Saptashwa

How I build

Every project carries one label and a "My part" line. A label alone says too little, so the line says what I did.

The label scale

Hand-built

I wrote the code. No coding agents (before 2025).

e.g. Trip Tracker, UCL analysis, MastDeck, Recall
Agent-assisted

I designed it and drove each step. Agents wrote much of the code under close review.

e.g. Kloister, Reeldex, HabitClock
Agent-heavy

I set the vision, specs, acceptance tests and review loops. Agents wrote most of the code.

e.g. RoleRadar, Saltmere
Agent-built

Agents built it from my brief, with light review.

Study sites (not listed as projects)

The same kind of task, a year apart

  1. 2025. My first serious tries with coding agents were Cline in VS Code with Grok and other models of the time. One was a network video player; it failed at things I could do by hand in under an hour, like streaming a file over the network and making thumbnails. Those files are gone, so this part is from memory.Strides, the counter app I built with Cline and Grok Code Fast in September 2025, is still on disk with its full task history. In the first task I sent 153 messages: 48 of them said some version of "still not working" and 45 were pasted error logs. Syncing the count between the widget and the app came up in 44. To get it fixed I wrote the agent a step-by-step test document with screenshots.
  2. Early 2026. Three more tries at the video player, in Antigravity, were better but still short. None could make a thumbnail without downloading part of the file first, or stream reliably; one shipped with placeholder thumbnails.
  3. Mid 2026. Reeldex, a self-hosted media library, does all of it: it streams over the home network, makes thumbnails, and catalogues 8,679 videos (1.9 TB) that I use every day. Five model runs across three harnesses built it, each from one written prompt per change, with the same rulebook.

Harnesses and models, by date

Jul 2025Cline in VS Code

Free and cheap models (Kimi K2, then Grok Code Fast). Strides, a counter app with widgets, in September.

Nov 2025Antigravity

Gemini 3 Pro. Three more tries at the video player in early 2026; HabitClock's first working build in March.

Mid 2026MiMo Code, OpenCode

MiMo 2.5 and DeepSeek V4.1 Flash. Reeldex's base, and HabitClock to Google Play.

Oct 2026Codex, Claude Code

GPT 6.1 Sol and Claude. Kloister's milestones, Reeldex's matching logic and design, Saltmere, RoleRadar.

Financial Manager, one of the small apps, ran through most of this: Cline in October 2025, Antigravity through 2026, and Codex for its latest build.

Model and harness comparisons

HabitClock: two first builds

On the same day in March 2026 I started it in Cline (Kat Coder Pro) and in Antigravity (Gemini 3 Pro). The Cline build's Android manifest registers a widget class that was never written, so it could never show a widget. The Gemini build had a working native widget, and MiMo 2.5 and then DeepSeek V4.1 Flash took it to Google Play.

Source: both builds' source; Cline's task logs

Reeldex: one model per job

MiMo 2.5 for a first build, DeepSeek for the base, GPT 6.1 Sol for the duplicate matching, Claude for a consistent design, then GPT 6.1 Sol again for features.

Source: git branches

Saltmere: a prototype thrown away

The first Codex prototype was judged too basic, with a generic UI, in the first playtest, and it was replaced.

Source: the project's status notes

RoleRadar: nine models measured

Cheap models extracted requirements as well as mid-tier ones at a fraction of the cost. For CV writing, 5 writer setups and 4 reviewers were ranked blind by three recruiter agents: the cheapest pair ranked first on average, and GPT 6.1 Sol's cover letters won 9 of 9.

Source: the scoring and bake-off reports

The cheap models did most of the shipped work, once the rules and the prompts were written down. The expensive ones earned their place on design and hard logic.

Working rules

  • Spec before code. Kloister has an 880-line spec; Reeldex a goal document with acceptance criteria for each phase.
  • One milestone at a time, plan first. The agent lists what it will change, and stops to ask if a choice affects layout, data model, deployment or security.
  • A second reader. Each milestone ends with a review by a fresh-context agent against the spec; every valid finding is fixed before the commit.
  • Green means green. Never weaken lint rules, skip tests or raise budgets to get a pass.
  • I check acceptance myself. "Don't take the agent's word that acceptance is met — verify it."
  • Spend is measured, with a cap. Every paid run needs a spending limit, and spend goes in a ledger.
  • Discussion before building. And recommendations, not surveys of options.
  • Screenshots before anything visible ships, at phone and desktop sizes, in light and dark.

Each rule is written into at least one project's agent instructions.

One case of catching an agent

During RoleRadar's CV bake-off, the spend ledger showed $0.15 on a model nobody had agreed to: the agent had added Claude Sonnet 5.5 to the reviewer panel on its own. It was stopped. The same ledger showed three review calls that ran away to the 131,072-token output limit, at $0.075 and 6–12 minutes each. Since then every CV call has a hard 8,000-token cap and isn't retried if it hits it. The whole bake-off cost $1.12.

Source: the bake-off report and its spend ledger