When you ask an agent to fix a trade-off, faster but not sloppier or lighter
but not weaker, it usually splits the difference. This
agent skill teaches it TRIZ, the theory of inventive
problem solving. The agent finds the contradiction behind your problem and
looks it up in Altshuller's contradiction matrix and the 40 inventive
principles. It turns the results into steps that use what you already have,
checks them with you, and hands back a short plan and a full one.
Install
Ask your agent. Paste this into Claude Code or any agent that supports
skills:
Install the TRIZ skill from https://github.com/magery-ai/triz-skill
Claude Code (plugin marketplace):
/plugin marketplace add magery-ai/triz-skill
/plugin install triz@triz-skill
Claude.ai: download triz-skill-<version>.zip from the
latest release and
upload it under Settings > Capabilities > Skills.
Any agent with the skills CLI:
npx skills add magery-ai/triz-skill --skill triz
Agents without skill support (Cursor rules, Copilot, Windsurf, Cline):
copy portable/AGENTS.md into your project.
Manual install and other options are in INSTALL.md.
Just want the matrix? The same data is browsable as a free
online contradiction matrix, with
a page for each of the 40 principles.
Then describe a problem where improving one thing makes another worse, for
example: "Our pages need to be fast, but the cache shows stale prices."
See it work
You: Our support team has to answer tickets faster, but every time we
push for speed the answers get sloppier and customers write back.
Agent: Let's work through this with TRIZ. Two quick questions first,
one at a time. What does "fast" mean for you today, and what is the target?
You: First reply takes about 9 hours. We want under 1 hour.
Agent: What have you already tried, and what can't change?
You: We set a 1-hour target. Replies got faster, but reopen rates went
from 8% to 19%. No new hires this year, we use Zendesk, and senior agents
are our best asset.
Agent: Here's my restatement. A team of fixed size has to cut first reply
time from 9 hours to under 1 hour without raising the reopen rate above 8%.
Ideal result: every customer gets a correct answer at once, with no extra
staff. Resources: Zendesk macros and triggers, ticket history, senior
agents' knowledge.
The contradiction behind it: improving speed (parameter 25, Loss of
time) worsens reliability (27, Reliability) of the answers. In physical
form: the reply must be instant, so agents don't investigate, and
considered, so it is correct. Is that the problem, and does the conflict
match what you see?
You: Exactly that.
Agent: Speed here is the time customers wait (25, Loss of time) or
tickets handled per hour (39, Productivity). Reading both matrix cells against
27 Reliability, #10 shows up in both, so I'll lead with it, then #1 from
the second cell:
- #10 Preliminary action: do the investigating before the ticket arrives.
- #1 Segmentation: split tickets that can be answered instantly from those
that can't.
Concrete options, ranked:
- Seniors write and own macros for the 20 most common ticket types, which
cover about 60% of volume (#10). Low effort, high fit.
- A Zendesk trigger routes tickets that match a macro to an "instant" queue
and everything else to "investigate", with an honest 4-hour promise (#1).
- An intake form asks for the order number and a screenshot before a human
sees the ticket (#10).
One question before we commit: who would object to option 1, and why?
You: Seniors might hate writing canned replies.
Agent: What if they own them instead: each senior maintains five macros
and reviews reopen data for them weekly? Does that answer the objection?
You: Yes, they'd like owning the quality numbers.
How it works
flowchart LR
A[1. Intake<br/>restate, ideal result,<br/>resources] --> B[2. Contradiction<br/>improving A<br/>worsens B]
B --> C[3. Directions<br/>matrix and<br/>separation]
C --> D[4. Actions<br/>steps that use<br/>your resources]
D --> E[5. Validation<br/>two Socratic<br/>questions]
E --> F[6. Plan<br/>short in chat,<br/>full in a file]
E -. rethink .-> B
E -. more options .-> D
The agent stops to check with you three times: after it restates the problem,
after it names the contradiction, and before it commits to an option. A
guided session takes about 5-8 of your replies.
Every plan has three required fields: Resources (what it uses, and
anything new with its cost), Watch for (the top risk, a measurable signal
and what to do) and Done when (a metric, a target and a date). The full
plan is saved to triz-plans/YYYY-MM-DD-<slug>.md in your project.
| Mode | When to use it | What changes |
|---|
| Guided (default) | You want to think it through | Every check, at most two questions at each |
| Fast | Say "just give me options" | One check, and you can skip validation |
| Autonomous | No human in the loop | The agent answers its own checks, lists its assumptions and flags low-confidence plans |
Results
We ran all 13 eval cases three times with the skill and three times without
it, and had a separate agent grade each pair of conversations blind. With
the skill, every plan in every run named the conflict, set a dated target
and gave a risk threshold, and the grader preferred it in 38 of 39
comparisons. Without it, the core idea was often just as good, but only 4
of 39 plans had a dated target. The skill takes about five more of your
replies. These are simulated users and the same model family plays every
role, so read it as a strong signal, not a benchmark. See the
full results and method.
FAQ
Can't my agent do this without the skill?
It knows what TRIZ is, but left alone it tends to trade one side off against
the other or give a generic list. The skill makes it name the contradiction,
look up the actual matrix cell, tie each option to your resources and end with
a plan you can measure. In our evals, plans without the skill had a dated
target in 4 of 39 runs; with it, 39 of 39. See Results.
Is it only for engineering?
No. A mapping guide translates software, business and personal problems into
TRIZ parameters, and the examples above cover all three kinds.
How accurate is the contradiction matrix?
Three independent copies agree on 1,492 of 1,521 cells. The other 29 are
marked as disputed, and the agent shows both values. The cells have not yet
been checked against the 1997 book. See Sources.
Does it need Python?
No. A small lookup script (standard library only) makes matrix lookups
faster. Agents that can't run scripts read the data files directly.
Where does the plan go?
The short plan stays in the chat. The full plan, with the analysis and the
validation log, goes to triz-plans/ in your project, or into the chat if the
agent has no file access.
Contributing
Worked examples, eval cases and matrix corrections are welcome. See
CONTRIBUTING.md for the repository layout and how to run
the tests, and CHANGELOG.md for what changed in each release.
To report a wrong matrix cell, open a
Wrong matrix cell
issue.
Sources
The skill uses only the classic 1971 Altshuller 39x39 contradiction matrix and
the 40 inventive principles. Descriptions and data are written in this
project's own words and format.
Matrix values come from the triz40.com
online matrix (Source: TRIZ40 - www.triz40.com), cross-checked cell by cell
against a second copy. Differences are logged in
skills/triz/data/matrix-discrepancies.md. A third copy with a documented
origin, the table on the G.S. Altshuller Foundation
site (from Altshuller's Algorithm of Invention, 1973), agrees on 1,492 of
1,521 cells; the 29 cells where it differs are marked disputed, and the
lookup script says so. The cells have not yet been checked against the primary
reference, Altshuller's 40 Principles (Shulyak and Rodman, 1997);
skills/triz/data/matrix-status.json tracks each cell's status. If you have
the book and find a difference, please open a
Wrong matrix cell
issue.
License
MIT
Star history
