AEGIS · Module 4 · Deploy
Artifact 4.2
Prompt & Template Library
Curated, version-controlled prompts for sanctioned workflows — with metadata, caveats, and evaluation evidence behind every one.
- Client
- [CLIENT NAME]
- Engagement
- [ENGAGEMENT ID]
- Version
- v1.0
- Issued
- 2026-05-18
Delivered by TechFides under the AEGIS Governance Operating Services engagement. This document is proprietary to the client named above. Redistribution beyond the engagement steering committee requires written consent.
Purpose
Intent — Untracked prompts are the shadow-AI of the AI-governed enterprise. This library makes prompts first-class assets — owned, tested, and versioned like code.
Why this matters
A prompt is not a private preference — it is an operating instruction that affects output quality, data handling, and regulatory posture. If workflow WF-03 classifies support tickets, the prompt that does the classification is a compliance control, not a Slack copy-paste.
What this library replaces
- Slack-shared “this prompt works well” copy-pastes.
- Private docs where individual employees keep their favorite prompts.
- Vendor-supplied prompt suggestions that are not evaluated against client data.
- Undocumented prompt changes made inside workflow automations.
Library Composition
Intent — Snapshot of how the current library is distributed across functions. Rebalance if one function dominates — signals the library is being written for them, not the enterprise.
Legal
14
prompts
Sales
22
prompts
Customer
18
prompts
Engineering
31
prompts
Marketing
12
prompts
Finance
9
prompts
Ops
16
prompts
HR
5
prompts
Metadata Schema
Intent — Every prompt in the library carries this envelope. The prompt text is only one field of thirteen.
| Field | Description |
|---|---|
| Prompt ID | PR-NNN stable identifier referenced by workflows and SOPs. |
| Name | Short, descriptive, action-oriented. |
| Workflow ID | The Artifact 4.1 workflow this prompt belongs to, or 'Library' for general-use. |
| Owner | Named accountable individual — same rule as workflow owners. |
| Function | Legal · Sales · Engineering · Ops · Finance · Marketing · HR · Customer. |
| Model | Approved model + minimum acceptable version. Reviewed when vendor releases new versions. |
| Data Class | Highest data classification permitted — per Artifact 2.1. |
| Purpose | One-sentence description of what this prompt produces and for whom. |
| Inputs | Required and optional inputs with format — anything outside this shape is rejected. |
| Expected Output | Structure and content of what the model should return. Informs validation. |
| Caveats & Known Failure Modes | What this prompt is known NOT to do well — and when to fall back to a human path. |
| Evaluation | How we test it: gold-standard set, sample cadence, tolerance threshold. |
| Last Tested | Date of the most recent evaluation pass. |
Prompt Entries
Intent — Seven example entries covering the workflows in Artifact 4.1 plus one general-library prompt. Full library in client engagement typically runs to 80–150 entries.
Contract Clause Redline
LegalWorkflow · WF-01- Model
- Claude Opus · enterprise tier
- Data Class
- P1 · counterparty contract, non-privileged
You are reviewing a counterparty contract against the attached playbook. For each clause that deviates from the playbook, produce: (1) the clause text verbatim, (2) the deviation type (material / minor / acceptable), (3) a suggested rewrite citing the playbook rule by rule-ID, and (4) a risk note of one sentence. Do not produce output for clauses that match the playbook. Return results as a JSON array keyed by clause location.
RFP Response Drafting
SalesWorkflow · WF-02- Model
- ChatGPT Enterprise · gpt-5
- Data Class
- P2 · internal proposal library + RFP prompt
For each question in the attached RFP, retrieve the most relevant approved snippet from the proposal library using the passed context. Produce a draft answer that stitches the snippet into the client's language. If no snippet scores above 0.72 similarity, return "GAP — needs manual response" for that question. Never invent capabilities or SLAs. Cite snippet ID for every answer.
Support Ticket Classification
CustomerWorkflow · WF-03- Model
- Claude Haiku · enterprise tier
- Data Class
- P2 · customer ticket content, PII scrubbed at ingress
Classify the attached support ticket into one of the 14 taxonomy categories and assign a priority (P1–P4) per the priority rubric. Output JSON with keys: category, priority, confidence, signals. If confidence < 0.6, set priority to P2 and mark for human triage.
PR Description Generation
EngineeringWorkflow · WF-04- Model
- GitHub Copilot · Business
- Data Class
- P2 · code and commit log
Given the diff and commit messages for a pull request, produce a PR description with sections: Summary (1–3 bullets), Test plan (checklist), and Risks (bullets if any). Do not fabricate test results. Do not include TODOs. Do not exceed 200 words.
Meeting Summary
OpsWorkflow · WF-05- Model
- Claude Sonnet · enterprise tier
- Data Class
- P2 · meeting transcript, non-privileged
Produce a summary of the attached meeting transcript with sections: Decisions made, Action items (attributed to named attendees with due dates where stated), Open questions. Capture direct quotes only when the attendee is explicitly named. Do not infer sentiment. Output in markdown.
Pricing Draft Rationale
FinanceWorkflow · WF-06- Model
- Claude Opus · enterprise tier (tenant-isolated)
- Data Class
- P1 · deal attributes + pricing rules
Given the attached opportunity attributes and pricing ruleset, produce a draft price sheet and a rationale section explaining which rules applied, the resulting band, and any ambiguity. Do not suggest prices outside the computed band. If a required attribute is missing, return "PENDING — missing: <attribute list>" and stop.
Customer Research Brief
MarketingWorkflow · Library- Model
- Perplexity Enterprise
- Data Class
- P2 · public web + internal account dossier
For the named account, produce a one-page research brief with sections: Company snapshot, Recent news (last 90 days, cited), Technology footprint, Likely objections, Suggested talk tracks. Cite every external claim with a link. Do not include contact data or personal information about employees.
Evaluation
Intent — Prompts are tested like code. An untested prompt in a governed workflow is a latent incident.
Gold-standard sets
Every prompt that drives a governed workflow has a held-out gold-standard evaluation set — real-but-scrubbed inputs with correct outputs. Typical size: 30–100 cases per prompt. Sets are owned by the prompt owner and versioned.
Evaluation cadence
- Quarterly— every prompt run against its gold set. Accuracy regression > 5% vs. baseline triggers rework.
- On model change — vendor version bump or tier migration forces a full re-evaluation before the new model is enabled.
- On incident — any workflow incident involving a prompt triggers immediate re-evaluation plus prompt revision or retirement.
Red-team prompts
A portion of each gold set is adversarial — inputs designed to probe failure modes (ambiguity, edge cases, prompt injection, policy-violating requests). Red-team performance is tracked separately and reported to the governance committee.
Change Control
Intent — Prompts change. So do model versions. Version control is the mechanism that lets you roll back when a change degrades the output.
- Prompts live in a Git-backed store with clear ownership per prompt.
- Every change is a PR reviewed by Owner + one peer from the same function.
- Evaluation results must be attached to any change PR — no change merges without a before/after score.
- Rollback is the default on any regression — fix forward only after root cause is understood.
- Retirement is flagged 30 days ahead; workflows and SOPs referencing the prompt are updated in the same release.
Regulatory Mapping
Intent — Prompt libraries aren't named in most regulations, but the evidentiary role they play in model risk, bias testing, and incident RCA is explicit.