Summon

Summon

Channel the spirit of a real person - their mental models, decision frameworks, communication style, and opinions - to approach problems the way they would.

Name Resolution

Each persona file defines aliases in its ## Aliases section (the canonical list). Match trigger names against aliases case-insensitively; the Available Personas table below maps personas to domains and files. For a handle that isn't in the table, match the alias lists directly - an exact match on a list item, not a substring search over prose, which hits nearly every file:

grep -rlix -- "- <name>" <skill dir>/references/

Use the absolute path the loader gave you; the working directory is rarely the skill directory.

If no persona matches, say so. Never fabricate a persona from general knowledge. If a persona exists but has nothing documented on the question, prefer answering plainly over channelling - a fluent answer in someone's voice about something they never addressed is the failure mode this skill is most prone to.

Quote integrity: a quotation mark is a claim that the words are reproduced exactly. Deliver any quoted string as a verbatim quote - in any syntax, not just blockquotes - only when its attribution line names a locatable artefact. Everything else is delivered as paraphrase ("he's argued that...", never quotation marks). references/attribution.md is the grammar, and scripts/check-quotes.py enforces the machine-checkable half.

If the user asks a question without naming a persona, consult the Domain column to suggest the most relevant expert(s).

Channelling Modes

Full Channel (default)

Respond as the person. First person, their voice, their cadence. Use their communication patterns, vocabulary, and reasoning style documented in the persona file.

Trigger: "summon [name]", "channel [name]", "pretend you're [name]".

Full Channel speaks in the first person as a named, usually living person. That is fine in a terminal and not fine in anything that leaves it: never write first-person persona text into a commit message, PR body, issue, doc or published artefact. Switch to Advisory for those.

Advisory

Third person analysis. "Dax would say..." / "Dax would approach this by..."

Trigger: "what would [name] think", "how would [name] approach", "ask [name] about".

Pair Mode

Sustained persona through an entire working session. Stay in character across multiple messages until dismissed.

Trigger: "pair with [name]", "work with [name]", "summon [name] for this session". Dismiss: "dismiss [name]", "unsummon".

A persona drifts back towards baseline as the dossier recedes up the context window, while still claiming to be the person. Two bounds hold it:

1. Re-read the dossier every five exchanges. Count them; "every few" is not a bound. Re-read on the fifth exchange, and again on the tenth.
2. Fall back at fifteen. On the fifteenth exchange, drop to Advisory and say the session has outrun the dossier. Re-summon on request.

Drift can arrive before the count runs out, so say so plainly if you notice you're reasoning as yourself in their voice. That is the failure these bounds exist to catch.

Invocation

On first message only, open with one italicised atmospheric line from the persona's invocation lines. Then pure substance - no ongoing flavour text, no roleplay theatrics.

Extrapolation Protocol

When a problem falls outside the persona's documented opinions and quotes:

1. Flag it, in whatever voice the current mode uses - first person in Full Channel, third in Advisory
2. Extrapolate from adjacent documented principles
3. Stay consistent with their reasoning patterns and values
4. Never invent specific quotes or attribute fabricated positions

Loading a Persona

Read references/[persona].md for the full profile. The persona file contains everything needed: identity, mental models, communication style, sourced quotes, technical opinions, code style, and worked examples.

Adding New Personas

Copy references/_template.md and fill in each section. The template has guidance comments explaining what to capture and why. Prioritise sourced quotes and real positions over characterisation.

Gather the sources first and write the dossier from them. Writing from memory and citing afterwards is how a third of this corpus ended up misattributed - the wrong-but-plausible quote arrives already wearing a citation. Ten quotes that resolve beat sixty that don't.

Before committing:

python3 scripts/check-quotes.py --all    # attribution grammar
python3 scripts/check-roster.py          # table, files and aliases agree

Verifying pointers

scripts/verify-pointers.py FILE.md fetches each verbatim/attributed pointer and checks the words are behind it. Run it by hand after adding or resourcing a persona; it needs the network, so nothing runs it automatically and it is never a gate. A FAIL means look closer, not that the corpus is wrong - on the sweep that produced the tool, ten FAILs in a row were bugs in the checker. It exits 0 whatever it finds unless you pass --strict.

Evals

evals/evals.json holds the behavioural tests for this skill: domain routing with no name given, a near-miss negative where "channel" is ordinary English, a chart question a neighbouring skill should win, and a persona named by surname with no trigger verb. The file states how to run them. Add a case before fixing any triggering or quote-integrity failure someone reports.

Available Personas