- Home
- Skills
- Testing & Debugging
- prompt doctor
More screenshots
Works with the AI Agents you already use
prompt doctor
The problem When a prompt breaks, teams patch it the way committees write documents: more instructions, more politeness, another "IMPORTANT:" line.
$5.90
prompt doctor
Example session with this skill installed
My support-ticket triage agent keeps misbehaving. The system prompt worked great for two weeks, now it randomly answers in prose when the downstream parser needs JSON, sometimes invents account limits that aren't in the knowledge base, and yesterday it started replying in formal Spanish for no reason. My team keeps "fixing" it by adding more instructions but it's getting worse, not better. Can you fix it properly?
- Read your context and instructions
- Compiled the prompt doctor
Verdict: Two compounding diseases found — a contradiction war (4 rounds of "fixes" stacked conflicting instructions) and an example teaching the wrong format. Repaired in 3 one-variable rounds; baseline test set included. Also caught one non-prompt disease: the Spanish replies came from stale context, not the prompt.
Diagnosis (5 layers):
| Layer | Finding | Severity |
|---|---|---|
| 3 Conflict | "ALWAYS return strict JSON" (v1) + "if unsure, explain your reasoning in plain text first" (v3 patch) — model picks either, hence 3-out-of-5 behavior | block |
| 4 Few-shot | The one example shows prose-first-then-JSON — teaches the exact behavior the parser can't eat | block |
| 2 Context | Knowledge base excerpts pasted undated; agent invented "account limits" to fill gaps (no invention ban) | fix |
| 5 Format | Format contract was prose ("return JSON") with no skeleton, no start/end markers | fix |
| 1 Instruction | "Be professional" — undefined; drift correlates with user messages in other languages | review |
Not-the-prompt check: ✅ input clean ✅ scope specific ❌ context stale — Spanish drift traced to undated locale-mixed KB excerpts. Flagged separately: context pipeline needs date labels. A prompt rewrite can't fix that.
Repair Log (one variable per round):
| Round | Variable changed | Finding treated | Test result |
|---|---|---|---|
| 1 | Removed the v3 reasoning clause; added priority rule: format > accuracy > style | Conflict war | 5/5 parser-valid JSON |
| 2 | Replaced example with strict-JSON-only shape + one boundary case (angry ticket → still JSON) | Example teaching prose | 5/5 incl. boundary |
| 3 | Added JSON skeleton + "facts not in KB → unverified flag — never invent limits" | Format contract + invention ban | missing-info case flags instead of inventing |
Rebuilt prompt (ready to paste):
You are the triage agent for Acme support tickets.
PRIORITY ORDER: (1) format contract, (2) accuracy rules, (3) style.
INPUT: the ticket text and KNOWLEDGE BASE excerpts below. Facts not present in the excerpts do not exist — never state invented limits, dates, or policy numbers; mark them unverified instead.
OUTPUT CONTRACT (highest priority): return ONLY valid JSON. Start with { and end with }. No code fences, no commentary, no text before or after. Shape:
{"category": "billing|technical|account|other", "priority": "low|normal|high", "summary": "<=30 words", "unverified_claims": ["...any KB-absent facts you were tempted to use..."]}
TONE (lowest priority): professional; match the ticket's language; short fields, no slang.
Test contract: 4 frozen baseline inputs (typical / angry boundary / policy-absent / long multi-issue) — binary pass criteria, same set after any future edit. Tuned for Claude-family; re-baseline before porting to GPT.
Team ritual to unlearn: stop patching by adding instructions. Every future change: one variable, same test set, logged.
Connects securely to your tools. The creator never sees your data.
About this skill
The problem When a prompt breaks, teams patch it the way committees write documents: more instructions, more politeness, another "IMPORTANT:" line. The prompt bloats, the behavior gets random — because the real disease (a contradiction buried between two patches, an example teaching the wrong format, acceptance criteria living in someone's head, a knowledge base pasted without dates) survives every rewrite. And when a "fix" seems to work, nobody knows which change did it, because everything changed at once. Meanwhile, generic prompt-engineer skills hand you a shiny new prompt and the same fate. What it does
- Diagnoses by layer before touching anything — instruction / context / conflict / few-shot / format, each failure named and severity-ranked
- Rules out non-prompt diseases first — bad input, stale context, inherently vague tasks; a rewrite can't cure those, and saying so is part of the job
- Repairs one variable per round — every change logged against the finding it treats, with the test result attached
- Rebuilds with a strict output contract — priority-ordered constraints, JSON skeleton with start/end markers, invention bans for facts the model doesn't have
- Ships a frozen baseline test set — typical case, boundary case, missing-information case; binary pass criteria, same set after every future edit Why this beats prompting it yourself Ask a generic model to "improve this prompt" and you get prompt theater: a persona bolted on top, more words, more please-and-thank-you — while the contradiction on line 4 and the example that teaches prose-instead-of-JSON survive untouched. Worse, the rewrite changes ten things at once, so even when it "works" you've learned nothing about why. This skill names the failing layer, changes one variable per round, and hands you a test that proves the fix — or tells you honestly that the prompt was never the sick part. Use cases
- An agent system prompt that "sometimes ignores its instructions" — conflict and priority diagnosis, ranked constraint set
- A data-extraction prompt returning prose when the parser needs JSON — output contract + schema enforcement
- A prompt that hallucinated facts on real documents — invention bans, unverified-claim slots
- Un-bloating a mega-prompt that grew by committee into a maintainable, testable form Known limitations
- Diagnoses from the prompt text and example outputs you provide; can't observe live API parameters (temperature, tools, schema settings)
- Doesn't run the model itself — the test contract tells you exactly what to run; verification is on your side
- Model-family tuning matters (Claude XML conventions vs GPT Structured Outputs); name your target model or advice stays generic
- Optimizes for correctness and consistency first, not token cost What this skill refuses to do
- Rewrite from scratch when a diagnosis would do — the disease survives rewrites
- Change multiple variables in one round and call the result a fix
- Declare improvement without a test you can reproduce
- Blame the prompt when the input, stale context, or an impossible task is the real problem
- Guarantee behavior on a model family it wasn't tuned for
Cannot observe live API parameters or temperature. Does not execute the model; requires user-side verification using the provided test contract. Target model must be specified for tuned advice.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 4 days ago