Audience Classifier

Audience Classifier — Who Is This For?

A problem statement that doesn't fit its audience has failed regardless of its other qualities. Statement-grader's audience-fit axis depends on the audience being correctly identified upstream — that's this skill's job. The audience tag drives the entire compression subdomain's routing.

For the canonical mapping of audience → compression skill, see [[minto-scqa]] and the compression skills directly.

The Five Audience Tags

| Tag | Who | Channel typical | Time budget | Routes to | |---|---|---|---|---| | LLM | A language model that will execute the resulting prompt | API call, in-context | Model-consumed (~10K tokens fine) | cursed-speech | | Exec | Decision-maker, time-pressured, trust intact | Slack, email, brief, deck | 30 sec – 5 min | bluf-shaper (short) or executive-distiller (memo) | | Peer | Collaborator with parallel context | Doc, email, meeting | 2–15 min | scqa-formatter | | Self | The user thinking through their own problem | Note, scratch buffer | Whatever's needed | scqa-formatter (skip Q step) or Polya restatement | | Public | Mixed-audience reader who may be skeptical | Essay, post, talk, public memo | 5 min – 30 min | executive-distiller (with Concern-priming) or scqa-formatter (Concern-priming variant) |

Detection Signals

Read the statement plus any context the user provided (channel, recipient, purpose). Apply these:

| Signal | Likely audience | |---|---| | Output is described as "a prompt", "ask Claude", "send to GPT/Claude" | LLM | | Recipient is named as a title role (CEO, CMO, board, VP) | Exec | | Phrase like "memo for the team", "share with engineering", "draft for the working group" | Peer | | Phrase like "I'm thinking through", "trying to figure out", "for my own clarity" (no specific recipient) | Self | | Phrase like "publish", "post", "essay", "talk", "newsletter", "public statement" | Public | | Recipient is a colleague at peer level with similar context | Peer | | Reader is described as time-pressured or "needs the answer fast" | Exec | | Reader is described as skeptical, opposed, or requiring persuasion | Public (Concern-priming variant) | | Output is described as "for me", "for myself", "to remember" | Self |

If multiple audiences are plausible (e.g., "I want to think this through AND share with the team"), default to the more constrained audience — sharing imposes more structure than self-thinking, so route Peer.

Sub-Context Tags

Beyond the primary audience tag, surface sub-context that downstream compression skills can use:

| Sub-context | Values | Used by | |---|---|---| | Time budget | ≤30s / 30s–5min / 5–30min / open | bluf-shaper vs executive-distiller choice | | Skepticism level | trusting / neutral / skeptical | SCQA variant selection (standard vs Concern-priming) | | Channel | slack / email / doc / deck / verbal / API | Length and formatting bounds | | Familiarity | high context / partial context / cold | Background-section depth |

These don't change the primary tag; they tune the compression skill's behavior.

Process

1. Scan for explicit signals in the statement and any context (channel, recipient, purpose).
2. Apply the detection signal table — count matches per audience tag.
3. Resolve conflicts. If multiple tags get matches, apply the "more constrained wins" rule. If genuinely tied, ask the user (this is the right time for L4 user-assist).
4. Tag sub-context — time budget, skepticism, channel, familiarity.
5. Return primary tag + sub-context.

Output Format

AUDIENCE — [first 60 chars of statement...]
─────────────────────────────────────────────
Primary tag:    [LLM | exec | peer | self | public]
Confidence:     [high | medium | low]
Sub-context:
  Time budget:    [≤30s | 30s–5min | 5–30min | open]
  Skepticism:     [trusting | neutral | skeptical]
  Channel:        [slack | email | doc | deck | verbal | API]
  Familiarity:    [high | partial | cold]
Routes to:      [recommended compression skill, given the primary tag + sub-context]
Rationale:      [which signals matched]

Edge Cases

| Pattern | Handling | |---|---| | Statement specifies two audiences explicitly (e.g., "for the CEO and the LLM that will analyze it") | Run compression twice. Tag both audiences; orchestrator's deep mode will produce both shapes. | | Statement has no explicit audience signals | Default to self (the user is thinking, no recipient specified) with low confidence. Recommend the user specify if a different audience is intended. | | Audience is a class of person, not an individual ("for engineering managers") | Treat as peer (collective collaborator). Skepticism level depends on the topic. | | Audience is an LLM but the user wrote it as if for a human | Tag LLM but flag the prose-shaped framing — cursed-speech will need to restructure for the canonical order. | | Audience is "everyone" or "the world" | Tag public; flag low confidence; recommend narrowing scope. Public-as-mass-audience usually conceals a more specific intended reader. |

Output Contract for six-eyes

Called from Phase 2 (Diagnose), in parallel with problem-typology and stakes-assessor. Returns:

• Primary tag
• Sub-context tags
• Recommended compression skill (for Phase 5 routing)

six-eyes Phase 5 uses the primary tag to pick the compression skill; the sub-context tunes its behavior. If primary tag is LLM, also signal that cursed-speech will need the audited statement plus the downstream task description.

Connections

• bluf-shaper, scqa-formatter, executive-distiller, cursed-speech (binding-vow) — downstream consumers of the audience tag
• problem-typology (binding-vow) — runs in parallel; some typologies (adaptive, wicked) suggest specific audience patterns
• stakes-assessor (binding-vow) — runs in parallel; high stakes often imply exec audience or public
• minto-scqa (binding-vow) — canonical audience-fit decision rules for SCQA variant selection
• statement-grader (binding-vow) — uses the audience tag to score the audience-fit axis

Sources

• Minto, B. (1987, 2003). The Pyramid Principle. (Audience-driven structure decisions.)
• See [[minto-scqa]] for the canonical audience → structure mapping.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: npbuilds
• Source: npbuilds/skill-library
• License: MIT
• Homepage: https://skill-library-prod.web.app/

Install and usage instructions live in the source repository linked above.