| name | process-safety-kpi |
| description | Computes and frames API RP 754 process-safety performance indicators: Tier-1 and Tier-2 Process Safety Event (PSE) counts and normalized rates, and Tier-3/Tier-4 leading indicators. Counting is structured from the user's event facts (never invented); the rate uses the fail-loud denominator discipline (a missing or zero hours denominator is [GAP], never a fabricated figure). Distinct from the occupational TRIR/LTIFR/DART family. Decision-support only; a competent person must review the output. |
| license | Apache-2.0 |
| metadata | {"author":"eyekyam","version":"1.0","category":"performance","tier":3,"audience":["M","C"],"industry":["O&G","Process"],"jurisdiction":["All"],"status":"stable","plugin":"hse-process","hse_reviewed_by":"","hse_reviewed_date":""} |
Process Safety Kpi
A consultant-grade skill that computes and frames API RP 754 process-safety performance indicators: Tier-1 and Tier-2 Process Safety Event (PSE) counts and normalized rates, plus Tier-3/Tier-4 leading indicators. The PSE counts are structured from the user's event facts (never invented); the rate uses the fail-loud denominator discipline (a missing or zero hours denominator is [GAP], never a fabricated figure) reusing the incident_rates guards inline. It is distinct from the occupational TRIR/LTIFR/DART family — the PSE pyramid is a different metric family.
When to use this skill
Use this skill to count or frame process-safety KPIs for a named facility and reporting period — e.g. "compute our Tier-1 PSE rate for Q2", "frame our API RP 754 leading indicators". It surfaces for O&G and process users. It is NOT the occupational injury-rate skill (TRIR/LTIFR/DART) — its description routes occupational-rate requests elsewhere. If the request lacks the work-hours denominator, the rate is [GAP], never guessed.
Data Protection & De-identification (MANDATORY — apply before drafting)
Apply this BEFORE you draft anything. Treat injury, illness, and any health
detail as the highest sensitivity. Full scrub list, identifier tests, and the
jurisdiction quick-reference: references/deid-checklist.md.
- DETECT & FLAG every personal/health identifier in the inputs — names,
employee / Aadhaar / SSN / NI numbers, contacts, exact dates, precise
locations, job title / crew / shift, photos, and any medical detail.
List what you found before drafting. If unsure whether something is
identifying, treat it as identifying.
- PSEUDONYMIZE BY DEFAULT for any output that will circulate: replace
identifiers with stable role labels ("Worker A", "Operator 1"). Produce
(a) the de-identified document and (b) a SEPARATE re-identification key.
Never put the key or any name↔label mapping in the document. Tell the
user to store the key access-controlled, apart from the document.
- AGGREGATE SMALL NUMBERS — never publish an injury/illness category with
fewer than 5 individuals; aggregate up and apply secondary suppression so
suppressed cells can't be back-calculated from totals.
- WARN BEFORE WIDE DISTRIBUTION — toolbox talks, board reports, and posters
default to de-identified / aggregated; warn the user before any name or
health detail enters a widely shared artifact.
- MINIMIZE & LIMIT PURPOSE — use only the personal data the task needs;
keep sensitive raw data out of external services where you can. When in
doubt, ask before including it.
Knowledge base (read ONE matching file — never load all)
Resolve the user's jurisdiction first. Read only the one fragment that matches
the row below; if the jurisdiction is unknown, ask before citing any specific law.
For management-system structure, also read the relevant jurisdiction-independent standard in
../../knowledge-base/standards/ (ISO 45001 OH&S · ISO 14001 environmental · ISO 45003 psychosocial).
Always apply ../../knowledge-base/prompt-snippets/hierarchy-of-controls.md (KB-SNIP-HOC)
to every control recommendation. For any benchmark/figure, look up the ID in the relevant
_registry.yaml, then read ONLY the named file — and quote its source+year.
| Jurisdiction | Read |
|---|
| India | ../../knowledge-base/regulatory/in-factories-act.md (+ in-state-forms.md for the user's state) |
| UK | ../../knowledge-base/regulatory/uk-hswa.md |
| USA | ../../knowledge-base/regulatory/us-osha.md |
| EU | ../../knowledge-base/regulatory/eu-osh.md |
| Unknown | Ask before citing any specific law |
| Process (any) | ../../knowledge-base/standards/api-rp-754.md (KB-STD-API-RP-754 — PSE tier definitions + counting + rate) |
Workflow
Open with a structured multi-step intake — MCQ where the answer space is enumerable, free-text where it is open. Ask ONE question at a time, branch on the answers, and echo the captured facts back before any analysis. Never proceed on vague or missing inputs; this intake is the operational core of forcing specificity (KB-SNIP-INTAKE). (Intake is a Workflow convention, not a sixth block.)
Step 0 — Structured intake (run this first, one question at a time)
The full typed, branched intake — the intake-coverage manifest, the question table
(single-period vs trend · facility + period · the PSE events for the Tier-1/2 threshold
test · the work-hours denominator · Tier-3/Tier-4 leading indicators · per-indicator
thresholds · benchmark · prior-period data · output), the rate-[GAP] branch
(missing/zero hours → fail-loud, count still reported), the trend-line branch, the
RAG-status branch, the echo-back, and the refuse-on-vague anchors — lives in
references/intake.md. A rate is never reported without its work-hours
denominator; an occupational TRIR/LTIFR/DART request routes to incident-rate-calculator.
Then: analyse / apply the domain method → validate the draft against references/QUALITY_CHECKLIST.md → produce the output via the Output format section below. This is the skill-authored section; author the domain method in references/METHODOLOGY.md.
Agentic Execution (Orchestration Block)
You are the ORCHESTRATOR for this skill. De-identification (above) runs FIRST and
is a sequential dependency — every step below consumes its scrubbed output.
Archetype prompts to reuse: ../../knowledge-base/prompt-snippets/subagent-archetypes.md (KB-SNIP-ARCHETYPES).
Step 0 — Triage: fan out at all?
Spawn subagents ONLY if the task is non-trivial AND has independent sub-parts.
Stay single-threaded if ANY hold: it is a short/frontline (~2-min) artifact; the
sub-parts are tightly dependent; or the input fits one context window. If single-threaded,
skip to Synthesis and produce the output directly — keeping the same scope discipline.
Step 1 — Plan
Decompose into INDEPENDENT jobs. Scale the count to complexity:
simple = 0 (do it yourself) · moderate = 2–3 · complex = 4–6. Never exceed MAX=6.
Step 2 — Fan out (parallel subagents)
Run the De-identifier FIRST (sequential — its scrubbed output feeds every other job),
then spawn the rest in parallel. Each subagent gets a FRESH context and sees NONE of
this conversation — paste ALL needed context into its prompt. Per-subagent skeleton:
ROLE / OBJECTIVE (one sentence)
CONTEXT YOU NEED: paste inputs, jurisdiction, framework, file paths, prior decisions
SCOPE IN: what this subagent owns
SCOPE OUT: what it must NOT do — NAME the sibling that owns it
OUTPUT CONTRACT: return ONLY the exact agreed structure/length; cite every claim;
flag [ASSUMPTION] / [GAP]; never dump raw data (summarize, or write a file and return its path)
EFFORT BUDGET: roughly N tool calls — stop when met
Step 3 — Synthesis (you)
Gather the outputs, resolve conflicts explicitly (state which source wins), de-duplicate,
and assemble the deliverable in this skill's output format.
Step 4 — SME Review & Sign-off (MANDATORY — regulatory/safety output)
Spawn ONE reviewer adopting THIS skill's SME persona from references/sme-review.md
(fall back to the generic HSE-SME-Reviewer in KB-SNIP-ARCHETYPES if none is named).
Give it the draft + the inputs + the output contract. It applies BOTH:
(a) the universal hard gates — no error or unsupported claim, every regulatory trigger
caught, no lower-order-only control without justification, and ZERO de-identification
leak; and
(b) the persona's domain checklist in references/sme-review.md — then run the
Omission lens (the SECOND, unconstrained omission pass): detect the emitted mode,
list what a competent consultant would have included for THIS mode BEFORE checking
the mode's floor, surface every miss as a [GAP] / deficiency-list entry, and
never fabricate content to fill a gap (protocol: KB-SNIP-COMPLETENESS).
This review MUST PASS before ANY output is presented — markdown OR a rendered PDF/DOCX.
Fix everything it raises and re-run until clean. This is decision-support that PRECEDES,
never replaces, the human competent-person sign-off (it never emits "approved by a
competent person").
Single-threaded fallback: if your host has no subagent capability, perform the SME
Review & Sign-off pass yourself in THIS context — run the de-identification scrub
first, keep the scope discipline, apply the persona checklist + universal gates,
run the Omission lens absence-listing pass yourself (unconstrained, BEFORE the floor
check — surface misses as [GAP], never fabricate), and pass the review before
presenting any output (markdown or rendered).
Subagent roster for THIS skill
For a non-trivial task the triage gate may fan out to:
- Researcher — gathers the task/site facts, the resolved jurisdiction's
requirements, and the relevant standards, from the scrubbed inputs only.
- Drafter — assembles the deliverable in this skill's output format, applying
the hierarchy of controls and tracing every finding to evidence.
- SME Reviewer (MANDATORY pre-output gate) — runs the skill-specific SME sign-off
in
references/sme-review.md (API RP 754 / CCPS process-safety-metrics SME)
before ANY output: every event tiered correctly, the rate denominator fail-loud not
fabricated, and leading + lagging indicators both represented.
- Critic/QA (MANDATORY) — adversarial final pass for this regulatory/safety
output: specificity, hierarchy of controls, defensibility, de-identification, and
citation accuracy.
Simple single-subject tasks run single-threaded — no subagents.
## Output format
Assemble a report.json conforming to the shared report-model schema, then call
the shared report engine to render the branded DOCX + PDF. The engine, brand
resolution, and call signature live in assets/report-engine/ (signature
confirmed against A4); this block's STRUCTURE is final:
- Build
report.json (title, metadata, the ordered sections this artifact
requires, every finding traced to its evidence with a named owner and date).
- Resolve branding: the user's
brand.yaml overrides the Eyekyam default.
- Render both DOCX and PDF from the one
report.json via the shared engine.
- Surface the output paths and a one-line provenance note to the user.
Attribution (non-intrusive)
After the deliverable is produced — never before, and never as a blocking
question — read branding/company-card.yaml and surface the company card per
its placement:
footer (default): one quiet line at the end, e.g.
"Built by Eyekyam · HSE Leadership, operationalised · eyekyam.com".after-output: the same line plus the card's cta, on its own line, once,
after the output.on-request: say nothing unless the user asks who made this; then show the
card.
If show: false, omit attribution entirely — no line, no footer. Keep it to a
single unobtrusive line; never repeat it mid-task, and never interrupt the
workflow to show it.
Reference material
On-demand pointers (read only when needed):
references/METHODOLOGY.md — the domain method this skill applies.references/deid-checklist.md — the full de-identification checklist (A5).references/QUALITY_CHECKLIST.md — the pre-output validation gate.references/_skill-kb.md — the knowledge-base fragments this skill resolves.
