The task states the goal in one verb phrase. It names the deliverable and its scope.
A task without scope invites a generic answer. One sentence of scope often fixes that.
Task line | Verdict | Why |
|---|---|---|
| Weak | No deliverable, scope, or audience. |
| Strong | Names the block, the question, and the length. |
| Strong | Names the artifact and the bug class. |
Context holds facts that change a correct answer. It includes interfaces, parameters, and failure symptoms.
Each fact should earn its place. A fact that cannot change the answer is clutter.
Constraints state limits: length, tools, forbidden actions, and what to assume. They shrink the answer space before generation.
Effective limits vary by model, so results vary by model: measure it on the task at hand.
Output format names the shape: headings, keys, or a table. Examples show one input and its desired output.
One short example often beats a paragraph of format prose. Two examples cover the normal case and one edge.
example section sizes in characters (stated example counts)
task 100, context 400, constraints 150, format 100, examples 250
total = 100 + 400 + 150 + 100 + 250 = 1000 characters
shares: task 10%, context 40%, constraints 15%, format 10%, examples 25%
shares sum: 10 + 40 + 15 + 10 + 25 = 100%
rough token estimate at 4 characters per token: 1000 / 4 = 250 tokens
the 4-characters-per-token rule is approximate and tokenizer-dependentScope and shape break first. Length, forbidden actions, and output keys stay open, so answers vary run to run.
40%. The whole holds about 250 tokens at 4 characters per token, stated as an approximate rule.