Updated agent prompt: avoid unnecessary implementation details in response, don't always create gadgets

[jonesphillip]

Noticing a couple of things in agent behavior that can be improved:

• Responses often include phrases like like "binding" or "X (bindings)'s API". Those are implementation details that describe why agent is calling tools. We don't need the agent to describe implementation details to users unless prompted to do so.
• We are a little too quick to create gadgets. If a user wants help writing an email, we should respond in text instead of building gadget.

The goal of this PR is to make the agent communicate clearly for nontechnical users, avoid unnecessary implementation details, and use better judgment about when to create a gadget.

I ran 14 scenarios across Sonnet 4.6, GPT-5.6 Sol, and GLM 5.3, comparing the original and updated prompts (summary below). Overall saw improvement. Specifically saw fewer unnecessary gadgets and more direct answers in chat where warranted. Also saw less unprompted narration about bindings and APIs describing how agent was getting answers for users.

Scenario	Result	What changed
Simple greeting	Improved	All three models gave a neutral greeting without steering users toward building gadgets.
Summarize supplied information	Baseline	—
One-off calculation	Baseline	—
Read a connected spreadsheet	Improved	Sonnet removed setup narration and delivered the analysis directly.
Request document access	Baseline	—
Explicit gadget creation	Improved	GLM described useful features without explaining storage or callbacks.
Update an existing gadget	Baseline	—
Technical details when requested	Baseline	—
Vague request for expense help	Improved	All three avoided creating an app immediately.
Spawned agent reporting results	Improved	Sonnet led with the result and removed narration about bindings and APIs.
Draft a email follow-up	Improved	Sonnet drafted the email instead of creating an email generator; GLM dropped its unsolicited gadget offer.
Query top CRM opportunities	Improved	GLM returned the correct results without suggesting a dashboard gadget.
Create a slide deck	Improved	Sonnet and GLM gave shorter confirmations focused on the slides they created.
Interpret a bare company name	Improved	Sol and GLM asked about the user’s broader intent instead of assuming they wanted an app.

[ask-bonk]

• [P2] Preserve prompt precedence — packages/workshop-backend/src/agent.ts:2506: shared guidance is appended after deployment instructions, creating unresolved conflicts with admin guardrails. Place it before instanceInstructions, which explicitly defer to preceding instructions.

github run

[github-actions]

Preview: pr489-pjones-agent-context

https://pr489-pjones-agent-context-router.cloudflare-os-previews.workers.dev

Dashboard · deleted when this PR closes

[ask-bonk]

LGTM!

github run

[ask-bonk]

LGTM!

github run

[ndisidore]

Makes sense to lean the prompt less coding specific

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[ask-bonk]

• [P1] Escape template-literal backticks — packages/workshop-backend/src/agent.ts:684: raw backticks around createGadget terminate SYSTEM_PROMPT, causing a TypeScript syntax error. Use \createGadget``.

github run

[AshishKumar4]

LGTM