Two Morning Plans from the Same Terminal
Two Morning Plans from the Same Terminal
My first mistake of the day usually happens before I write code: I ask an agent what to work on without giving it enough evidence. It produces a tidy plan, I start the first task, and twenty minutes later I discover yesterday’s unfinished migration. The plan was coherent. It just belonged to a different repository.
Here are two ways to spend the first fifteen minutes in your editor and terminal: a chat-first planning session and a repository-first planning session. The time boxes below are a suggested routine, not benchmark results.
Minutes 0–3: Describe the Day or Inspect the Repository
The chat-first approach starts with a familiar prompt: “Review this project and plan my day.” You add a ticket title, mention a deadline, and let the agent explore. This is comfortable when you are still waking up, but it leaves scope and priorities implicit.
The repository-first approach starts with a small evidence packet. Before asking for recommendations, inspect git status --short, git log -5 --oneline, and git diff --stat. Add the acceptance criteria from one current ticket and your actual available work time.
- Chat-first gives the agent freedom before establishing today’s constraints.
- Repository-first identifies unfinished changes before proposing any new work.
I also tell the agent that local changes may be intentional. It must not reset, stash, edit files, install dependencies, or run migrations during planning. Reading the repository is not permission to modify it.
Minutes 3–7: Request Suggestions or Demand Evidence
In the chat-first lane, the next response often contains plausible maintenance work: improve tests, simplify a module, investigate performance. These can be useful ideas, but none automatically deserves this morning’s attention.
In the repository-first lane, I use a tighter prompt: “Plan only; do not edit files. Read the supplied ticket and inspect relevant code. Separate observed facts from assumptions. Propose three tasks maximum, each with supporting file paths, a completion check, and a dependency. Ask before running commands beyond read-only inspection.”
- Chat-first optimizes for generating options when priorities remain unclear.
- Repository-first makes each recommendation traceable to a concrete requirement.
If a test failure matters, run the narrow relevant test yourself or explicitly approve its execution. Tests can write snapshots, create databases, or call external services. Record the command and result; do not let “probably failing” become a planning fact.
Minutes 7–11: Accept a Checklist or Choose a Deliverable
A chat-first plan might say, “Fix authentication, add coverage, update documentation.” That is three categories, not an executable morning.
A repository-first plan should be smaller: “Reproduce the expired-session redirect bug with one failing test, patch the redirect branch, then run the authentication test suite.” The deliverable is an observable behavior, and the first action is obvious.
I ask for one primary task, one fallback if blocked, and one explicit non-goal. I also reserve time for review. A plan that consumes every available minute assumes the agent will never misunderstand anything.
Minutes 11–15: Save the Plan and Pick the Agent
Copy the approved plan into your daily notes or a local planning file after checking that it contains no credentials. Keep the acceptance check beside the task. Start an implementation session with that bounded context rather than carrying forward every discarded suggestion.
Fluxo fits here as infrastructure: one key and one base URL, https://api.fluxorelay.com, across the Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok families and more. It works with Claude Code, Codex, Gemini CLI, and any OpenAI SDK. For example, you could compare Claude Sonnet 5.5 and GPT-6.1 Sol on the same planning packet without changing the acceptance criteria.
Billing follows real tokens, including cache reads and writes, with no monthly fee. Repeated context still deserves attention. Users also earn a referral share when people they invite pay. If this setup suits your terminal workflow, consider creating a Fluxo account.
The Morning Verdict: Repository-First Wins for Execution
Chat-first wins when you need to explore possible directions. Repository-first wins when you need to ship today: it exposes unfinished work, limits speculation, and produces a verifiable next step. My default is evidence first, three tasks maximum, and no edits until I approve the plan.