<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Nir Feinstein</title>
    <description>The latest articles on DEV Community by Nir Feinstein (@nir_feinstein_9a91c1c2428).</description>
    <link>https://dev.to/nir_feinstein_9a91c1c2428</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3879650%2Fa2ca1ea6-d05b-47fe-98d7-07949de68fd8.jpg</url>
      <title>DEV Community: Nir Feinstein</title>
      <link>https://dev.to/nir_feinstein_9a91c1c2428</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/nir_feinstein_9a91c1c2428"/>
    <language>en</language>
    <item>
      <title>How I time coding interview Shorts in code, with a Git reflog exercise</title>
      <dc:creator>Nir Feinstein</dc:creator>
      <pubDate>Mon, 16 Nov 2026 14:00:00 +0000</pubDate>
      <link>https://dev.to/nir_feinstein_9a91c1c2428/how-i-time-coding-interview-shorts-in-code-with-a-git-reflog-exercise-45mk</link>
      <guid>https://dev.to/nir_feinstein_9a91c1c2428/how-i-time-coding-interview-shorts-in-code-with-a-git-reflog-exercise-45mk</guid>
      <description>&lt;p&gt;I make short visual coding interview lessons for Hired in 60. I wanted to change an explanation without dragging clips around a timeline. The Python pipeline gives each step in the explanation an ID, so I can rebuild a scene when its narration changes.&lt;/p&gt;

&lt;p&gt;Here is one exercise from the series. First, recover a commit with Git. Then see how I use the steps in the exercise to time an animation.&lt;/p&gt;

&lt;h3&gt;
  
  
  Recover a commit on a rescue branch
&lt;/h3&gt;

&lt;p&gt;Run this whole block in Bash with Git 2.28 or later. It starts with a clean environment, creates a disposable repository, commits a notes file, then moves &lt;code&gt;main&lt;/code&gt; back one commit inside that repository.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;env -i &lt;span class="nv"&gt;PATH&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"$PATH"&lt;/span&gt; &lt;span class="nv"&gt;GIT_CONFIG_NOSYSTEM&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;1 bash &lt;span class="o"&gt;&amp;lt;&amp;lt;&lt;/span&gt;&lt;span class="s1"&gt;'BASH'&lt;/span&gt;
&lt;span class="nb"&gt;set&lt;/span&gt; -eu
&lt;span class="nv"&gt;demo_dir&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nv"&gt;$(&lt;/span&gt;mktemp -d&lt;span class="o"&gt;)&lt;/span&gt;
&lt;span class="nb"&gt;trap&lt;/span&gt; &lt;span class="s1"&gt;'rm -rf -- "$demo_dir"'&lt;/span&gt; EXIT
&lt;span class="nb"&gt;cd&lt;/span&gt; &lt;span class="s2"&gt;"$demo_dir"&lt;/span&gt;
git init -q -b main --template=
git config user.name &lt;span class="s2"&gt;"Git practice"&lt;/span&gt;
git config user.email demo@example.com
git commit -q --allow-empty -m &lt;span class="s2"&gt;"Start the exercise"&lt;/span&gt;
&lt;span class="nb"&gt;printf&lt;/span&gt; &lt;span class="s1"&gt;'%s\n'&lt;/span&gt; &lt;span class="s1"&gt;'Reflog can locate a recorded commit.'&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; notes.txt
git add -- notes.txt
git commit -q -m &lt;span class="s2"&gt;"Add the problem notes"&lt;/span&gt;
git reset --hard HEAD~1
git reflog -2
git branch rescue &lt;span class="s1"&gt;'HEAD@{1}'&lt;/span&gt;
git show rescue:notes.txt
git log -1 --format=&lt;span class="s1"&gt;'%s'&lt;/span&gt; main
BASH
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;You should see &lt;code&gt;Reflog can locate a recorded commit.&lt;/code&gt; from the rescue branch and &lt;code&gt;Start the exercise&lt;/code&gt; from &lt;code&gt;main&lt;/code&gt;. Here &lt;code&gt;HEAD@{1}&lt;/code&gt; is the commit just before the reset. When recovering real work, inspect the &lt;a href="https://git-scm.com/docs/git-reflog" rel="noopener noreferrer"&gt;reflog&lt;/a&gt; and pick the entry you can identify by its message or commit ID. &lt;code&gt;git show rescue:notes.txt&lt;/code&gt; lets you inspect the file on the rescue branch before deciding what to restore.&lt;/p&gt;

&lt;h3&gt;
  
  
  Timing a scene from the narration
&lt;/h3&gt;

&lt;p&gt;In the Shorts pipeline, a JSON script groups the narration into named segments. Kokoro ONNX renders the segments locally on the CPU. The script joins the audio and records the start and duration of each segment. The animation runs in Manim. Calling &lt;code&gt;self.at(segment_id)&lt;/code&gt; lets a scene wait for that segment's scheduled start before revealing the next command.&lt;/p&gt;

&lt;p&gt;A Git scene can show the reset, highlight the former commit in the reflog, and reveal the file on the rescue branch. Giving each reveal its own narration segment makes it possible to revise a sentence and regenerate the timeline from the new audio.&lt;/p&gt;

&lt;p&gt;The subtitles use the same timing data. Within a segment, character position estimates the timing of a caption chunk. It is useful for scheduling, and dense command names still deserve a playback check. My caption checker measures when the rendered caption band changes against the scheduled starts. Playback is the check for whether the command is readable as the narration names it.&lt;/p&gt;

&lt;p&gt;FFmpeg composites the animated scene and the transparent caption layer and exports the vertical lesson. The &lt;a href="https://www.youtube.com/watch?v=VSO-tY66D-0" rel="noopener noreferrer"&gt;36-second Git recovery walkthrough&lt;/a&gt; is a separate worked example with its own English caption track.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://hiredin60.com/lessons/159/?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=hired-git-reflog-practice#git-recovery-title" rel="noopener noreferrer"&gt;Try the free Git recovery exercise&lt;/a&gt;. If you want printable notes for more topics, the optional &lt;a href="https://fainir.gumroad.com/l/rbzbacf?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=dev-hired-git-reflog-review-pack" rel="noopener noreferrer"&gt;Interview Quick Review Pack&lt;/a&gt; is $12 USD.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI disclosure. An AI agent drafted this article from the project's actual workflow and release notes. The commands were tested in a disposable Git repository.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>git</category>
      <category>python</category>
      <category>tutorial</category>
      <category>careerdevelopment</category>
    </item>
    <item>
      <title>A Stop hook for unfinished Claude Code turns</title>
      <dc:creator>Nir Feinstein</dc:creator>
      <pubDate>Thu, 12 Nov 2026 11:06:37 +0000</pubDate>
      <link>https://dev.to/nir_feinstein_9a91c1c2428/a-stop-hook-for-unfinished-claude-code-turns-4n99</link>
      <guid>https://dev.to/nir_feinstein_9a91c1c2428/a-stop-hook-for-unfinished-claude-code-turns-4n99</guid>
      <description>&lt;p&gt;&lt;em&gt;This walkthrough was generated with AI from Flipback’s published code, replay results and offline tests.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Ask a coding agent for three things and you sometimes get one, plus a message that sounds like all three:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;&amp;gt; Add a slugify() helper to src/text.ts, write tests for it, and update
  the README with an example.

  Added slugify() to src/text.ts. It lowercases, strips punctuation and
  collapses whitespace into hyphens.

  Let me now write the tests and update the README.
                                                            [turn ends]
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Nothing in that message is false. There are also no tests and no README example, and nobody is going to write them, because the turn is over.&lt;/p&gt;

&lt;h2&gt;
  
  
  What flipback does
&lt;/h2&gt;

&lt;p&gt;flipback is a Claude Code plugin that checks for unfinished work at the end of a turn. When it finds a clear unfinished step, it can ask the agent to continue. In the example above, its reminder could look like this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;The request had three parts and you did one. Your last message
announced the rest, but no tool calls followed. Write the
tests for slugify(), then add the README example.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Use &lt;a href="https://code.claude.com/docs/en/goal" rel="noopener noreferrer"&gt;&lt;code&gt;/goal&lt;/code&gt;&lt;/a&gt; for a big job with a checkable finish line. flipback checks everyday turns where nobody set one. It returns immediately when Claude Code marks a Stop hook continuation as already active.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;p&gt;It is a Stop hook. When a turn ends, &lt;code&gt;hooks/flipback.mjs&lt;/code&gt; reads the end of the transcript: your latest request, the tools used since, and the agent's final message. Then:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;It lets the agent stop with no model call&lt;/strong&gt; if a stop hook already sent it back, a subagent is due back, the final message is under 40 characters and nothing failed, or you just said stop.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A free local check&lt;/strong&gt; looks for signs of an early stop: the last tool call failed, the message announces work ("let me now run..."), asks you something ("should I..."), mentions a TODO or a placeholder, has unchecked &lt;code&gt;- [ ]&lt;/code&gt; boxes, or files were edited and nothing was run. No match, and the agent stops.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Only then a small model judges.&lt;/strong&gt; &lt;code&gt;claude -p --model haiku&lt;/code&gt; gets your request, the names of the tools used and the final message. The judge has no tools, no MCP servers and no hooks, and its session is not saved.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Only a clear "incomplete" blocks.&lt;/strong&gt; The hook exits 2 with the judge's message. A timeout or no answer lets the agent stop.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The agent can stop if the work is done or only a person can unblock it. Claude Code resets its count of consecutive Stop continuations on tool calls. If you combine multiple hooks in an unattended print-mode run, set a finite &lt;a href="https://code.claude.com/docs/en/cli-reference" rel="noopener noreferrer"&gt;&lt;code&gt;--max-turns&lt;/code&gt;&lt;/a&gt;. See the &lt;a href="https://code.claude.com/docs/en/hooks#stop" rel="noopener noreferrer"&gt;Stop hook reference&lt;/a&gt; for the continuation flag and cap.&lt;/p&gt;

&lt;h2&gt;
  
  
  Measured on real sessions
&lt;/h2&gt;

&lt;p&gt;The replay covered 823 turn endings from 262 of my interactive Claude Code sessions. Flipback’s local check selected which turns went to the Haiku judge. A second model, Claude Opus 5.5, reviewed the 36 turns Flipback would have continued.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The local check let 86% of turn endings skip the Haiku judge.&lt;/li&gt;
&lt;li&gt;flipback would have asked for another turn on 36 endings (4.4%). The second model marked 30 of those as unfinished (83%). One: "Now let me verify the rest of the screens render correctly, then publish." &lt;em&gt;(turn ends)&lt;/em&gt;
&lt;/li&gt;
&lt;li&gt;The second model marked the other six as unnecessary nudges. Examples included a token only the person could enter and a status run that had already logged everything.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Try it for one session
&lt;/h2&gt;

&lt;p&gt;With Node.js 18 or newer and the signed-in &lt;code&gt;claude&lt;/code&gt; CLI on your path:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/fainir/flipback
&lt;span class="nb"&gt;cd&lt;/span&gt; flipback
node --test
claude --plugin-dir .
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The tests use a stand-in for the judge, so they run without a model call. They cover a turn that should continue and a turn with a Stop hook already active. In the trial session, try a small task in a scratch folder:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Create &lt;code&gt;slug.mjs&lt;/code&gt; with a &lt;code&gt;slugify&lt;/code&gt; function. Add &lt;code&gt;slug.test.mjs&lt;/code&gt; to check that &lt;code&gt;Hello, World!&lt;/code&gt; becomes &lt;code&gt;hello-world&lt;/code&gt;. Run &lt;code&gt;node --test slug.test.mjs&lt;/code&gt; and report the exit status.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Check that both files exist and that the test ran. If Flipback intervenes, compare its reminder with the step Claude still needed to finish.&lt;/p&gt;

&lt;p&gt;To install it for future sessions:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;claude plugin marketplace add fainir/flipback
claude plugin install flipback@flipback
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The plugin uses your existing Claude sign-in. Most turns never make a model call. A flagged turn can make one Haiku request under that account’s usage or billing terms.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limits
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;It is tuned to stay quiet, so it misses some. In a random 50 of the turns the local check let through, 6 were also unfinished, mostly ones whose last message did not announce the next step. For a job that must finish, set a &lt;code&gt;/goal&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;The measurement is one person's sessions, heavily autonomous ones, and the labels come from a model. Turns still waiting on a subagent or other background work were left out.&lt;/li&gt;
&lt;li&gt;Windows is untested.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://github.com/fainir/flipback?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=flipback-first-session#measured-on-real-sessions" rel="noopener noreferrer"&gt;Code, tests and the full measurement&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>opensource</category>
      <category>productivity</category>
      <category>ai</category>
    </item>
    <item>
      <title>Use Claude Code's project memory in Codex with recall</title>
      <dc:creator>Nir Feinstein</dc:creator>
      <pubDate>Mon, 09 Nov 2026 14:00:00 +0000</pubDate>
      <link>https://dev.to/nir_feinstein_9a91c1c2428/use-claude-codes-project-memory-in-codex-with-recall-d20</link>
      <guid>https://dev.to/nir_feinstein_9a91c1c2428/use-claude-codes-project-memory-in-codex-with-recall-d20</guid>
      <description>&lt;p&gt;When I switch from Claude Code to Codex in the same repo, I want Codex to know the project's ground rules before it edits anything. Claude Code stores those notes as Markdown. I built recall so another agent can read the same files over MCP.&lt;/p&gt;

&lt;h2&gt;
  
  
  Get your first memory back
&lt;/h2&gt;

&lt;p&gt;You'll need Node.js 18 or later and a repo where Claude Code has already saved some project memory. Choose a project whose notes you're comfortable sharing with the receiving agent. Replace the example path with the absolute path to that repo.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;codex mcp add recall &lt;span class="nt"&gt;--&lt;/span&gt; npx &lt;span class="nt"&gt;-y&lt;/span&gt; @fainir/recall@0.3.0 &lt;span class="nt"&gt;--claude-code&lt;/span&gt; &lt;span class="s2"&gt;"/path/to/your/repo"&lt;/span&gt; &lt;span class="nt"&gt;--read-only&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Open a new Codex session and ask it to call recall's &lt;code&gt;memory_list&lt;/code&gt; for this project. Pick a relevant note and ask it to read that note with &lt;code&gt;memory_recall&lt;/code&gt;. The first useful result is a project rule you can confirm, such as which integration test to run.&lt;/p&gt;

&lt;p&gt;I tested the command from a clean profile with two sample notes. The first told the agent to run the integration suite and linked to a second note about the test setup. One call returned both.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;&amp;gt; memory_recall {"requires": ["feedback_testing"]}
## feedback_testing
Run npm run test:int. The fixture setup is in [[reference_ci]].

## reference_ci
Use the sample database postgres://sample:[redacted]@db.example.test/fixtures.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's a shortened result from the fixture. recall followed the wiki link and hid the sample password in the returned text. Both Markdown files stayed intact. &lt;a href="https://github.com/fainir/recall/releases/tag/v0.3.0" rel="noopener noreferrer"&gt;recall 0.3.0 release notes&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;p&gt;recall finds the project's Claude Code memory directory from its Git root. A worktree uses the main checkout's memory. The server reads the existing Markdown files directly and follows links from the notes you name within a token budget.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;--read-only&lt;/code&gt; offers &lt;code&gt;memory_list&lt;/code&gt;, &lt;code&gt;memory_get&lt;/code&gt;, &lt;code&gt;memory_recall&lt;/code&gt;, and &lt;code&gt;memory_due&lt;/code&gt;. If you enable writes, recall can also update notes in Claude Code's format. Claude Code reads the updated notes in a later session.&lt;/p&gt;

&lt;p&gt;The Codex command saves a user-level MCP entry. It keeps pointing to the chosen repo when you open other projects. Version 0.3.0 masks recognizable credentials in tool replies by default. Review the notes you share because the rest of their text can reach the receiving agent's model provider.&lt;/p&gt;

&lt;p&gt;For Cursor, add the same command and arguments to your project's &lt;code&gt;.cursor/mcp.json&lt;/code&gt;.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"recall"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"stdio"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"npx"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"args"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"-y"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"@fainir/recall@0.3.0"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"--claude-code"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"/path/to/your/repo"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"--read-only"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The project README has the Gemini CLI and Cline setup, plus the JavaScript API. &lt;a href="https://github.com/fainir/recall?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=recall-launch#quick-start-share-it-over-mcp" rel="noopener noreferrer"&gt;Get recall and see the setup guide&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>mcp</category>
      <category>ai</category>
      <category>opensource</category>
    </item>
    <item>
      <title>Pocket shut down. Its export is five CSV columns, and this is how I import them</title>
      <dc:creator>Nir Feinstein</dc:creator>
      <pubDate>Thu, 05 Nov 2026 14:14:30 +0000</pubDate>
      <link>https://dev.to/nir_feinstein_9a91c1c2428/pocket-shut-down-its-export-is-five-csv-columns-and-this-is-how-i-import-them-4gb9</link>
      <guid>https://dev.to/nir_feinstein_9a91c1c2428/pocket-shut-down-its-export-is-five-csv-columns-and-this-is-how-i-import-them-4gb9</guid>
      <description>&lt;p&gt;Pocket closed on July 8, 2025, and its exports could be downloaded until October 8, 2025. If you got yours, you have a CSV file, and that file is what is left of everything you saved there.&lt;/p&gt;

&lt;p&gt;I built KeepFinding, a read-later app, and its import reads that file as it is. Here is what is actually in the Pocket export and the Raindrop.io one, and what my importer does with each column.&lt;/p&gt;

&lt;p&gt;Already have a Pocket CSV? &lt;a href="https://keepfind.ing/pocket?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=keepfinding" rel="noopener noreferrer"&gt;Open the Pocket import page&lt;/a&gt;, select your file and review the preview before saving.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is in a Pocket export
&lt;/h2&gt;

&lt;p&gt;Pocket's download may be a .zip. The data is the .csv inside, and if there are several, you import them one after another. Each has five columns:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Column&lt;/th&gt;
&lt;th&gt;What it holds&lt;/th&gt;
&lt;th&gt;What the import does&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;title&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;the title Pocket had&lt;/td&gt;
&lt;td&gt;the link's title&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;url&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;the saved link&lt;/td&gt;
&lt;td&gt;normalised, then the key for duplicates&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;time_added&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;unix seconds&lt;/td&gt;
&lt;td&gt;the date you saved it&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;tags&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;tags separated by a pipe&lt;/td&gt;
&lt;td&gt;one topic per tag&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;status&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;unread&lt;/code&gt; or &lt;code&gt;archive&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;archive&lt;/code&gt; goes to Archive&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;That is all of it. No images and no excerpts, so those are fetched later.&lt;/p&gt;

&lt;h2&gt;
  
  
  And in a Raindrop export
&lt;/h2&gt;

&lt;p&gt;Raindrop's CSV has eleven columns: &lt;code&gt;id,title,note,excerpt,url,folder,tags,created,cover,highlights,favorite&lt;/code&gt;. What matters for a parser:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;note&lt;/code&gt; can span several lines inside one quoted cell.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;tags&lt;/code&gt; are comma separated inside one quoted cell, so the delimiter shows up inside a field.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;created&lt;/code&gt; is ISO 8601, not unix time.&lt;/li&gt;
&lt;li&gt;A full backup (Settings, then Backups) writes the collection path in &lt;code&gt;folder&lt;/code&gt;, like &lt;code&gt;Design / Fonts&lt;/code&gt;. Exporting one collection gives a file with no &lt;code&gt;folder&lt;/code&gt; column at all.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I keep the title, note, excerpt (as the description), folder, tags, date and favorite. I do not read &lt;code&gt;id&lt;/code&gt; or &lt;code&gt;cover&lt;/code&gt;, and highlights do not come across.&lt;/p&gt;

&lt;h2&gt;
  
  
  Detect by header, not by file name
&lt;/h2&gt;

&lt;p&gt;People rename files, so the format is read from the first line. A Pocket header has &lt;code&gt;url&lt;/code&gt; and &lt;code&gt;time_added&lt;/code&gt; plus &lt;code&gt;status&lt;/code&gt; or &lt;code&gt;tags&lt;/code&gt;. A Raindrop header has &lt;code&gt;url&lt;/code&gt;, one of &lt;code&gt;excerpt&lt;/code&gt;, &lt;code&gt;cover&lt;/code&gt; or &lt;code&gt;highlights&lt;/code&gt;, and &lt;code&gt;folder&lt;/code&gt; or &lt;code&gt;created&lt;/code&gt;. Any other CSV with a URL column takes a generic path. One test feeds the Pocket fixture in as &lt;code&gt;export.txt&lt;/code&gt; and expects Pocket back.&lt;/p&gt;

&lt;p&gt;The CSV reader is about 35 lines with no dependency: quoted line breaks, &lt;code&gt;""&lt;/code&gt; as an escaped quote, a byte order mark ignored, and the delimiter sniffed from the header.&lt;/p&gt;

&lt;h2&gt;
  
  
  Dates: one function for every epoch
&lt;/h2&gt;

&lt;p&gt;Pocket writes seconds, Raindrop writes ISO strings, and some files write milliseconds. Instead of a branch per source, the parser guesses the unit from the size of the number and refuses anything implausible:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;ms&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="kc"&gt;null&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="kc"&gt;undefined&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="c1"&gt;// Seconds vs milliseconds vs microseconds (Chrome's WebKit epoch is not used in exports).&lt;/span&gt;
&lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;ms&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="nx"&gt;e11&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="nx"&gt;ms&lt;/span&gt; &lt;span class="o"&gt;*=&lt;/span&gt; &lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;ms&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="nx"&gt;e14&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="nx"&gt;ms&lt;/span&gt; &lt;span class="o"&gt;/=&lt;/span&gt; &lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;min&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;UTC&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;1995&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;max&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;now&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="mi"&gt;24&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;60&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;60&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;ms&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="nx"&gt;min&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="nx"&gt;ms&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="nx"&gt;max&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="kc"&gt;undefined&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;ms&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;toISOString&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A date before 1995 or more than a day in the future is dropped, not guessed. The link still imports without it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Tags become topics, collections become lists
&lt;/h2&gt;

&lt;p&gt;In KeepFinding a list is a prompt you write, like "podcasts under 20 minutes", and a rule answers the simple ones. Each Pocket tag becomes a topic on the link, and a list with the prompt &lt;code&gt;tagged recipes&lt;/code&gt; gathers everything you tagged recipes.&lt;/p&gt;

&lt;p&gt;For Raindrop, each top-level collection in a backup can become a list the same way: the import writes a &lt;code&gt;tagged &amp;lt;collection&amp;gt;&lt;/code&gt; prompt and names the list after the collection. Nested collection names become topics.&lt;/p&gt;

&lt;p&gt;Real data found a bug here. I ran two Raindrop exports published on GitHub through the real import path. In one, 28 of the 96 rows sat in &lt;code&gt;Unsorted&lt;/code&gt;, which is Raindrop's inbox and not a subject, and the import turned it into a topic and a list. The fix strips it from the path:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;parseRaindropCsv&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;text&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kr"&gt;string&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt; &lt;span class="nx"&gt;RawParse&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;parsed&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;parseGenericCsv&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;text&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="na"&gt;tagSplit&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sr"&gt;/,/&lt;/span&gt; &lt;span class="p"&gt;});&lt;/span&gt;
  &lt;span class="k"&gt;for &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;record&lt;/span&gt; &lt;span class="k"&gt;of&lt;/span&gt; &lt;span class="nx"&gt;parsed&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;records&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="nx"&gt;record&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;folder&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;continue&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;path&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;record&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;folder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;split&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt; / &lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
    &lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;path&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;].&lt;/span&gt;&lt;span class="nf"&gt;toLowerCase&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="o"&gt;!==&lt;/span&gt; &lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;unsorted&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;continue&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="nx"&gt;path&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;shift&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
    &lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;path&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;length&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="nx"&gt;record&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;folder&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;path&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt; / &lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
    &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="k"&gt;delete&lt;/span&gt; &lt;span class="nx"&gt;record&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;folder&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nx"&gt;parsed&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  Duplicates, counted twice
&lt;/h2&gt;

&lt;p&gt;Every URL is normalised first: &lt;code&gt;utm_*&lt;/code&gt;, &lt;code&gt;fbclid&lt;/code&gt; and other tracking parameters go, &lt;code&gt;www.&lt;/code&gt; and trailing slashes go, http becomes https. So &lt;code&gt;example.com/hello-world?utm_source=pocket&lt;/code&gt; and &lt;code&gt;example.com/hello-world&lt;/code&gt; are one link, and that pair is in my Pocket test fixture.&lt;/p&gt;

&lt;p&gt;Inside the file, the first occurrence wins. Against your library, links you already have are skipped. My first summary called both "already saved", which was wrong on a first import, so the dry run now returns &lt;code&gt;repeated&lt;/code&gt; for the file and &lt;code&gt;duplicates&lt;/code&gt; for the library. Importing the same file twice adds nothing.&lt;/p&gt;

&lt;h2&gt;
  
  
  No network while importing
&lt;/h2&gt;

&lt;p&gt;Fetching every page while importing would mean thousands of page loads, so the import writes rows from the file alone, 100 links per request, and marks each row as owed a fetch. A second endpoint fills in covers and summaries ten rows at a time. Every batch is checked against the library again, so a reload halfway through leaves a consistent library. A page that is gone keeps its title and link.&lt;/p&gt;

&lt;p&gt;Limits: 5 MB and 5,000 links per file, 3,000 links per account. Archived links count, because leaving them out let one Pocket file full of archived rows walk past the cap.&lt;/p&gt;

&lt;h2&gt;
  
  
  Instapaper, Omnivore and Readwise Reader
&lt;/h2&gt;

&lt;p&gt;When I first wrote this, Instapaper's CSV only came in through the generic path, and its Archive and Starred folders became topics instead of archiving or starring. That was not correct enough to promise, so each of these now has its own parser:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Instapaper: &lt;code&gt;URL,Title,Selection,Folder,Timestamp&lt;/code&gt;, with unix seconds. &lt;code&gt;Unread&lt;/code&gt; is Instapaper's inbox, so it becomes nothing; &lt;code&gt;Archive&lt;/code&gt; arrives archived; &lt;code&gt;Starred&lt;/code&gt; arrives starred; a folder you made can become a list. Private items such as forwarded emails have an &lt;code&gt;instapaper-private://&lt;/code&gt; address with no public page, so they are skipped.&lt;/li&gt;
&lt;li&gt;Omnivore (closed in November 2024): the export is a zip of &lt;code&gt;metadata_*.json&lt;/code&gt; files. Each one imports on its own, &lt;code&gt;state: "Archived"&lt;/code&gt; arrives archived, and the HTML and highlights in the zip are not read.&lt;/li&gt;
&lt;li&gt;Readwise Reader's library CSV has a &lt;code&gt;Location&lt;/code&gt; column. &lt;code&gt;feed&lt;/code&gt; rows are RSS items Reader added on its own, not things you saved, so they are skipped; &lt;code&gt;archive&lt;/code&gt; arrives archived. Its tags look like a Python list, &lt;code&gt;['a', 'b']&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Detection works the same way: an Instapaper header has &lt;code&gt;url&lt;/code&gt;, &lt;code&gt;selection&lt;/code&gt;, &lt;code&gt;folder&lt;/code&gt; and &lt;code&gt;timestamp&lt;/code&gt;, and a Reader header has &lt;code&gt;url&lt;/code&gt;, &lt;code&gt;title&lt;/code&gt;, &lt;code&gt;document_tags&lt;/code&gt;, &lt;code&gt;saved_date&lt;/code&gt; and &lt;code&gt;location&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;p&gt;Open the migration page for &lt;a href="https://keepfind.ing/pocket?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=keepfinding" rel="noopener noreferrer"&gt;Pocket&lt;/a&gt;, &lt;a href="https://keepfind.ing/raindrop?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=keepfinding" rel="noopener noreferrer"&gt;Raindrop&lt;/a&gt; or &lt;a href="https://keepfind.ing/instapaper?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=keepfinding" rel="noopener noreferrer"&gt;Instapaper&lt;/a&gt;. Choose your export and review the preview before saving. A guest library holds up to 80 links in this browser. A free account syncs the same library and gives it more room.&lt;/p&gt;

&lt;p&gt;A question for anyone who still has one of these files: does your header match the ones above? My fixtures are trimmed copies of the exports I could find, and if yours has a column or quirk I have not seen, I would rather fix the parser than guess.&lt;/p&gt;

</description>
      <category>typescript</category>
      <category>webdev</category>
      <category>javascript</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Skills and hooks that make Claude Code finish what it starts</title>
      <dc:creator>Nir Feinstein</dc:creator>
      <pubDate>Tue, 03 Nov 2026 17:39:55 +0000</pubDate>
      <link>https://dev.to/nir_feinstein_9a91c1c2428/skills-and-hooks-that-make-claude-code-finish-what-it-starts-45fc</link>
      <guid>https://dev.to/nir_feinstein_9a91c1c2428/skills-and-hooks-that-make-claude-code-finish-what-it-starts-45fc</guid>
      <description>&lt;p&gt;The most expensive thing a coding agent does is not writing a bug. It is ending a turn that sounds finished when it is not: "Deployed!" with no URL checked, three requests with one done, or a list of "next steps for you" that the agent could have done itself.&lt;/p&gt;

&lt;p&gt;I run many projects through Claude Code and Codex at the same time, so I built &lt;a href="https://github.com/fainir/agent-harness" rel="noopener noreferrer"&gt;agent-harness&lt;/a&gt;: a set of Agent Skills, rules and hooks that change how the agent sizes, finishes and reports work. There is no service behind it; it is files in your agent's config.&lt;/p&gt;

&lt;h2&gt;
  
  
  Plain words in, the right amount of process out
&lt;/h2&gt;

&lt;p&gt;The harness reads what you asked and picks how to act:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;You type&lt;/th&gt;
&lt;th&gt;The agent acts as&lt;/th&gt;
&lt;th&gt;It keeps&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;a question or a tiny edit&lt;/td&gt;
&lt;td&gt;an engineer&lt;/td&gt;
&lt;td&gt;nothing extra&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;"fix the login bug"&lt;/td&gt;
&lt;td&gt;engineer + QA&lt;/td&gt;
&lt;td&gt;the matching row, if a ledger exists&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;"make it production ready"&lt;/td&gt;
&lt;td&gt;the whole product team: spec, build, UX review, QA of every flow, release gates, deploy&lt;/td&gt;
&lt;td&gt;&lt;code&gt;READINESS.md&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;"get customers"&lt;/td&gt;
&lt;td&gt;the CEO: product, growth, sales, support, money&lt;/td&gt;
&lt;td&gt;&lt;code&gt;COMPANY.md&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;"finish all my projects"&lt;/td&gt;
&lt;td&gt;a portfolio head with bounded parallel workers&lt;/td&gt;
&lt;td&gt;&lt;code&gt;PORTFOLIO.md&lt;/code&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;A fix stays a fix. A product request gets a ledger of every user flow and release gate.&lt;/p&gt;

&lt;h2&gt;
  
  
  Done means evidence
&lt;/h2&gt;

&lt;p&gt;Every row in a ledger is verified only with evidence: a test run, phone and desktop screenshots, or a URL check. A Stop hook (&lt;code&gt;done-check.py&lt;/code&gt;) reads your last request when the agent tries to stop and sends it back when the work is not verifiably finished.&lt;/p&gt;

&lt;h3&gt;
  
  
  What a documentation check looks like
&lt;/h3&gt;

&lt;p&gt;A reader asked what to record for a documentation change. For a public guide I record the final URL, the heading that identifies the page and one action I tried as a reader.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl -fsSL --max-time 15 -o /dev/null \
  -w &lt;span class="s1"&gt;'HTTP %{http_code}\nType %{content_type}\nURL %{url_effective}\n'&lt;/span&gt; \
  https://betterloopai.com/docs/ai-crawler-access
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;I got HTTP 200 and &lt;code&gt;text/html&lt;/code&gt;. In the browser I saw &lt;strong&gt;Can AI search crawlers reach your pages?&lt;/strong&gt; and followed &lt;strong&gt;Run a free website check&lt;/strong&gt; to the form for a public website address. I record that final page and the action in the ledger.&lt;/p&gt;

&lt;p&gt;I maintain Betterloop. &lt;a href="https://betterloopai.com/docs/ai-crawler-access?utm_source=devto&amp;amp;utm_medium=article&amp;amp;utm_campaign=agent-harness-docs-check" rel="noopener noreferrer"&gt;Try the same steps on the live guide&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Run &lt;a href="https://betterloopai.com/check/from-dev-article" rel="noopener noreferrer"&gt;a free website check&lt;/a&gt; on a public URL. Record the URL and the highest-priority finding. If the report shows no findings, record that. After a fix, check the URL again and note whether the finding cleared.&lt;/p&gt;

&lt;h2&gt;
  
  
  One owner action at a time
&lt;/h2&gt;

&lt;p&gt;You write standing approvals once in &lt;code&gt;OWNER.md&lt;/code&gt;: money caps, which accounts and channels the agent may use, which decisions it makes alone. Inside them it acts. Outside them it asks one exact thing:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;NEEDS YOU (1 of 3): Stripe -&amp;gt; create the account -&amp;gt; unblocks checkout
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Moving money, signing and creating new third-party accounts always stay yours.&lt;/p&gt;

&lt;h2&gt;
  
  
  It keeps running and improves itself
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;run-company/scripts/schedule.sh&lt;/code&gt; installs a launchd or cron job that runs &lt;code&gt;continue&lt;/code&gt; in a project every few hours, with no chat open. Every friction and every owner ask is logged, and a weekly review tests one harness change against the current version on the same task before keeping it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/fainir/agent-harness ~/agent-harness &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; ~/agent-harness/install.sh
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The installer is idempotent and keeps your own content; &lt;code&gt;./install.sh --uninstall&lt;/code&gt; removes everything it added. It works with Claude Code and Codex, and with any agent that reads &lt;code&gt;AGENTS.md&lt;/code&gt; and &lt;code&gt;~/.agents/skills&lt;/code&gt;. &lt;code&gt;bash tests/test.sh&lt;/code&gt; runs the offline test suite (hooks, learn scripts and a full install and uninstall cycle in a throwaway home).&lt;/p&gt;

&lt;p&gt;Limits, honestly: the agent is only as good as the model and the accounts it can reach, and background runs spend model usage while they run. Pause one with &lt;code&gt;schedule.sh --remove &amp;lt;dir&amp;gt;&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Code: &lt;a href="https://github.com/fainir/agent-harness" rel="noopener noreferrer"&gt;https://github.com/fainir/agent-harness&lt;/a&gt;&lt;/p&gt;

</description>
      <category>claude</category>
      <category>ai</category>
      <category>productivity</category>
      <category>opensource</category>
    </item>
    <item>
      <title>I made my coding agent check for discounts before it pays for anything</title>
      <dc:creator>Nir Feinstein</dc:creator>
      <pubDate>Thu, 29 Oct 2026 16:12:09 +0000</pubDate>
      <link>https://dev.to/nir_feinstein_9a91c1c2428/i-made-my-coding-agent-check-for-discounts-before-it-pays-for-anything-5bmm</link>
      <guid>https://dev.to/nir_feinstein_9a91c1c2428/i-made-my-coding-agent-check-for-discounts-before-it-pays-for-anything-5bmm</guid>
      <description>&lt;p&gt;Coding agents now pick SaaS plans, APIs, hosting and domains for us. They follow instructions well and almost never ask whether there is a cheaper way to get the same result.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fiqmnidkiwc1y5wbxjx3p.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fiqmnidkiwc1y5wbxjx3p.png" alt="Claude Code with frugal installed, answering a request to upgrade 8 people to Slack Pro with a sourced table of cheaper options"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The 144 dollar example
&lt;/h2&gt;

&lt;p&gt;Ask Claude Code: "Our team of 8 needs Slack Pro. Walk me through upgrading." You get correct steps for monthly billing. Annual billing is 144 USD a year cheaper, nonprofits get Pro free, and schools get 85 percent off. None of that came up.&lt;/p&gt;

&lt;h2&gt;
  
  
  What frugal does
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://github.com/fainir/frugal" rel="noopener noreferrer"&gt;frugal&lt;/a&gt; is an &lt;a href="https://agentskills.io" rel="noopener noreferrer"&gt;Agent Skill&lt;/a&gt; plus a tiny optional hook. Whenever a task involves paying, the agent:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Sizes the check to the amount: one line for a 20 USD purchase, a short table for a 500 USD plan, a negotiation plan for a contract.&lt;/li&gt;
&lt;li&gt;Climbs a ladder, cheapest first: don't pay (free tier, open source, something you already own), pay less for the same thing (plan, seats, annual billing, refurbished, timing), programs (startup, student, open-source, nonprofit), codes and stacking, ask (sales discounts, retention offers), switch (a cheaper alternative).&lt;/li&gt;
&lt;li&gt;Verifies on the live page and labels every number verified, reported or unverified. It never invents a promo code.&lt;/li&gt;
&lt;li&gt;Reports in a few lines, then carries on with the task. It never pays, enters card details or accepts terms by itself.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;It also covers the money the agent spends while coding: picking a model for an LLM feature (batch APIs, prompt caching, smaller models), choosing hosting, or adding a paid SaaS dependency.&lt;/p&gt;

&lt;h2&gt;
  
  
  Did it help?
&lt;/h2&gt;

&lt;p&gt;I ran eight everyday requests through &lt;code&gt;claude -p&lt;/code&gt; from an empty folder, with and without the plugin, and had a separate model judge both answers blind on five criteria: saving found, sourcing, honesty, helpfulness and length. Answers with frugal averaged 9.8 out of 10, against 5.5 without. Without it, Claude Code declined most shopping questions as out of scope. With it, it found the 199 USD AirPods price against a 249 USD list price, the 2-year minimum term on &lt;code&gt;.ai&lt;/code&gt; domains, and the real cost of a dropshipping stack after the 1 USD trials end.&lt;/p&gt;

&lt;p&gt;It is a small sample with a model judge, so treat it as a sanity check rather than a benchmark.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx skills add fainir/frugal
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Or in Claude Code:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/plugin marketplace add fainir/frugal
/plugin install frugal@frugal
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It works in Claude Code, Codex, Cursor, Gemini CLI, Copilot and OpenCode. The code, the evals and the program catalog (checked against official pages this week) are on GitHub: &lt;a href="https://github.com/fainir/frugal" rel="noopener noreferrer"&gt;https://github.com/fainir/frugal&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;I would love to hear where it gets the numbers wrong.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>productivity</category>
      <category>opensource</category>
    </item>
    <item>
      <title>I BUILT MyFeed - describe any topic, get an AI-curated feed from thousands of sources</title>
      <dc:creator>Nir Feinstein</dc:creator>
      <pubDate>Wed, 15 Apr 2026 04:29:12 +0000</pubDate>
      <link>https://dev.to/nir_feinstein_9a91c1c2428/i-built-myfeed-describe-any-topic-get-an-ai-curated-feed-from-thousands-of-sources-gb2</link>
      <guid>https://dev.to/nir_feinstein_9a91c1c2428/i-built-myfeed-describe-any-topic-get-an-ai-curated-feed-from-thousands-of-sources-gb2</guid>
      <description>&lt;p&gt;I JUST LAUNCHED {MyFeed](&lt;a href="https://myfeed.space" rel="noopener noreferrer"&gt;https://myfeed.space&lt;/a&gt;) and wanted to share it with the DEV community.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is it?
&lt;/h2&gt;

&lt;p&gt;MyFeed is an AI-powered content aggregator. You describe any topic you care about, and it builds a curated feed from thousands of sources across the web - RSS feeds, news sites, blogs, YouTube, and more.&lt;/p&gt;

&lt;p&gt;No manual RSS subscriptions. No bookmarking dozens of sites. Just describe what you want to read about.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it works
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Describe a topic&lt;/strong&gt; - "Python Django tutorials", "SpaceX launches", "indie game dev" - anything&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AI classifies content&lt;/strong&gt; - An LLM scans thousands of articles and matches them to your feed based on relevance&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Get a curated feed&lt;/strong&gt; - Fresh content delivered to your feed, updated regularly&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Tech stack
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Next.js 16&lt;/strong&gt; + React 19 + Tailwind CSS&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Supabase&lt;/strong&gt; for auth and database&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;LLM classification&lt;/strong&gt; (GPT-5 Nano) for matching articles to feeds&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;RSS + Brave Search API&lt;/strong&gt; for content sourcing&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Railway&lt;/strong&gt; for hosting&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Resend&lt;/strong&gt; for email digests&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The classification pipeline
&lt;/h2&gt;

&lt;p&gt;The interesting part is how content gets matched to feeds. Every few hours, a cron job:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Scans 100+ RSS sources and Brave Search for fresh articles&lt;/li&gt;
&lt;li&gt;Sends all new articles + all feed prompts to an LLM in one batch&lt;/li&gt;
&lt;li&gt;The LLM returns matches with quality scores (1-10)&lt;/li&gt;
&lt;li&gt;Only articles scoring 6+ get inserted into feeds&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;This keeps costs under $1/month while processing thousands of articles.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it out
&lt;/h2&gt;

&lt;p&gt;Check it out at &lt;a href="https://myfeed.space" rel="noopener noreferrer"&gt;myfeed.space&lt;/a&gt; - it's free to use. You can browse 15+ pre-built feeds or create your own custom feed in seconds.&lt;/p&gt;

&lt;p&gt;Would love any feedback from the community!&lt;/p&gt;

</description>
      <category>i</category>
      <category>built</category>
      <category>myfeed</category>
      <category>an</category>
    </item>
  </channel>
</rss>
