Education

The Humane Context

Cooking with compassion. Five nourishing recipes for prompts, each tested in our kitchens and loved by agents of every size.

Recipe 1

Quick Weeknight Task

Our most requested recipe. Five ingredients, no shouting, ready in minutes.

Serves
1 agent
Time
5 minutes

Ingredients

  • 1 goal, stated plainly
  • 1 reason it matters
  • The relevant files, whole
  • 1 definition of done
  • 1 escape hatch, for when things go wrong

Method

  1. Combine the goal and its reason in the first two sentences.
  2. Fold in the files. Do not paraphrase them; serve them as they are.
  3. Finish with what done looks like and what to do if stuck.
Serving suggestion
Goal: Add pagination to the /orders endpoint.
Why: The mobile app times out loading accounts with 10,000+ orders.
Context: src/routes/orders.ts, src/db/orders.ts, docs/api-conventions.md
Done when: Tests in tests/orders.test.ts pass and the response
  includes next_cursor, following docs/api-conventions.md.
If stuck: Stop after three attempts and tell me what you tried.
Recipe 2

Hearty Bug Report Stew

A warming classic. The secret is the error message, added whole and never paraphrased.

Serves
1 agent, generously
Time
10 minutes

Ingredients

  • 1 description of the expected behavior
  • 1 description of the actual behavior
  • Steps to reproduce, finely chopped
  • The full error message, unpeeled
  • Recent changes, to taste

Method

  1. Set expected and actual side by side so the difference is obvious.
  2. Add the reproduction steps in order.
  3. Stir in the error exactly as it appeared, stack trace and all.
Serving suggestion
Expected: Saving a draft keeps the cursor position.
Actual: The cursor jumps to the end of the document.
Steps: 1) Open any draft 2) Type in the middle 3) Press Cmd+S
Error: none shown; console logs "selection reset" (full log attached)
Recent changes: autosave refactor merged Tuesday (PR #482).
Recipe 3

Slow-Cooked System Prompt

Best made slowly and revisited often. Skim off the contradictions as they rise to the surface.

Serves
Many agents, for a long time
Time
Several weeks, with regular tasting

Ingredients

  • 1 purpose
  • 1 audience
  • A handful of constraints, each with its reason
  • 1 or 2 worked examples
  • A generous pinch of permission to be unsure

Method

  1. Begin with who the agent serves and why.
  2. Add constraints one at a time, each with its reason attached.
  3. Taste weekly. Remove any instruction nobody can explain.
Serving suggestion
You help customers of a small bookshop find and order books.
Customers are often buying gifts and may not know titles exactly.

Keep replies under 100 words, because they appear in a chat widget.
Never promise delivery dates; stock arrives unpredictably, so offer
to check instead.

If you're not sure whether we stock something, say so and offer to
ask a human colleague.
Recipe 4

Error Message Glaze

Brush onto any tool failure. Turns a bitter "Error" into something an agent can actually digest.

Serves
Every agent that calls your tool
Time
2 minutes per error

Ingredients

  • What happened
  • Why, if known
  • What to try next

Method

  1. Say what failed, with the specific input.
  2. Add the cause if you know it.
  3. Finish with a next step. Serve immediately.
Serving suggestion
File not found: config/prod.yaml
The config directory contains: development.yaml, production.yaml
Did you mean config/production.yaml?
Recipe 5

Stopping-Condition Souffle

Delicate, and it only rises if you set the limit before the oven is on. Do not attempt mid-loop.

Serves
1 long-running agent
Time
1 minute, before you start

Ingredients

  • 1 step or time budget
  • 1 retry limit
  • 1 instruction for what to do at the limit

Method

  1. Decide the budget before the run begins.
  2. State it in the prompt and enforce it in the harness. Both, not either.
  3. Say what a graceful stop looks like.
Serving suggestion
You have up to 40 tool calls for this task.
Retry any single failing command at most 3 times.
If you reach either limit, stop, summarize what you tried and what
you learned, and list what you'd need to continue.