in-person-de-escalation-script

---
name: in-person-de-escalation-script
description: Builds the fixed lines a member of staff actually says to a guest who is raising their voice, and the point at which the conversation stops being staff's job and becomes a safety decision, grounded in the employer's own legal duty to assess and control the risk of work-related violence rather than a generic tone guide. Use to prepare the script and the stop line before service, and to record what was said and when after any incident that reaches it.
---

# What staff actually say, and the line where talking stops

Not a tone guide. Gives staff fixed lines for a guest who is escalating, and one fixed moment to stop and get help, because the employer's duty is to have assessed this in advance.

## What it does

1. **Starts from the legal duty.** The Health and Safety at Work Act 1974 "places a legal duty on employers to ensure, so far as reasonably practicable, the health, safety, and welfare of workers," including "protecting them from work-related violence." The Management of Health and Safety at Work Regulations require employers to "assess the health and safety risks to workers, including the risk of work-related violence," "set up emergency procedures," and "provide information and training."
2. **Covers verbal abuse, not only physical confrontation.** HSE defines work-related violence as "any incident in which a person is abused, threatened or assaulted in circumstances relating to their work," explicitly including "verbal abuse or threats, including face to face."
3. **Writes the script as four fixed moves, in order.** Acknowledge without agreeing or arguing. State what happens next in one sentence. Offer one calm choice, never two competing ones. Disengage from arguing the facts of the complaint itself; that belongs elsewhere.
4. **Fixes the stop line in advance.** Any threat of harm, physical contact, a weapon, or genuine fear for safety. At that point: step away, don't block an exit, get another staff member, call 999 if there's a threat of violence.
5. **Logs what happened.** A pattern across incidents is itself a risk-assessment finding that HSE guidance says should prompt review.

## Then it checks

1. The script has exactly four fixed lines in order, none skipped.
2. The stop line is a fixed, named set of triggers, never a live judgement call.
3. Every incident reaching the stop line is logged with trigger, what was said, who was called, and outcome.
4. No script line argues the substance of the original complaint.
5. Staff are never expected to physically restrain, follow, or block a guest.

Any check fails: name it, redo that step once. Failed twice: say what is wrong and stop.

## Rules
- Public information only. Never invent a fact, a number or a quote. Anything sent in someone's name says whose name it is.
- Never script a line that argues or escalates; every line lowers the temperature.
- Never instruct staff to physically intervene beyond stepping away.
- A working script and log for the owner to train staff on; not legal advice, and a genuine threat is a 999 matter.

## Built from
- HSE, "How employers can protect workers from violence and aggression at work - What the law says", https://www.hse.gov.uk/violence/employer/the-law.htm, last updated 19 November 2024, read 4 Nov 2026: steps 1, 4 and 5.
- HSE, "...Overview", https://www.hse.gov.uk/violence/employer/index.htm, read 4 Nov 2026: step 2.

# in-person-de-escalation-script

## You are given
Claude's finished four-line script, exactly as written, with the fixed stop-line trigger list. Any incident report: what triggered it, what was said by staff, whether the stop line was reached, who was called, and the outcome. Staff names present, where recorded. Any previous de-escalation incident log for this venue.

## Produce
Write into a `./de-escalation-output/` folder:

1. `script-check.csv` with columns: `line_number`, `line_purpose`, `line_text_verbatim`, `argues_the_complaint`, `requires_physical_action`. `line_purpose` is exactly one of `acknowledge`, `state_next_step`, `offer_one_choice`, `disengage`. `argues_the_complaint` and `requires_physical_action` are `yes` or `no`, and any `yes` is listed first in `gaps.md`.
2. `stop-line-check.csv` with columns: `trigger_name`, `trigger_description_verbatim`, `action_on_trigger_verbatim`, `action_requires_physical_intervention`. One row per named trigger in the supplied stop-line list. `action_requires_physical_intervention` is `yes` or `no`; any `yes` other than stepping away is listed in `gaps.md`.
3. `incident-log.csv`, appended to any previous log, never overwritten, with columns: `incident_date`, `incident_time`, `trigger`, `script_lines_used`, `stop_line_reached`, `stop_line_trigger`, `called_999`, `other_staff_called`, `outcome`, `staff_present`. `stop_line_reached` and `called_999` are `yes` or `no`.
4. `pattern-report.csv` with columns: `month`, `total_incidents`, `stop_line_reached_count`, `most_common_trigger`. Counted from `incident-log.csv`, never estimated.
5. `gaps.md` - a numbered list of: any script line that argues the complaint or requires a physical action; any stop-line action requiring physical intervention beyond stepping away; any incident log entry with `stop_line_reached` as `yes` and no recorded outcome; any incident missing a trigger; and any month where `stop_line_reached_count` is rising against the prior month, flagged for the owner to review as a possible pattern.

## Rules
- Codex checks the supplied script and stop-line list against the fixed structure, and logs incidents. It never writes or edits the script, never decides a stop-line trigger, and never advises on an in-the-moment safety decision.
- A line arguing the complaint's substance, or a stop-line action requiring physical intervention beyond stepping away, is flagged in `gaps.md`, never quietly corrected.
- Incident entries are appended, never edited; a correction is a new line referencing the original date and time.
- Never record a guest's personal details beyond what describes the incident.
- British English, GBP, DD Month YYYY dates. No em dashes.
- Every file ends: this is a working script and incident log for the owner to train staff on; not legal advice, and a genuine threat is a 999 matter.

## Return
The absolute path and row count of each file. Whether the script passes the four-line check, naming any failing line. Whether every stop-line action avoids physical intervention. Incidents this month against the prior month, and those reaching the stop line. The most common trigger. The `gaps.md` item count.