How to keep agent prompts short with SpecDD
This guide shows you how to keep agent prompts short while preserving the context an agent needs to do correct work in a spec-driven development workflow.
Short prompts are not vague when the project has good specs. With SpecDD, the prompt names the work and the repository files provide intent, authority, boundaries, tasks, and checks.
Short answer
Write durable context into .sdd specs and bootstrap files. Then prompt the specific action:
Plan the Itinerary validation change.
or:
Implement the open validation task.
Keep each prompt to one instruction. Use separate prompts for explanation, planning, implementation, and review.
Why this works
Long prompts often try to carry too many jobs:
- explain project rules
- identify the target
- define ownership
- list forbidden files
- describe behavior
- request implementation
- request tests
- request reporting
SpecDD moves most of that durable context into files. The agent can use bootstrap files for workflow rules, root specs for broad intent, local specs for behavior and authority, and tasks for current work. The prompt can stay focused.
Steps
1. Move repeated prompt context into specs
If you keep typing a rule, write it where it belongs.
Local behavior:
Must:
Missing place names are rejected before an itinerary item is stored.Local boundary:
Must not:
Change destination search behavior.Writable scope:
Can modify:
./itinerary.js
./itinerary.test.jsShared project conventions belong in:
.specdd/bootstrap.project.md
Do not use a long prompt as a substitute for missing project context.
2. Name the work in human terms
Good prompt:
Implement the missing-place validation task.
This identifies the task. The spec supplies the rest.
Vague prompt:
Improve trip validation.
This invites scope expansion.
3. Use one instruction per prompt
Use separate prompts for separate phases:
Explain the Itinerary spec.
Plan the Itinerary validation change.
Implement the approved Itinerary validation plan.
Review this change against the Itinerary spec.
This gives you a review point between understanding, planning, implementation, and review.
4. Avoid bundling optional work
A short prompt should not hide extra scope.
Too broad:
Implement the validation task, refactor related code, update nearby tests, close all open tasks, and clean up the module.
Better:
Implement the missing-place validation task.
If related cleanup is truly needed, create a separate task or spec update.
5. Use tasks to make short prompts precise
Tasks let short prompts refer to a durable local work packet:
Tasks:
[ ] Add missing-place validation.Then:
Complete the open Itinerary validation task.
The task should be specific enough that the agent does not have to infer the work from a vague prompt.
6. Use Done when to keep short prompts safe
Completion criteria prevent short prompts from becoming under-specified:
Done when:
Missing-place behavior is covered by a check.
Existing itinerary ordering still passes.The prompt can be short because “done” is defined in the spec.
7. Review the result against the contract
Short prompts need strong review. Check:
- changed files are allowed
- required behavior is satisfied
- prohibited behavior is preserved
- checks match
Done when - tasks are marked done only after verification
- missing context is added to specs
Prompt examples
For orientation:
Orient in this SpecDD project.
For explanation:
Explain the Itinerary spec.
For planning:
Plan the Itinerary validation change.
For implementation:
Implement the approved Itinerary validation plan.
For review:
Review this change against the Itinerary spec.
Common mistakes
- Shortening prompts before the relevant spec exists.
- Packing several instructions into one “short” prompt.
- Using vague names that do not point to a spec, task, feature, or behavior.
- Repeating project conventions in prompts instead of
.specdd/bootstrap.project.md. - Letting short prompts bypass plan review for risky work.
How to verify the result
Your prompts are short in the right way when:
- durable context lives in specs and bootstrap files
- prompts identify one action
- the agent can explain the target spec before work
- plans stay within local authority
- final changes satisfy
Done when - reviewers do not need chat history to understand the work