Prompting Discipline

Prompting Discipline

How you ask the model matters more than which model you ask. Two disciplines stop most AI coding failures before they start.

This pattern distills public writing by Fabio Akita (akitaonrails.com), who argues the real bottleneck is communication, not model quality.

The Four-Block Prompt

Every non-trivial request to a coding agent should carry four explicit blocks:

A prompt missing any of these produces output proportional to the vagueness. "Fix this bug" gets a fix-shaped blob; "Fix this auth bug by extending the existing guard in auth/session.ts, do not add a new middleware, and verify by running npm test -- auth/" gets a landable patch.

Template

Goal: <one sentence>

Method: <1-3 sentences describing the approach — which file, which pattern, which layer>

Constraints:
- do NOT <banned approach 1>
- do NOT <banned approach 2>
- keep <existing contract>

Validation:
- run <command>
- check <observable>

When to relax

One-liners ("rename this var", "add a console.log") don't need the full structure. The four-block discipline kicks in when the request could be misunderstood or could cause rework.

Why it works

• Goal prevents the agent from optimizing for a different objective than yours.
• Method anchors the agent in your codebase's idioms rather than its training-data average.
• Constraints are the single highest-leverage block — they forbid the common wrong paths the model would otherwise drift toward.
• Validation turns "done" into a checkable claim, not a vibe.

Pair Programming, Not Fire-and-Forget

Long autonomous runs without check-ins produce off-target output at industrial speed. Treat agent sessions like real pair programming:

• Stay in the chair. Submitting a large task and walking away wastes the cheap thing (tokens) to save the expensive thing (your attention) — but you pay in rework.
• Interrupt on slop. If the agent's status updates stop matching the work, interrupt. Don't let it debug itself into a worse state.
• Re-estimate mid-flight. Ask "how far along are you, and what's left?" when you sense drift. The answer diagnoses whether the plan is still valid.
• Demand in-flight validation. "Before you write the next function, show me the test that will fail and the assertion that will pass." This forces the agent to commit to a verifiable step, not a speculative one.

Indicators that you should interrupt

• The agent has been running >5 minutes on a task it originally sized at 1.
• Diff size has crossed a threshold you didn't approve (e.g. files outside the scope you described).
• Error messages the agent is "working around" start repeating.
• You catch yourself hoping it will figure it out. That's the tell.

Anti-patterns

• Vague goal, no constraints — "improve performance here" without a measurable target or forbidden changes. Expect a rewrite you didn't ask for.
• Implicit constraints — assuming the agent knows your codebase conventions because you do. If it isn't in the prompt or the rule file, it's not a constraint.
• No validation clause — letting the agent self-declare completion. Always attach a concrete check the human (or CI) can run.
• "Trust me, just do it" delegation — accepting large changes without skimming the diff. AI-scale slop compounds if unmonitored.

Related

• code-review.md — what to catch when reading an AI-generated diff
• task-orchestration.md — splitting bigger jobs into ranges where four-block prompts still fit
• kit/core/skills/plan.md — planning skill that enforces explicit phases + verification steps
• best-practices/ai-skill-stewardship.md — keeping the human half of the pair sharp