weft-grow
SkillDev toolsLets 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.
Account requirements not reviewed. Check the skill instructions before use; Ahel provides instructions and does not run this skill.
No other account needed.
Add Ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.
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.
- Read
weft-sdpandweft-running. The user's argument names the stage to grow or the saved example to rerun. - Read
weft treeandweft examples. Checkpoint work you may want back withweft checkpoint start-<what>. - 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. - 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 usesweft bakefollowed byweft run --fire '<trigger>=<wake-json>';--emit '<node>={"port":value}'supplies outputs without executing that node. Read the actual interface before choosing payloads. - During iteration, add
--seedto reuse compatible work.--seed-before <node>and--seed-until <node>bound reuse. Verify that the changed nodes ran, then inspect their values withweft events <color> --node <id> --full. A successful status alone does not prove the result. - 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--expectfor more focus nodes. Grow the next stage, then exercise a second real input. - After changes,
weft run <name> --detachreruns a relevant example on current code with its saved parameters. Leave out--seedwhen reviewing the new computation. Compare withweft diff <color> example:<name> --full; inspect the differences and judge them against the user's intent. Replace the accepted example withweft freeze <name> <color>only after accepting the new result. - 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. - 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