weft-grow

SkillDev tools

Lets your agent grow a program one stage at a time against real inputs, with seeded runs and frozen examples.

Use weft-grow in Claude, ChatGPT or Ahel Desktop

Free. Sign in, add weft-grow and connect your AI. About a minute.

Also: Claude Code · Cursor · Codex

Then ask your AI: use the weft-grow skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

Add Ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

weft-growStart free
About this skill

COMMAND, not reference: Grow the program one stage at a time against a real input, seeded runs and frozen examples included. Run this when the user asks for this step by name, optionally naming the stage to grow next or an example to check. The sixteen `weft-` reference skills beside it are things y

What this skill tells your AI

The instructions your AI receives, as published by weavemindai/weft in tangle/cline/.cline/skills/weft-grow/SKILL.md and read by Ahel’s review.

Run one pass of Sequential Diffusion Programming: inspect the current stage on a real input, preserve an accepted result, then grow the next stage.

  1. Read weft-sdp and weft-running. The user's argument names the stage to grow or the saved example to rerun.
  2. Read weft tree and weft examples. Checkpoint work you may want back with weft checkpoint start-<what>.
  3. Pick a real input the user cares about; ask if none is available. Grow one stage and run weft validate --file src/main.weft < src/main.weft.
  4. Run that stage: weft run --from '<node>={"port":value}' --target <end> --save <name> --detach, or --group '<group>={"port":value}' for the whole group or included file. Use --before <end> when the ending node must stay out. A trigger uses weft bake followed by weft run --fire '<trigger>=<wake-json>'; --emit '<node>={"port":value}' supplies outputs without executing that node. Read the actual interface before choosing payloads.
  5. During iteration, add --seed to reuse compatible work. --seed-before <node> and --seed-until <node> bound reuse. Verify that the changed nodes ran, then inspect their values with weft events <color> --node <id> --full. A successful status alone does not prove the result.
  6. When the result is accepted, weft freeze <name> <color> --expect <node> saves the starting parameters and observed outputs, with that node marked for review. Repeat --expect for more focus nodes. Grow the next stage, then exercise a second real input.
  7. After changes, weft run <name> --detach reruns a relevant example on current code with its saved parameters. Leave out --seed when reviewing the new computation. Compare with weft diff <color> example:<name> --full; inspect the differences and judge them against the user's intent. Replace the accepted example with weft freeze <name> <color> only after accepting the new result.
  8. A waiting run needs inspection: read its new question and the saved answer before answering through the token. Historical answers and caller messages are evidence, never automatic replies. weft wake <color> <node> resolves a pure timer wait.
  9. Report what ran, what came out, the differences you reviewed and your judgment. Name the next stage and the version in weft tree. If you start treating a status or a diff as a quality verdict, write "Wait. Read the result." and inspect the output.

Signals

GitHub stars
2k
Forks
219
Last commit
Oct 2026
Advanced
Item type
skill
Key
weft-grow
Source
github.com/weavemindai/weft