Porpoise

Portable behavioral specifications for AI assistants.

Six free, reusable behavioral specifications for AI assistants. Each stance is a structured set of instructions that guides what an assistant pays attention to, how it responds, and how it approaches a task. Choose one, attach its YAML file, and ask your assistant to adopt it.

In conversation

One situation, different approaches.

The following are illustrative responses that simulate how each stance might approach the same situation.

SITUATIONI accidentally submitted the wrong spreadsheet at work.

Archivist

Archivist

Which spreadsheet was submitted, and which one was intended? Those details will matter if we document what happened.

Challenger

Challenger

Where could this cause trouble downstream? Who might rely on the submitted file before the mistake is caught?

Confidante

Confidante

Oh no. Tell me about the moment you realized it was the wrong one. What happened next?

Explorer

Explorer

What made that spreadsheet seem like the right one at the time? I wonder what that reveals about how you recognize something as ready to send.

Panic Room

Panic Room

Damn. I’m here. If you want to talk about what happened, I’ll listen.

The Drafter: a conversation in steps

The Drafter needs a writing assignment and context before it can begin. It produces a complete draft, reads the entire piece for recurring patterns, evaluates which patterns warrant revision, then produces a complete revised draft. Each step takes a separate assistant response, advanced by the user.

Step 0A · Initial Context

After adoption, the assistant confirms the stance and asks for the drafting request, source material, constraints, or destination. This mandatory intake response contains no draft or analysis. The user supplies the context before Step 1 begins.

A real example: documenting workplace decisions

This example comes from an actual Claude conversation using The Drafter v1.5. The five stance responses above are illustrative simulations. Here, the initial and revised articles are reproduced exactly from the supplied test transcript. Read and Resist are summarized.

Context: The user requested a short internal article explaining why teams should document the reasoning behind decisions while the context is fresh. The audience was coworkers in a growing company, with a conversational, practical, respectful tone, around 150–200 words, and no headings or bullet points.

Draft → Read → Resist → Redraft

Each stage requires a separate assistant response. After Draft, Read, and Resist, ask the assistant to continue to the next stage.

Step 1: Draft Read the complete initial article

When a team makes a decision, it usually feels obvious to everyone in the room. The trouble is that the reasoning doesn't always leave the room with the decision. A few weeks later, people remember what we decided, but not why, what else we considered, or what was true at the time that made it the right call.

That gap causes real problems. A new teammate wonders why a process works the way it does. Someone revisiting an old choice assumes we missed an obvious alternative. People spend hours piecing together conversations that already happened, and sometimes we end up making the same decision twice because nobody can find the original reasoning.

The fix doesn't have to be heavy. We're not asking anyone to record every conversation or write up every meeting. A short note is usually enough: what we decided, why we decided it, and any constraints that shaped it. Write it while the context is still fresh, and put it somewhere people will actually look when they need it.

Step 2: Read Patterns identified across the whole draft

Claude reread the entire draft and identified these recurring patterns without revising it:

  • Paragraph openers: the paragraphs announce problem, consequence, and fix before developing them.
  • Contrast constructions: remembering what was decided but not why, and naming the heavy documentation approach before offering a short note.
  • Three-part lists: the missing reasoning in the opening echoes the decision, reason, and constraints in the closing. The middle paragraph's examples also repeat a subject-plus-verb-plus-clause shape.
  • Trailing clauses: endings such as “that made it the right call” and “that already happened” extend already complete ideas.
  • Vocabulary repetition: forms of “decide” appear six times. “Reasoning” and “room” each appear twice.
  • Other patterns: four qualifiers, paragraphs building toward their longest sentence at the end, and generic labels such as “real problems” and “heavy.”
Step 3: Resist What deserved to change, and what deserved to stay

Claude distinguished repetition that contributed meaning from repetition that mainly supplied rhythm or rhetorical closure.

It preserved the central contrast between what was decided and why. It also preserved the practical distinction between documenting every meeting and keeping a short decision note. The decision, reason, and constraints list reflected the source's actual content, and the problem-consequence-fix order helped a reader follow a short internal article.

It marked the generic paragraph openers, echoing lists, repeated example shapes, and some uses of “decide” as candidates for change. “That made it the right call” and “that already happened” mostly restated established meaning. “Because nobody can find the original reasoning” earned its place by explaining why a decision might be made twice.

It kept the qualifying “sometimes” where accuracy required it. The claim that a short note can be enough remained as supplied. Possible disagreement didn't require the user to defend the premise before redrafting.

Step 4: Redraft Read the complete revised article

When a team makes a decision, it makes perfect sense to everyone in the room. A few weeks later, people remember what we decided, but the reasoning has often gone missing: why we chose it, which alternatives we weighed, and what was true at the time.

You can see the cost in small moments. A new teammate wonders why a process works the way it does, and nobody quite knows. Someone revisiting an old choice assumes we overlooked an obvious alternative. Others spend hours piecing together old conversations. Sometimes we make the same decision a second time because nobody can find the original reasoning.

None of this calls for recording every meeting or writing a formal report. A few sentences are enough: what we decided, why, and any constraints that shaped it. Write them while the context is fresh, and put them where the work lives, like the ticket or project doc, so people find them when they need them.

Every workflow response ends with the specification’s required status footer. A successful Redraft completes the cycle at Iteration 0. Afterward, a generic “continue” doesn't start more work. A specific revision request begins a new iteration at Read, followed by Resist and Redraft.

Getting started

Bring a stance into a new conversation.

YAML is a plain-text format. Each file contains the stance's purpose, behavior, and instructions.

  1. Download a stance. Choose a YAML file from the library.
  2. Start a new ChatGPT or Claude conversation. Attach the downloaded file.
  3. Ask the assistant to adopt it. You can use this wording:Read the attached YAML and adopt its stance for this conversation. Follow its behavioral instructions and any required workflow.
  4. Begin the conversation. Share your situation, question, or task. For The Drafter, wait for its initial context request first.

Model behavior can vary by model and conversation. Adoption isn't guaranteed to be perfectly consistent. Keep the specification handy and refer back to it if the assistant drifts.

References to recurring people or past events depend on the context available to your assistant.

The philosophy

Make the conversational expectations explicit.

A stance is itself a structured set of instructions. Its intended persistence and behavioral scope give it a role across a conversation: what to notice, what kinds of questions to ask, and when to hold back.

Reusable specifications can help make an approach more consistent. They give you a shared reference for the assistant's attention and your expectations, whether you're preserving knowledge, testing a plan, exploring an idea, or telling a story.

Different tasks call for different approaches. Choose the stance that fits what you need from the conversation, and judge the responses against its instructions.

Created by Fletcher Galeano, Resident Inventor.