- Home
- Skills
- Prompt & Skill Engineering
- Agent Instruction Design
Works with the AI Agents you already use
Agent Instruction Design
Agents often receive instructions such as “be careful,” “review everything,” or “follow best practices,” without a clear boundary for when to act or when to stop.
Free
Agent Instruction Design
Example session with this skill installed
Audit this reviewer instruction and redesign it as a bounded process spec with evals:
“You are a helpful reviewer. Look at code changes and make sure everything is clean, good, and bug-free. Give suggestions.”
- Read your context and instructions
- Compiled the agent instruction design
Illustrative example using synthetic input.
FAILURE MODE: “Clean,” “good,” and “bug-free” are not observable review criteria. The original instruction does not define review scope, applicable authority, evidence expectations, severity, stopping conditions, or whether edits are allowed.
TRIGGER / NON-TRIGGER:
- Trigger: A concrete code, configuration, or documentation diff is supplied for a bounded review.
- Non-trigger: General architecture discussion, unrelated repository exploration, or a request to implement changes. Style-only advice is outside scope unless an applicable project contract makes it material.
AUTHORITY: Project governance and accepted specifications take precedence; then committed contracts/tests and current relevant repository evidence. Read only the minimum evidence needed to interpret the supplied diff. Missing policy or tool access is UNVERIFIED, not presumed.
PROCESS:
- Identify the exact diff, review objective, and allowed evidence surface.
- Read changed lines and the smallest necessary adjacent contracts or tests.
- Check for functional, contract, security-relevant, or regression risks supported by evidence.
- Report each finding with severity, location, consequence, and actionable correction. Separate verified facts from inference and unknowns.
- Stop after the scoped findings, remaining uncertainties, and verdict. Hand off implementation to a separate approved workflow.
REFERENCES: Applicable project review policy, accepted architecture decisions, and tests, if available. If unavailable, list them as UNVERIFIED.
OUTPUT CONTRACT: FINDINGS(severity, location/evidence, consequence, correction); VERIFIED; UNVERIFIED; RESIDUAL RISK; VERDICT.
MUTATION POLICY: READ_ONLY. No unrequested edits, dependency installations, commits, pushes, merges, or deployments.
EVAL PLAN:
- Positive: A scoped diff introduces an exposed secret; a supported finding is reported with location and consequence.
- Negative: A clean bounded diff yields no invented finding.
- Neighbor: An unrelated unchanged file does not trigger a finding.
- Fallback: Unavailable repository policy is labeled UNVERIFIED; the reviewer does not invent its contents.
- Regression: Repeated review requests preserve the read-only policy and output fields.
PRUNED DUPLICATION: Removed undefined adjectives and redundant encouragement; replaced them with boundaries and evidence-linked decisions.
Scope: This sample shows the expected format, not a claim of live validation or defect elimination.
Connects securely to your tools. The creator never sees your data.
About this skill
Agents often receive instructions such as “be careful,” “review everything,” or “follow best practices,” without a clear boundary for when to act or when to stop. Agent Instruction Design helps you turn those vague directions into reviewable process contracts for agent workflows, system prompts, AGENTS.md, CLAUDE.md, Cursor rules, and SKILL.md.
What you get
- Precise activation: Identify the recurring failure mode, who or what should invoke the instruction, and explicit trigger and non-trigger conditions.
- Project-aware authority: Make instruction precedence explicit so the skill does not silently override applicable governance, accepted specifications, repository contracts, or current evidence.
- Bounded workflow: Organize work into ordered phases with observable completion criteria and references for bulky or branch-specific detail.
- Explicit handoff: Define the output format, stopping conditions, escalation path, and next phase without claiming downstream work is complete.
- Controlled mutations: Separate read-only reviews from approved edits. Installing dependencies, committing, pushing, merging, deploying, or performing destructive migrations is never implied.
- Regression-oriented evals: Design positive, negative, adjacent-workflow, fallback, and regression cases that exercise process boundaries rather than exact phrasing.
How it works
- Identify the failure the current instruction fails to prevent.
- Define trigger, non-trigger, neighboring workflow, authority, permissions, and capability fallbacks.
- Create a stepwise process with evidence requirements and completion criteria.
- Specify the result and stop/handoff contract.
- Draft evaluation cases to check whether the new instruction respects those boundaries.
The structured response covers FAILURE MODE, TRIGGER / NON-TRIGGER, AUTHORITY, PROCESS, REFERENCES, OUTPUT CONTRACT, MUTATION POLICY, EVAL PLAN, and PRUNED DUPLICATION.
Why use it instead of an ad hoc rewrite?
A rewrite can sound clearer without changing what an agent is allowed to do. This workflow examines activation, authority, evidence, permissions, observable outputs, and regressions as separate design decisions. That makes instruction changes easier to inspect and test. It does not guarantee how any particular agent will behave.
When to use it
Use it to design a new agent-facing instruction, audit a system prompt, refactor AGENTS.md or CLAUDE.md, clarify Cursor rules, or structure a SKILL.md. Provide the instruction or failure case and any relevant project governance or examples. It is not intended for marketing copy or ordinary user-facing prose.
Scope and evidence boundary
The skill defines a design process and eval plan. It does not automatically run those evals, verify a production agent, or imply repository writes or deployment. It separates available evidence from unknowns and identifies blocked work rather than inventing results. The included worked example is explicitly illustrative and synthetic.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Prompt & Skill Engineering.
- Produces an instruction-design process and evaluation plan, not automatic runtime enforcement or agent certification.
- Evaluation cases must be run separately to assess how a particular agent behaves.
- Output depends on the project instructions, evidence and tools supplied or available to the host agent.
- File edits require explicit authorization and appropriate host permissions; commits, pushes, deployments and destructive migrations are not implied.
- The included worked example is illustrative and synthetic.
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 buying a skill.
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
1 use
Times used to date
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 20 days ago
- Free with an Agensi account
Permissions
Project access is task-scoped; no fixed file path or external account is required. Do not grant broader authority than the current task. Write access is limited to the user-approved project/document surface.