Directional Prompting

Directional Prompting

Two layers, both required.

Layer 1 — Outcome. Open with a block that names the destination. The goal, what "done" looks like, when to stop, the true invariants. This is the frame.

Layer 2 — Direction. Inside that frame, every sentence names the path forward with positive verbs. "Trace", "build", "use", "read", "return", "ask", "check". The correct behavior is described so clearly and completely that the wrong behavior has no room to exist.

Outcome without direction reads as wishful — the model knows where to go but not how to step. Direction without outcome wanders — the model walks crisp paths to nowhere. Both layers together: a model that knows the destination and walks toward it on every token.

Why both

Modern frontier models (Claude Opus 4.7, GPT-5.5) follow instructions literally. The Claude 4.7 guide: "Positive examples showing how Claude can communicate with the appropriate level of concision tend to be more effective than negative examples or instructions that tell the model what not to do." The GPT-5.5 guide: "GPT-5.5 is strongest when the prompt defines the target outcome, success criteria, constraints, and available context, then lets the model choose the path."

Both labs converge on the same shape. Name the destination. Name the path. Skip the prohibitions.

Layer 1 — The outcome block

Every non-trivial prompt opens with this:

Goal: <one sentence>

Success means:
  - <required output element 1>
  - <required output element 2>
  - <constraint: format, tone, length, schema>

Stop when: <explicit stopping condition>

Optional fourth field for agentic prompts:

Constraints: <only the true invariants — safety, required output fields, hard limits>

Rules for the block:

1. Goal is one sentence. If you need two sentences, the goal is two goals — split the prompt.
2. Success criteria are checkable. "Returns valid JSON matching schema X" beats "high quality output".
3. Stopping condition is explicit. "Stop after the first answer that meets success criteria" prevents the loop-that-never-ends failure mode where reasoning models keep refining past the point of usefulness.
4. Constraints carry real weight. Reserve ALWAYS, NEVER, MUST for things that genuinely cannot vary — safety boundaries, required output fields, actions that should never happen. Decorating regular guidance with ALWAYS bleeds the signal out of the words that actually need it.

Layer 2 — Directional execution

Inside the outcome frame, every sentence pulls forward.

The five rules

1. Lead with the verb of the correct action. "Trace", "build", "use", "read", "commit", "return", "write", "ask", "check". The first token of the sentence sets the trajectory.
2. Describe the destination, not the failure modes. "Return JSON matching this schema" beats "do not return prose". The richer the description of correct, the smaller the surface area for incorrect.
3. Replace prohibitions with positive replacements. Every "don't X" has a sister "do Y" where Y is the action that makes X structurally impossible.
4. Make the correct behavior the only behavior described. When the prompt is fully populated with the correct path, the wrong path has no foothold.
5. Cut hedges, warnings, and meta-commentary. "Be careful with...", "watch out for...", "make sure you don't..." — anxiety in text form. Delete and replace with the concrete positive action.

Bad → good

The audit pass

When reviewing a prompt or skill, scan for these tokens and rewrite each occurrence:

• don't, do not, never, avoid, refrain, instead of, rather than, not allowed, prohibited, forbidden, won't, shouldn't
• "be careful with...", "watch out for...", "make sure you don't..."
• Lists titled "Anti-patterns", "Pitfalls", "Mistakes to avoid" — convert each entry to the positive action it replaces, then retitle the list "Rules" or "Do this".

Each match is a prompt smell. Rewrite as the positive replacement. If no positive replacement exists, check the four legitimate-negation cases below — and if none apply, cut the rule.