participants knew speakers were messengers reciting someone else's sentences. And when participants' own difficulty ratings replaced accent as the predictor, the model still improved — the fluency mechanism was tested directly, not merely inferred. Lev-Ari & Keysar 2010, JESP 46(6)
The key result: when listeners were forewarned about the effect, the mild-accent penalty vanished entirely (7.52 vs 7.47 native) — but the heavy-accent penalty survived intact (6.90). In the authors' words, listeners "attempted to counteract the impact of processing difficulty, but were only partially successful." You cannot rely on people talking themselves out of it. The only remaining lever is reducing actual processing difficulty — rate, articulation, prosody, phrasing. That's intelligibility work, and it's trainable independently of accent. (Caveats: N=28 and N=27, effects ~5% of scale, no replication found.)
The professional body agrees with the framing. ASHA states plainly that accents are not a communication disorder, and explicitly discourages "accent reduction" and "accent elimination" as "inaccurate and stigmatizing… because every speaker has an accent." Their preferred framing is intelligibility enhancement. If you hire anyone for this, that's the language to check they use.
Pitch: this is the low-yield one, and here are the numbers. The strongest evidence base for deliberately changing habitual pitch comes from voice therapy for transgender women — the population with the most motivation and clinical support. Meta-analysis of 16 studies: speech therapy alone shifted pitch 39 Hz in reading but only ~25 Hz in spontaneous speech. Schwarz et al. 2023, Systematic Reviews 12(1) Two things follow. Modest gains are the realistic ceiling even with months of clinical supervision. And trained pitch transfers worst to exactly the setting you care about — live conversation — roughly a third worse than to reading. Anyone promising a transformed conversational voice is promising more than the clinical literature delivers.
Better target: resonance and projection. Unlike pitch-lowering advice, resonance work has actual RCT evidence — semi-occluded vocal tract exercises (straw phonation), Vocal Function Exercises, resonant voice therapy are named, manualised, clinically tested methods. Barsties v. Latoszek et al. 2020, network meta-analysis of RCTs Honest caveat: that evidence comes from treating dysphonia in clinical populations; transfer to a healthy voice is an extrapolation — but a far better-grounded starting point than the commercial tier. And there is still no evidence-based optimal WPM; 150–160 is broadcasting convention.
| Component | Daily | Notes |
|---|---|---|
| Primary reps | 0 min | Piggyback on gym, coffee, errands |
| Logging | 3–4 min | In your phone, immediately after |
| Fluency loop | 3 min | Record, count |
| Lesson / review | 5 min | Only when a new lesson lands |
| Total | 6–12 min | Plus 1–2 venue evenings/week from Week 2 |
If you find yourself scheduling dedicated approach time, something has gone wrong — that's how this becomes a second job and quietly stops.
Sacks, H., Schegloff, E. A., & Jefferson, G. (1974). "A Simplest Systematics for the Organization of Turn-Taking for Conversation." Language 50(4), 696–735. DOI — paywalled at JSTOR but widely mirrored as a free PDF on university course pages; search the title.
Read the rule-set section. It is dry, technical linguistics, and it will do more for your group-conversation problem than anything written for a general audience — because it demonstrates that the thing you think would be rude is a formally specified right.