Choosing an instruction profile for project work
Select a profile above the prompt before sending a request. Write your actual task normally; do not paste the preset into the prompt. Bridge saves the selection for that conversation and includes its instructions with each request. Change profiles when the kind of work changes. Use one at a time.
Saved profiles are available in the updated Coding QA build. An older installed build may not show the selector. Selecting a preset in LM Studio does not select it in Bridge.
Quick selection guide
| Stage or situation | Profile | Example request |
|---|---|---|
| Decide what to build or compare approaches | Reasoning & Architecture | Compare two ways to add project checkpoints. Recommend the smallest reliable design and acceptance checks. |
| Implement a defined change | Coding & Implementation | Inspect the relevant files, implement this change, preserve unrelated work, and run the relevant checks. |
| Something crashes, fails, or behaves incorrectly | Diagnostics & Repair | Reproduce this error, identify its cause, make the smallest repair, and verify the original failure is resolved. |
| Review permissions, credentials, inputs, or release security | Security Review | Review this feature’s trust boundaries and credential handling. Report confirmed findings separately from hypotheses. |
| Position a product or prepare launch material | Marketing & Positioning | Draft three listing headlines for this audience using only these verified features. Recommend one measurable launch experiment. |
| Independently assess completed work | Critical Review | Review these changes against the acceptance criteria. Identify material gaps and return a readiness verdict with evidence. |
| Ordinary questions with no standing specialization | None | Explain what this setting does. |
A useful request contains five things
- Goal: the concrete outcome you want.
- Context: the project, relevant feature/files, and current behavior.
- Constraints: what to preserve, avoid, or keep within budget.
- Acceptance: observable behavior and checks that establish success.
- Scope: inspect only, propose, or implement with the available approvals.
Example: “In Noah360’s disposable test copy, investigate the task reminder failure. Preserve notification permissions and unrelated changes. Reproduce the error first, propose the smallest fix, then run relevant tests. Do not change the production app or publish a build.” Choose Diagnostics & Repair before sending.
For ASO copy, choose Marketing & Positioning and supply the intended audience, actual features, screenshots, store limits, and current metrics. The preset cannot invent market research or provide live data without an available research tool.
A complete working sequence
Reasoning → Coding → Diagnostics if needed → Security where relevant → Critical Review → Marketing for release. Skip stages that do not apply. For an existing bug, begin with Diagnostics rather than redesigning the project. For a simple known change, begin with Coding.
Keep related implementation work in one conversation. For an independent Critical Review, start a new conversation and provide the requirements, diff or changed files, and test evidence. Another profile on the same model is a different instruction set, not an independent expert or guaranteed second opinion.
Settings and boundaries
| Profile | Temperature | Output allowance |
|---|---|---|
| Reasoning & Architecture | 0.3 | 4,096 tokens |
| Coding & Implementation | 0.2 | 4,096 tokens |
| Security Review | 0.2 | 4,096 tokens |
| Diagnostics & Repair | 0.2 | 4,096 tokens |
| Marketing & Positioning | 0.7 | 3,072 tokens |
| Critical Review | 0.3 | 4,096 tokens |
These are editable starting values, not guarantees of quality. Profile temperature and output allowance override request defaults. Thinking follows Bridge’s local task mode, not the profile name. Use Automatic or Reasoning mode for complex local work; explicit Fast mode keeps thinking off. Match the context window to the model runtime. Larger output allowances leave less space for input and tool results.
Edit profile changes the shared definition for every conversation using it. Selection is per conversation and survives restart. The current version does not import/synchronize LM Studio presets or automatically choose an instruction profile. Selecting None preserves normal request behavior.
Profiles do not install tools, enable permissions, remove approvals, create live browsing, or authorize cloud use. Local-only mode remains the cloud boundary. Approve commands and edits only after reviewing the concrete action. A project working directory is not an OS sandbox.
When to get help
A failed test is evidence, not automatically a reason to abandon Bridge. Allow one informed repair. If the same failure remains after two attempts, stop and prepare a handoff. Also stop for lost requirements, unexplained changes, missing execution evidence, or an uncertain recovery path.
Bring the task, expected/actual behavior, exact errors and command results, attempted fixes, changed files, and recovery location to your backup assistant. “Tests passed” needs an actual result; an exit code of zero alone does not establish that the tests cover the requested behavior.